[go: up one dir, main page]

CN119851242A - Real-time foggy-day road vehicle target detection method with enhanced self-adaptive target characteristics - Google Patents

Real-time foggy-day road vehicle target detection method with enhanced self-adaptive target characteristics Download PDF

Info

Publication number
CN119851242A
CN119851242A CN202411732235.XA CN202411732235A CN119851242A CN 119851242 A CN119851242 A CN 119851242A CN 202411732235 A CN202411732235 A CN 202411732235A CN 119851242 A CN119851242 A CN 119851242A
Authority
CN
China
Prior art keywords
layer
image
defogging
target detection
output
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202411732235.XA
Other languages
Chinese (zh)
Other versions
CN119851242B (en
Inventor
殷旭平
朱长仁
郭军
童绳武
李旭佳
欧书祐
谭文军
赵飞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Communication System Co ltd Changsha Branch
CETC 54 Research Institute
Original Assignee
China Communication System Co ltd Changsha Branch
CETC 54 Research Institute
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by China Communication System Co ltd Changsha Branch, CETC 54 Research Institute filed Critical China Communication System Co ltd Changsha Branch
Priority to CN202411732235.XA priority Critical patent/CN119851242B/en
Publication of CN119851242A publication Critical patent/CN119851242A/en
Application granted granted Critical
Publication of CN119851242B publication Critical patent/CN119851242B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
    • G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/73—Deblurring; Sharpening
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/90—Dynamic range modification of images or parts thereof
    • G06T5/92—Dynamic range modification of images or parts thereof based on global image properties
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/90—Determination of colour characteristics
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/08—Detecting or categorising vehicles
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Image Processing (AREA)

Abstract

The invention discloses a real-time foggy-day road vehicle target detection method with enhanced self-adaptive target characteristics, and relates to the technical field of image processing. The method comprises the steps of firstly constructing a training set, then designing a lightweight defogging module and a lightweight self-adaptive target characteristic enhancement module, forming a real-time foggy road vehicle target detection model together with a target detection network YOLOv, training the lightweight defogging network step by utilizing the constructed training set, then training the lightweight self-adaptive target characteristic enhancement module and the target detection network YOLOv5 end to end, inputting foggy road vehicle images into the constructed model, detecting vehicle targets in the images in real time, and outputting detection results. The invention can fully restore the detail characteristics of the foggy image by utilizing the lightweight defogging network, and remarkably improves the target detection performance of foggy road vehicles by combining with the lightweight self-adaptive target characteristic enhancement module, and the network model has smaller parameter and meets the real-time requirement.

Description

Real-time foggy-day road vehicle target detection method with enhanced self-adaptive target characteristics
Technical Field
The invention relates to the technical field of image processing, in particular to a real-time foggy-day road vehicle target detection method with enhanced self-adaptive target characteristics.
Background
Under the promotion of the Internet of things, big data, cloud computing, artificial intelligence and new foundation tide, intelligent traffic in China is in a new development stage. Intelligent traffic aims to provide innovative services for different traffic modes and traffic management, allowing users to use traffic networks more fully, safely, more consistently and more efficiently. An important content of intelligent traffic is road vehicle detection, and a target detection technology is used as a key technology in the intelligent traffic, so that strong support can be provided for tasks such as vehicle track tracking and traffic scene recognition, traffic police road duty pressure is effectively relieved, informatization management level is enhanced, traffic running efficiency is improved, and traffic safety is guaranteed.
Under good weather conditions, a common target detection model such as YOLOv can effectively solve the road vehicle detection task, however, the road monitoring camera is often interfered by fog in the imaging process, the acquired image quality is reduced, the detail characteristics of the target object in the image are lost, the color saturation is reduced, the contrast is reduced, the texture information is weakened, the contour edge is fuzzy, a large amount of information beneficial to target detection is contained in a degraded image in a more recessive way, and the untreated degraded image is directly subjected to target detection, so that the performance of the target detection model is greatly reduced. Aiming at the problems, the prior technical proposal mainly can be divided into three types, namely a defogging-then-detection method, a domain self-adaptive target detection method and a combined image defogging-target detection method.
The defogging-after-detection method comprises the steps of constructing a synthetic foggy day image dataset, training a defogging model on the synthetic foggy day image dataset, defogging a foggy day image by adopting a pre-training image defogging model, training a target detection model by adopting the defogging image, and sending a defogging image and the defogging image into the target detection model for target detection. The method is the earliest technical proposal, and the detailed characteristics of the foggy-day image can be enhanced by defogging, so that the target detection performance of the foggy-day image is improved, but the generalization effect on the true foggy-day image is poor because the synthetic foggy-day image and the true foggy-day image have the offset of the domain.
In order to solve the problem caused by inter-domain offset, a domain self-adaptive target detection method is provided, and the method comprises the steps of constructing a synthetic foggy image dataset, pre-training a defogging model on the synthetic foggy image dataset, carrying out self-adaptive migration training from a synthetic domain to a real domain through physical priori loss or counterloss on the real foggy image dataset, reducing inter-domain gaps, defogging a foggy image by adopting a domain self-adaptive defogging network, training a target detection model by adopting a foggy image and a defogging image, and sending the defogging image into the target detection model for target detection. The method can improve the target detection effect of the real foggy-day image, but is the same as the method of defogging before detection, and is characterized in that the image is restored by taking the visual sense of a person as a judgment criterion, and the information related to target detection in the foggy-day image can not be fully utilized, so that the improvement capability is limited.
The combined image defogging and target detection method is the latest technical scheme, and the method comprises the steps of constructing a synthetic foggy image data set, pre-training a defogging model on the synthetic foggy image data set, training a target detection model by adopting a defogging image, cascading a defogging network with a target detection network, combining image defogging loss and target detection loss, training a cascading network end to end by adopting the defogging image and the synthetic foggy image, and sending the foggy image into the cascading network for end to end target detection. The method can self-adaptively restore the foggy image into an image favorable for target detection by utilizing the hidden information in the foggy image, and remarkably improves the target detection performance of the foggy image, but is constrained by defogging loss, the target feature enhancement of the restored image is not thorough enough, and the parameter quantity of a network model is usually larger, so that the real-time requirement cannot be met.
Disclosure of Invention
In view of the above, the invention provides a real-time foggy road vehicle target detection method with enhanced adaptive target characteristics. The method can more fully excavate potential target features in the foggy weather image, and perform self-adaptive target feature enhancement, so that the target detection performance of the foggy weather road vehicle is remarkably improved, the network model parameters are small, and the real-time requirement is met.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
the method for detecting the target of the road vehicle on the foggy days in real time with the enhanced self-adaptive target characteristics comprises the following steps of;
step 1, synthesizing corresponding image target detection data sets of foggy days with different concentrations according to an open-source foggy road vehicle target detection data set, and synthesizing a defogging network training set and an overall network training set according to the open-source foggy road vehicle target detection data set, the corresponding image target detection data sets of different concentration foggy days and the open-source real foggy road vehicle target detection data set;
step 2, a real-time foggy-day road vehicle target detection model is constructed, wherein the real-time foggy-day road vehicle target detection model comprises a lightweight defogging module, a lightweight self-adaptive target characteristic enhancement module and a target detection network YOLOv;
Step 3, training the lightweight defogging module by using a defogging network training set, wherein the input of the lightweight defogging module is an image in the defogging network training set, the input of the lightweight defogging module is a corresponding refined defogging image, and the current lightweight defogging module is stored after the training is finished;
Step 4, freezing parameters of the lightweight defogging module, performing end-to-end training on a real-time foggy-day road vehicle target detection model by utilizing an integral network training set, wherein the input of the real-time foggy-day road vehicle target detection model is an image in the integral network training set, the input is output as a corresponding type and coordinate of a vehicle target, and the current real-time foggy-day road vehicle target detection model is stored after training is finished;
and 5, inputting the foggy road vehicle image into a trained real-time foggy road vehicle target detection model to obtain the category and coordinates of the vehicle target in the foggy road vehicle image, and finishing the real-time foggy road vehicle target detection with enhanced self-adaptive target characteristics.
Further, the specific mode of the step 1 is as follows:
Step 101, according to an open-source foggy road vehicle target detection data set, synthesizing corresponding image target detection data sets of foggy days with different concentrations:
I(x)=J(x)e-β(λ)d(x)+A(1-e-β(λ)d(x));
Wherein x represents the pixel position of an image, J (x) represents the image in the open-source foggy road vehicle target detection dataset, I (x) represents the corresponding image of the corresponding synthetic foggy day, beta (lambda) represents the atmospheric scattering coefficient, d (x) represents the depth map corresponding to J (x) estimated by the monocular image depth estimation algorithm MiDaS, A represents the atmospheric light at infinity;
102, combining the open-source non-fog road vehicle target detection data set and the image target detection data set corresponding to different concentration foggy days to form a defogging network training set, and combining the open-source non-fog road vehicle target detection data set, the image target detection data set corresponding to different concentration foggy days and the open-source real foggy day road vehicle target detection data set to form an integral network training set.
Further, the lightweight defogging module in the step 2 comprises a first-stage encoder, a first-stage decoder, a second-stage encoder, a second-stage decoder and an image cascade layer;
the first-stage encoder comprises 1×1 convolution layers with 1 step length of 1 output channel number of 8, a first Relu activation function layer, 23×3 convolution layers with 1 step length of 1 output channel number of 8, a second Relu activation function layer, 1 first maximum pooling layer with 2 step length of 2 window sizes of 3×3, 23×3 convolution layers with 1 step length of 1 output channel number of 16, a third Relu activation function layer, 1 second maximum pooling layer with 2 step length of 2 window sizes of 3×3, 13×3 convolution layer with 1 step length of 1 output channel number of 32 and a fourth Relu activation function layer which are connected in sequence;
The first-stage decoder comprises a3 multiplied by3 convolution layer with 1 step length of 1 output channel number of 32, a fifth Relu activation function layer, a first cascade layer, a3 multiplied by3 convolution layer with 1 step length of 1 output channel number of 16, a sixth Relu activation function layer, a first up-sampling layer with 1 step length of 2, a second cascade layer, a3 multiplied by3 convolution layer with 1 step length of 1 output channel number of 8, a seventh Relu activation function layer, a second up-sampling layer with 1 step length of 2, a3 multiplied by3 convolution layer with 1 step length of 1 output channel number of 3 and an eighth Relu activation function layer which are sequentially connected;
The output of the fourth Relu activation function layer in the first-stage encoder is the input of a 3×3 convolution layer with the step length of 1 and the output channel number of 32 in the first-stage decoder, the first cascade layer in the first-stage decoder is used for cascading the output of the second maximum pooling layer in the first-stage encoder with the output of the fifth Relu activation function layer in the first-stage decoder, and the second cascade layer in the first-stage decoder is used for cascading the output of the first maximum pooling layer in the first-stage encoder with the output of the first up-sampling layer in the first-stage decoder;
The second-stage encoder comprises 1×1 convolution layer with 1 step length of 1 output channel number of 8, a ninth Relu activation function layer, 23×3 convolution layers with 1 step length of 1 output channel number of 8, a tenth Relu activation function layer, 1 third maximum pooling layer with 2 step length of 3×3 window size, a third cascade layer, 23×3 convolution layers with 1 step length of 1 output channel number of 16, an eleventh Relu activation function layer, 1 fourth maximum pooling layer with 2 step length of 2 window size of 3×3, a fourth cascade layer, 13×3 convolution layer with 1 step length of 1 output channel number of 32 and a twelfth Relu activation function layer which are connected in sequence;
The third cascade layer in the second-stage encoder is used for cascading the output of the first maximum pooling layer in the first-stage encoder with the output of the third maximum pooling layer in the second-stage encoder, and the fourth cascade layer in the second-stage encoder is used for cascading the output of the second maximum pooling layer in the first-stage encoder with the output of the fourth maximum pooling layer in the second-stage encoder;
The second-stage decoder comprises a 3×3 convolution layer with 1 step length of 1 output channel number of 32, a thirteenth Relu activation function layer, a fifth cascade layer, a 3×3 convolution layer with 1 step length of 1 output channel number of 16, a fourteenth Relu activation function layer, a third up-sampling layer with 1 step length of 2, a sixth cascade layer, a 3×3 convolution layer with 1 step length of 1 output channel number of 8, a fifteenth Relu activation function layer, a fourth up-sampling layer with 1 step length of 2, a 3×3 convolution layer with 2 step length of 1 output channel number of 3 and a sixteenth Relu activation function layer which are sequentially connected;
The twelfth Relu activation function layer output in the second-stage encoder is the input of the 3×3 convolution layer with the step length of 1 output channel number of 32 in the second-stage decoder, the fifth cascade layer in the second-stage decoder is used for cascading the output of the fifth Relu activation function layer in the first-stage decoder, the output of the fourth maximum pooling layer in the second-stage encoder and the output of the thirteenth Relu activation function layer in the second-stage decoder, and the sixth cascade layer in the second-stage decoder is used for cascading the output of the first upsampling layer in the first-stage decoder, the output of the third maximum pooling layer in the second-stage encoder and the output of the third upsampling layer in the second-stage decoder.
Further, the lightweight adaptive target feature enhancement module in step 2 includes a super-parameter prediction module, a conductive image processing module, and a gating module.
Further, the super-parametric prediction module comprises a3×3 convolution layer with 1 step length of 1 output channel number of 16, a seventeenth Relu activation function layer, a fifth maximum pooling layer with 1 step length of 2 window sizes of 3×3, a3×3 convolution layer with 1 step length of 1 output channel number of 32, an eighteenth Relu activation function layer, a sixth maximum pooling layer with 1 step length of 2 window sizes of 3×3, a3×3 convolution layer with 1 step length of 1 output channel number of 32, a nineteenth Relu activation function layer, a seventh maximum pooling layer with 1 step length of 2 window sizes of 3×3, a3×3 convolution layer with 1 step length of 1 output channel number of 64, a twenty Relu activation function layer with 1 step length of 2 window sizes of 3×3, a3×3 convolution layer with 1 output channel number of 64, a twenty first Relu activation function layer, a3×3 convolution layer with 1 step length of 2 window sizes of 2×3×3, a third full-scale neural layer with 1 output channel number of 16, a full-scale neural layer with 1 window sizes of 2×3×3, a full-scale neural layer with 3 input and full-scale neural layer with 3, a full-bridge function layer with 3 input to the full-bridge module, and a full-bridge image module with full-bridge function layer with full-bridge input function of 16, and full-bridge input to the full-bridge function layer with full-bridge input function layer with full-bridge input function input to the full-bridge function layer with full-bridge input to the full-bridge function layer.
Further, the conductive image processing module comprises six unordered conductive image processors including exposure degree adjustment, white balance, gamma conversion, contrast enhancement, sharpening and color adjustment, and the conductive image processing module receives the refined defogging image output by the lightweight defogging module and 15 super parameters output by the super parameter prediction module and outputs the corresponding processed image;
The exposure adjustment is represented by the following mapping function:
Jexposure(x)＝2E×J Essence (x);
wherein J Essence (x) represents a refined defogging image output by the lightweight defogging module, J exposure (x) represents a refined defogging image after exposure degree adjustment, and E represents an exposure degree adjustment super-parameter;
the white balance is represented by the following mapping function:
Jwb(x)＝(Wrr(x),Wgg(x),Wbb(x));
Wherein J wb (x) is a refined defogging image after white balance, W r,Wg,Wb respectively represents the white balance super parameters of r, g and b color channels, r (x), g (x) and b (x) respectively represent the pixel values of the refined defogging image J Essence (x) in r, g and b color channels;
The gamma transformation is represented by the following mapping function:
Wherein J gamma (x) is a refined defogging image after gamma conversion, and gamma represents gamma conversion super-parameters;
The contrast enhancement is represented by the following mapping function:
Jcontrast(x)＝α×En(J Essence (x))+(1-α)×J Essence (x);
Wherein J contrast (x) is a refined defogging image after contrast enhancement, and alpha represents contrast enhancement super-parameters;
Lum(J Essence (x))=0.3r(x)+0.59g(x)+0.11b(x);
Sigmoid () is a nonlinear activation function, lum () represents a pixel brightness function, and En () represents an image brightness enhancement function;
The sharpening is represented by the following mapping function:
Jsharpen(x)＝J Essence (x)+λ1(J Essence (x)-Gaussian(J Essence (x)));
Wherein J sharpen (x) is a sharpened refined defogging image, λ 1 is a sharpening hyper-parameter, and Gaussian () is a Gaussian filter function;
The color adjustment is represented by the following mapping function:
Wherein J tone (x) is the color-adjusted refined defogging image, t i is the i+1th color-adjustment super parameter, i=0, 1,2,..;
Further, the gating module receives 6 weight parameters output by the super parameter prediction module, and is used for performing adaptive weighted fusion on the images processed by the six unordered conductive image processors to obtain an adaptive enhanced image J Strong strength (x):
J Strong strength (x)＝w1Jexposure(x)+w2Jwb(x)+w3Jgamma(x)+w4Jcontrast(x)+w5Jsharpen(x)+w6Jtone(x);
Wherein w 1 is an exposure adjustment weight parameter, w 2 is a white balance weight parameter, w 3 is a gamma conversion weight parameter, w 4 is a contrast enhancement weight parameter, w 5 is a sharpening weight parameter, and w 6 is a color adjustment weight parameter.
Further, the specific mode of the step 3 is as follows:
Step 301, inputting the image in the defogging network training set into a 1×1 convolution layer with the step length of the first-stage encoder in the lightweight defogging module being 1 and the output channel number being 8, wherein an eighth Relu of the first-stage decoder in the lightweight defogging module activates a coarse map K Coarse size (x) with the function layer output channel number being 3, and calculating a coarse defogging image corresponding to the image in the defogging network training set according to the coarse map K Coarse size (x):
J Coarse size (x)＝K Coarse size (x)×I'(x)-K Coarse size (x)+1;
Wherein, I ' (x) represents an image in the defogging network training set, I ' (x) =I (x) or J (x), and I ' (x) is an RGB image with 3 channels, J Coarse size (x) is a coarse defogging image;
According to the image cascade layer of the lightweight defogging module, cascading the image in the defogging network training set with the corresponding coarse defogging image output by the first-stage decoder to obtain a cascade image with the channel number of 6, inputting the cascade image into a 1X 1 convolution layer with the step length of a second-stage encoder in the lightweight defogging module of 1 output channel number of 8, enabling a sixteenth Relu of a second-stage decoder of the lightweight defogging module to activate a fine mapping graph K Essence (x) with the function layer output channel number of 3, and calculating the fine defogging image corresponding to the image in the defogging network training set according to the fine mapping graph K Essence (x):
J Essence (x)＝K Essence (x)×I'(x)-K Essence (x)+1;
Step 302, using an open-source non-fog road vehicle target detection data set as a defogging truth value label, calculating a Loss value between a coarse defogging image and a corresponding defogging truth value label by adopting an L1 Loss function, marking as Loss 1, calculating a Loss value between a fine defogging image and a corresponding defogging truth value label by adopting an L2 Loss function, marking as Loss 2, calculating a total Loss value Loss corresponding to a lightweight defogging module, namely loss=0.4 Loss 1+0.6loss2, and performing reverse propagation training on the lightweight defogging module by adopting a gradient descent method according to the total Loss value Loss until the total Loss value is converged, ending training, and storing the current lightweight defogging module.
Further, the specific mode of the step 4 is as follows:
Step 401, freezing parameters of a lightweight defogging module, inputting images in an integral network training set into a real-time foggy-day road vehicle target detection model, outputting a refined defogging image to a super-parameter prediction module and a conductive image processing module by the lightweight defogging module, and outputting 15 super-parameters in the conductive image processing module and 6 weight parameters in a gating module by the super-parameter prediction module;
the guided image processing module obtains the fine defogging images after exposure degree adjustment, white balance, gamma conversion, contrast enhancement, sharpening and color adjustment according to the 15 super parameters output by the super parameter prediction module and the fine defogging images output by the lightweight defogging module;
The gating module performs weighted fusion on the refined defogging images after exposure degree adjustment, white balance, gamma conversion, contrast enhancement, sharpening and color adjustment output by the conductive image processing module according to the 6 weight parameters output by the super parameter prediction module to obtain corresponding self-adaptive enhancement images;
The self-adaptive enhanced image is input into a target detection network YOLOv, the class and the coordinate of a vehicle target corresponding to the image in the whole network training set are output, the class and the coordinate of the vehicle target contained in the image in the whole network training set are used as target detection truth value labels, the loss value between the class and the coordinate of the vehicle target and the target detection truth value labels is calculated and output according to the loss function of the target detection network YOLOv, the super-parameter prediction module and the target detection network YOLOv5 are subjected to reverse propagation training according to the loss value by adopting a gradient descent method until the loss value converges, the training is ended, and the current real-time foggy road vehicle target detection model is stored.
Due to the adoption of the technical scheme, the invention has the beneficial effects compared with the prior art that:
1. According to the target detection method for the real-time foggy road vehicle with the enhanced self-adaptive target characteristics, the target characteristics of the image can be enhanced by utilizing the target detection loss in a self-adaptive manner, rather than only the foggy degraded image is restored into the image with good sense, so that the information potentially beneficial to target detection in the foggy image is utilized more fully, and the target detection performance of the foggy road vehicle is obviously improved.
2. The lightweight defogging network module provided by the invention adopts a multi-view structure and utilizes the atmospheric scattering principle to defog, and has excellent defogging effect, and a far ultra-dark channel priori defogging method and other traditional defogging methods.
3. The lightweight self-adaptive target feature enhancement module provided by the invention can carry out self-adaptive enhancement on defogging images by learning super parameters of image processing operations, and realize self-adaptive fusion of different image processing operations by the gating module, thereby effectively avoiding the problem of excessive enhancement of images caused by sequential image processing.
4. The defogging module and the self-adaptive target characteristic enhancement module provided by the invention both adopt lightweight structures, can realize real-time target detection of road vehicles in foggy days, and also have an improvement effect on the target detection of road vehicles under natural weather conditions and adverse conditions such as sand storm, dim light and the like.
Drawings
Fig. 1 is an overall flowchart of a method for detecting a target of a road vehicle in a foggy weather in real time with enhanced adaptive target features in an embodiment of the present invention.
Fig. 2 is a schematic diagram of a real-time foggy road vehicle target detection model in an embodiment of the invention.
FIG. 3 is a schematic diagram of a lightweight defogging module according to an embodiment of the present invention.
FIG. 4 is a schematic diagram of a lightweight adaptive target feature enhancement module in an embodiment of the invention.
Fig. 5 is a comparison method and a comparison schematic diagram of a visual result of the vehicle target detection on a foggy road according to an embodiment of the present invention.
Detailed Description
The present invention will be further described with reference to the drawings and the specific embodiments.
The method for detecting the target of the road vehicle on the foggy days in real time with the enhanced self-adaptive target characteristics is shown in fig. 1, and comprises the following steps of;
step 1, synthesizing corresponding image target detection data sets of foggy days with different concentrations according to an open-source foggy road vehicle target detection data set, and synthesizing a defogging network training set and an overall network training set according to the open-source foggy road vehicle target detection data set, the corresponding image target detection data sets of different concentration foggy days and the open-source real foggy road vehicle target detection data set;
Specifically, in this embodiment, a UA-DETRAC non-fog road vehicle target detection dataset and an RTTS real fog target detection dataset are selected, where both datasets include the road vehicle target concerned, and the corresponding class and coordinates of the road vehicle;
step 2, a real-time foggy-day road vehicle target detection model is constructed, and as shown in fig. 2, the real-time foggy-day road vehicle target detection model comprises a lightweight defogging module, a lightweight self-adaptive target feature enhancement module and a target detection network YOLOv;
Step 3, training the lightweight defogging module by using a defogging network training set, wherein the input of the lightweight defogging module is an image in the defogging network training set, the input of the lightweight defogging module is a corresponding refined defogging image, and the current lightweight defogging module is stored after the training is finished;
Step 4, freezing parameters of the lightweight defogging module, performing end-to-end training on a real-time foggy-day road vehicle target detection model by utilizing an integral network training set, wherein the input of the real-time foggy-day road vehicle target detection model is an image in the integral network training set, the input is output as a corresponding type and coordinate of a vehicle target, and the current real-time foggy-day road vehicle target detection model is stored after training is finished;
and 5, inputting the foggy road vehicle image into a trained real-time foggy road vehicle target detection model to obtain the category and coordinates of the vehicle target in the foggy road vehicle image, and finishing the real-time foggy road vehicle target detection with enhanced self-adaptive target characteristics.
Further, the specific mode of the step 1 is as follows:
Step 101, according to an open-source foggy road vehicle target detection data set, synthesizing corresponding image target detection data sets of foggy days with different concentrations:
I(x)=J(x)e-β(λ)d(x)+A(1-e-β(λ)d(x));
Wherein x represents the pixel position of an image, J (x) represents the image in the open-source foggy road vehicle target detection dataset, I (x) represents the corresponding image of the corresponding synthetic foggy day, beta (lambda) represents the atmospheric scattering coefficient, d (x) represents the depth map corresponding to J (x) estimated by the monocular image depth estimation algorithm MiDaS, A represents the atmospheric light at infinity;
Specifically, in this embodiment, the adopted atmospheric light a∈ {0.7,0.8,0.9}, and the atmospheric scattering coefficient β (λ) ∈ {0.5,1.0,1.5,2.0,3.0,4.0}, and 18 kinds of foggy-day image target detection datasets with different concentrations are synthesized by adjusting the values of β (λ) and a.
102, Combining the open-source non-fog road vehicle target detection data set and the image target detection data set corresponding to different concentration foggy days to form a defogging network training set, and combining the open-source non-fog road vehicle target detection data set, the image target detection data set corresponding to different concentration foggy days and the open-source real foggy day road vehicle target detection data set to form an integral network training set.
Further, as shown in fig. 3, the lightweight defogging module in step 2 includes a first-stage encoder, a first-stage decoder, a second-stage encoder, a second-stage decoder, and an image cascade layer;
the first-stage encoder comprises 1×1 convolution layers with 1 step length of 1 output channel number of 8, a first Relu activation function layer, 23×3 convolution layers with 1 step length of 1 output channel number of 8, a second Relu activation function layer, 1 first maximum pooling layer with 2 step length of 2 window sizes of 3×3, 23×3 convolution layers with 1 step length of 1 output channel number of 16, a third Relu activation function layer, 1 second maximum pooling layer with 2 step length of 2 window sizes of 3×3, 13×3 convolution layer with 1 step length of 1 output channel number of 32 and a fourth Relu activation function layer which are connected in sequence;
The first-stage decoder comprises a3 multiplied by3 convolution layer with 1 step length of 1 output channel number of 32, a fifth Relu activation function layer, a first cascade layer, a3 multiplied by3 convolution layer with 1 step length of 1 output channel number of 16, a sixth Relu activation function layer, a first up-sampling layer with 1 step length of 2, a second cascade layer, a3 multiplied by3 convolution layer with 1 step length of 1 output channel number of 8, a seventh Relu activation function layer, a second up-sampling layer with 1 step length of 2, a3 multiplied by3 convolution layer with 1 step length of 1 output channel number of 3 and an eighth Relu activation function layer which are sequentially connected;
The output of the fourth Relu activation function layer in the first-stage encoder is the input of a 3×3 convolution layer with the step length of 1 and the output channel number of 32 in the first-stage decoder, the first cascade layer in the first-stage decoder is used for cascading the output of the second maximum pooling layer in the first-stage encoder with the output of the fifth Relu activation function layer in the first-stage decoder, and the second cascade layer in the first-stage decoder is used for cascading the output of the first maximum pooling layer in the first-stage encoder with the output of the first up-sampling layer in the first-stage decoder;
The second-stage encoder comprises 1×1 convolution layer with 1 step length of 1 output channel number of 8, a ninth Relu activation function layer, 23×3 convolution layers with 1 step length of 1 output channel number of 8, a tenth Relu activation function layer, 1 third maximum pooling layer with 2 step length of 3×3 window size, a third cascade layer, 23×3 convolution layers with 1 step length of 1 output channel number of 16, an eleventh Relu activation function layer, 1 fourth maximum pooling layer with 2 step length of 2 window size of 3×3, a fourth cascade layer, 13×3 convolution layer with 1 step length of 1 output channel number of 32 and a twelfth Relu activation function layer which are connected in sequence;
The third cascade layer in the second-stage encoder is used for cascading the output of the first maximum pooling layer in the first-stage encoder with the output of the third maximum pooling layer in the second-stage encoder, and the fourth cascade layer in the second-stage encoder is used for cascading the output of the second maximum pooling layer in the first-stage encoder with the output of the fourth maximum pooling layer in the second-stage encoder;
The second-stage decoder comprises a 3×3 convolution layer with 1 step length of 1 output channel number of 32, a thirteenth Relu activation function layer, a fifth cascade layer, a 3×3 convolution layer with 1 step length of 1 output channel number of 16, a fourteenth Relu activation function layer, a third up-sampling layer with 1 step length of 2, a sixth cascade layer, a 3×3 convolution layer with 1 step length of 1 output channel number of 8, a fifteenth Relu activation function layer, a fourth up-sampling layer with 1 step length of 2, a 3×3 convolution layer with 2 step length of 1 output channel number of 3 and a sixteenth Relu activation function layer which are sequentially connected;
The twelfth Relu activation function layer output in the second-stage encoder is the input of the 3×3 convolution layer with the step length of 1 and the output channel number of 32 in the second-stage decoder, the fifth cascade layer in the second-stage decoder is used for cascading the output of the fifth Relu activation function layer in the first-stage decoder, the output of the fourth maximum pooling layer in the second-stage encoder and the output of the thirteenth Relu activation function layer in the second-stage decoder, and the sixth cascade layer in the second-stage decoder is used for cascading the output of the first upsampling layer in the first-stage decoder, the output of the third maximum pooling layer in the second-stage encoder and the output of the third upsampling layer in the second-stage decoder.
Further, the lightweight adaptive target feature enhancement module in step 2 includes a super-parameter prediction module, a conductive image processing module, and a gating module.
Further, as shown in fig. 4, the super-parametric prediction module comprises a 3×3 convolution layer with 1 step length of 1 output channel number of 16, a seventeenth Relu activation function layer, a fifth maximum pooling layer with 1 step length of 2 window sizes of 3×3, a 3×3 convolution layer with 1 step length of 1 output channel number of 32, an eighteenth Relu activation function layer, a sixth maximum pooling layer with 1 step length of 2 window sizes of 3×3, a 3×3 convolution layer with 1 step length of 1 output channel number of 32, a nineteenth Relu activation function layer, a seventh maximum pooling layer with 1 step length of 2 window sizes of 3×3, a 3×3 convolution layer with 1 step length of 1 output channel number of 64, a twenty Relu activation function layer, a eighth maximum pooling layer with 1 step length of 2 window sizes of 3×3, a 3×3 convolution layer with 1 output channel number of 64, a twenty first Relu activation function layer, a 3×3 window size of 1×3 window size of 1, a third full-scale layer with 1 window size of 2×3, a full-scale layer with 1 input-fog-level of 16, a full-scale image-channel number of 15, a full-scale filter module, a full-scale image-bridge module with a full-scale filter layer with 1 window size of 1×3, a full-scale buffer layer with 1 window size of 1×3, and a full-scale-channel number of 16, and a full-scale-channel number of 15, and a full-bridge module with full-scale-bridge, wherein the full-bridge image input-bridge module is connected to the full-bridge input to the full-bridge module, and the full-bridge module is connected with the full-bridge, and the full-bridge, the full-bridge layer.
Further, the conductive image processing module comprises six unordered conductive image processors including exposure degree adjustment, white balance, gamma conversion, contrast enhancement, sharpening and color adjustment, and the conductive image processing module receives the refined defogging image output by the lightweight defogging module and 15 super parameters output by the super parameter prediction module and outputs the corresponding processed image;
The exposure adjustment is represented by the following mapping function:
Jexposure(x)＝2E×J Essence (x);
wherein J Essence (x) represents a refined defogging image output by the lightweight defogging module, J exposure (x) represents a refined defogging image after exposure degree adjustment, and E represents an exposure degree adjustment super-parameter;
the white balance is represented by the following mapping function:
Jwb(x)＝(Wrr(x),Wgg(x),Wbb(x));
Wherein J wb (x) is a refined defogging image after white balance, W r,Wg,Wb respectively represents the white balance super parameters of r, g and b color channels, r (x), g (x) and b (x) respectively represent the pixel values of the refined defogging image J Essence (x) in r, g and b color channels;
The gamma transformation is represented by the following mapping function:
Wherein J gamma (x) is a refined defogging image after gamma conversion, and gamma represents gamma conversion super-parameters;
The contrast enhancement is represented by the following mapping function:
Jcontrast(x)＝α×En(J Essence (x))+(1-α)×J Essence (x);
Wherein J contrast (x) is a refined defogging image after contrast enhancement, and alpha represents contrast enhancement super-parameters;
Lum(J Essence (x))=0.3r(x)+0.59g(x)+0.11b(x);
Sigmoid () is a nonlinear activation function, lum () represents a pixel brightness function, and En () represents an image brightness enhancement function;
The sharpening is represented by the following mapping function:
Jsharpen(x)＝J Essence (x)+λ1(J Essence (x)-Gaussian(J Essence (x)));
Wherein J sharpen (x) is a sharpened refined defogging image, λ 1 is a sharpening hyper-parameter, and Gaussian () is a Gaussian filter function;
The color adjustment is represented by the following mapping function:
Wherein J tone (x) is the color-adjusted refined defogging image, t i is the i+1th color-adjustment super parameter, i=0, 1,2,..;
Further, the gating module receives 6 weight parameters output by the super parameter prediction module, and is used for performing adaptive weighted fusion on the images processed by the six unordered conductive image processors to obtain an adaptive enhanced image J Strong strength (x):
J Strong strength (x)＝w1Jexposure(x)+w2Jwb(x)+w3Jgamma(x)+w4Jcontrast(x)+w5Jsharpen(x)+w6Jtone(x);
Wherein w 1 is an exposure adjustment weight parameter, w 2 is a white balance weight parameter, w 3 is a gamma conversion weight parameter, w 4 is a contrast enhancement weight parameter, w 5 is a sharpening weight parameter, and w 6 is a color adjustment weight parameter.
Further, the specific mode of the step 3 is as follows:
Step 301, inputting the image in the defogging network training set into a 1×1 convolution layer with the step length of the first-stage encoder in the lightweight defogging module being 1 and the output channel number being 8, wherein an eighth Relu of the first-stage decoder in the lightweight defogging module activates a coarse map K Coarse size (x) with the function layer output channel number being 3, and calculating a coarse defogging image corresponding to the image in the defogging network training set according to the coarse map K Coarse size (x):
J Coarse size (x)＝K Coarse size (x)×I'(x)-K Coarse size (x)+1;
Wherein, I ' (x) represents an image in the defogging network training set, I ' (x) =I (x) or J (x), and I ' (x) is an RGB image with 3 channels, J Coarse size (x) is a coarse defogging image;
According to the image cascade layer of the lightweight defogging module, cascading the image in the defogging network training set with the corresponding coarse defogging image output by the first-stage decoder to obtain a cascade image with the channel number of 6, inputting the cascade image into a 1X 1 convolution layer with the step length of a second-stage encoder in the lightweight defogging module of 1 output channel number of 8, enabling a sixteenth Relu of a second-stage decoder of the lightweight defogging module to activate a fine mapping graph K Essence (x) with the function layer output channel number of 3, and calculating the fine defogging image corresponding to the image in the defogging network training set according to the fine mapping graph K Essence (x):
J Essence (x)＝K Essence (x)×I'(x)-K Essence (x)+1;
Step 302, using an open-source non-fog road vehicle target detection data set as a defogging truth value label, calculating a Loss value between a coarse defogging image and a corresponding defogging truth value label by adopting an L1 Loss function, marking as Loss 1, calculating a Loss value between a fine defogging image and a corresponding defogging truth value label by adopting an L2 Loss function, marking as Loss 2, calculating a total Loss value Loss corresponding to a lightweight defogging module, namely loss=0.4 Loss 1+0.6loss2, and performing reverse propagation training on the lightweight defogging module by adopting a gradient descent method according to the total Loss value Loss until the total Loss value is converged, ending training, and storing the current lightweight defogging module.
Specifically, the simplified atmospheric scattering model can be expressed as:
I(x)=J(x)t(x)+A(1-t(x));
Wherein I (x) represents a foggy image, J (x) represents a foggy image, a represents atmospheric light intensity, t (x) represents a transmission map, and x represents an image pixel.
Defogging images can be obtained according to an atmospheric scattering model:
combining the parameters t (x) and a to be predicted into the unknown K (x), the formula can be transformed into:
J(x)=K(x)I(x)-K(x)+1;
Calculating a mapping parameter K (x) of the foggy day image, and defogging the foggy day image by combining the images.
Further, the specific mode of the step 4 is as follows:
Step 401, freezing parameters of a lightweight defogging module, inputting images in an integral network training set into a real-time foggy-day road vehicle target detection model, outputting a refined defogging image to a super-parameter prediction module and a conductive image processing module by the lightweight defogging module, and outputting 15 super-parameters in the conductive image processing module and 6 weight parameters in a gating module by the super-parameter prediction module;
the guided image processing module obtains the fine defogging images after exposure degree adjustment, white balance, gamma conversion, contrast enhancement, sharpening and color adjustment according to the 15 super parameters output by the super parameter prediction module and the fine defogging images output by the lightweight defogging module;
The gating module performs weighted fusion on the refined defogging images after exposure degree adjustment, white balance, gamma conversion, contrast enhancement, sharpening and color adjustment output by the conductive image processing module according to the 6 weight parameters output by the super parameter prediction module to obtain corresponding self-adaptive enhancement images;
The self-adaptive enhanced image is input into a target detection network YOLOv, the class and the coordinate of a vehicle target corresponding to the image in the whole network training set are output, the class and the coordinate of the vehicle target contained in the image in the whole network training set are used as target detection truth value labels, the loss value between the class and the coordinate of the vehicle target and the target detection truth value labels is calculated and output according to the loss function of the target detection network YOLOv, the super-parameter prediction module and the target detection network YOLOv5 are subjected to reverse propagation training according to the loss value by adopting a gradient descent method until the loss value converges, the training is ended, and the current real-time foggy road vehicle target detection model is stored.
Specifically, as shown in fig. 5,6 images on the left side of fig. 5 are visual result diagrams for detecting the targets of the foggy road vehicles under different concentration foggy days, and 6 images on the right side of fig. 5 are visual result diagrams for detecting the targets of the foggy road vehicles under different concentration foggy days.
In a word, the invention can utilize the lightweight defogging module to defog the foggy road vehicle image rapidly, recover the overall detail information of the foggy road image, and guide the lightweight self-adaptive target feature enhancement module by adopting the target detection task to adaptively enhance the effective features of the foggy road vehicle target, thereby greatly improving the detection performance of the foggy road vehicle target. Compared with the prior art, the method has the advantages of better detection effect and stronger robustness on the target of the foggy road vehicle, meets the real-time requirement, can be applied to intelligent traffic, and provides a new solution for the target detection of the foggy road vehicle.
Those skilled in the art will recognize that the embodiments described are for the purpose of aiding the reader in understanding the principles of the invention and should be understood to be not limited to the embodiments described. Various modifications and variations of the present invention will be apparent to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims (9)

1.自适应目标特征增强的实时雾天道路车辆目标检测方法，其特征在于，包括以下步骤；1. A real-time foggy road vehicle target detection method with adaptive target feature enhancement, characterized in that it comprises the following steps; 步骤1，根据开源的无雾道路车辆目标检测数据集合成不同浓度雾天的对应图像目标检测数据集，根据开源的无雾道路车辆目标检测数据集、合成不同浓度雾天的对应图像目标检测数据集以及开源的真实雾天道路车辆目标检测数据集组成去雾网络训练集以及整体网络训练集；Step 1: synthesize the corresponding image target detection datasets of foggy days with different concentrations based on the open source fog-free road vehicle target detection dataset, and form the defogging network training set and the overall network training set based on the open source fog-free road vehicle target detection dataset, the synthesized corresponding image target detection datasets of foggy days with different concentrations, and the open source real foggy road vehicle target detection dataset; 步骤2，构建实时雾天道路车辆目标检测模型，所述实时雾天道路车辆目标检测模型包括轻量级去雾模块、轻量级自适应目标特征增强模块以及目标检测网络YOLOv5；Step 2, constructing a real-time foggy road vehicle target detection model, wherein the real-time foggy road vehicle target detection model includes a lightweight defogging module, a lightweight adaptive target feature enhancement module, and a target detection network YOLOv5; 步骤3，利用去雾网络训练集对轻量级去雾模块进行训练，轻量级去雾模块的输入为去雾网络训练集中的图像，输出为对应的精去雾图像，训练结束后保存当前轻量级去雾模块；Step 3, use the defogging network training set to train the lightweight defogging module, the input of the lightweight defogging module is the image in the defogging network training set, and the output is the corresponding refined defogging image. After the training is completed, the current lightweight defogging module is saved; 步骤4，冻结轻量级去雾模块参数，利用整体网络训练集对实时雾天道路车辆目标检测模型进行端对端训练，实时雾天道路车辆目标检测模型的输入为整体网络训练集中的图像，输出为对应的车辆目标的类别和坐标，训练结束后保存当前实时雾天道路车辆目标检测模型；Step 4, freeze the parameters of the lightweight defogging module, use the overall network training set to perform end-to-end training on the real-time foggy road vehicle target detection model, the input of the real-time foggy road vehicle target detection model is the image in the overall network training set, and the output is the category and coordinates of the corresponding vehicle target. After the training is completed, save the current real-time foggy road vehicle target detection model; 步骤5，将雾天道路车辆图像输入到训练好的实时雾天道路车辆目标检测模型中，得到雾天道路车辆图像中的车辆目标的类别和坐标，完成自适应目标特征增强的实时雾天道路车辆目标检测。Step 5, input the foggy road vehicle image into the trained real-time foggy road vehicle target detection model, obtain the category and coordinates of the vehicle target in the foggy road vehicle image, and complete the real-time foggy road vehicle target detection with adaptive target feature enhancement. 2.根据权利要求1所述的自适应目标特征增强的实时雾天道路车辆目标检测方法，其特征在于，步骤1的具体方式为：2. The real-time foggy road vehicle target detection method with adaptive target feature enhancement according to claim 1 is characterized in that the specific manner of step 1 is: 步骤101，根据开源的无雾道路车辆目标检测数据集，合成不同浓度雾天的对应图像目标检测数据集：Step 101, based on the open-source fog-free road vehicle target detection dataset, synthesize the corresponding image target detection datasets of foggy days with different concentrations: I(x)＝J(x)e-β(λ)d(x)+A(1-e-β(λ)d(x))；I(x)=J(x)e -β(λ)d(x) +A(1-e -β(λ)d(x) ); 其中，x表示图像的像素位置，J(x)表示开源的无雾道路车辆目标检测数据集中的图像，I(x)表示对应的合成雾天的对应图像，β(λ)表示大气散射系数，d(x)表示通过单目图像深度估计算法MiDaS估算得到的与J(x)对应的景深图，A表示无穷远处的大气光；通过调整β(λ)与A的值，合成不同浓度雾天的对应图像目标检测数据集；Where x represents the pixel position of the image, J(x) represents the image in the open-source fog-free road vehicle target detection dataset, I(x) represents the corresponding image of the synthetic foggy day, β(λ) represents the atmospheric scattering coefficient, d(x) represents the depth map corresponding to J(x) estimated by the monocular image depth estimation algorithm MiDaS, and A represents the atmospheric light at infinity. By adjusting the values of β(λ) and A, the corresponding image target detection datasets of foggy days with different concentrations are synthesized. 步骤102，将开源的无雾道路车辆目标检测数据集、合成不同浓度雾天的对应图像目标检测数据集组成去雾网络训练集；将开源的无雾道路车辆目标检测数据集、合成不同浓度雾天的对应图像目标检测数据集以及开源的真实雾天道路车辆目标检测数据集组成整体网络训练集。Step 102, the open source fog-free road vehicle target detection dataset and the corresponding image target detection dataset of synthetic fog with different concentrations are combined into a defogging network training set; the open source fog-free road vehicle target detection dataset, the corresponding image target detection dataset of synthetic fog with different concentrations and the open source real foggy road vehicle target detection dataset are combined into an overall network training set. 3.根据权利要求1所述的自适应目标特征增强的实时雾天道路车辆目标检测方法，其特征在于，步骤2中的轻量级去雾模块包括第一级编码器、第一级解码器、第二级编码器、第二级解码器以及图像级联层；3. The real-time foggy road vehicle target detection method with adaptive target feature enhancement according to claim 1, characterized in that the lightweight defogging module in step 2 comprises a first-stage encoder, a first-stage decoder, a second-stage encoder, a second-stage decoder and an image cascade layer; 其中，第一级编码器包括依次连接的1个步长为1输出通道数为8的1×1卷积层、第一Relu激活函数层、2个步长为1输出通道数为8的3×3卷积层、第二Relu激活函数层、1个步长为2窗口大小为3×3的第一最大池化层、2个步长为1输出通道数为16的3×3卷积层、第三Relu激活函数层、1个步长为2窗口大小为3×3的第二最大池化层、1个步长为1输出通道数为32的3×3卷积层以及第四Relu激活函数层；Among them, the first-level encoder includes a 1×1 convolution layer with a stride of 1 and an output channel number of 8, a first Relu activation function layer, two 3×3 convolution layers with a stride of 1 and an output channel number of 8, a second Relu activation function layer, a first maximum pooling layer with a stride of 2 and a window size of 3×3, two 3×3 convolution layers with a stride of 1 and an output channel number of 16, a third Relu activation function layer, a second maximum pooling layer with a stride of 2 and a window size of 3×3, a 3×3 convolution layer with a stride of 1 and an output channel number of 32, and a fourth Relu activation function layer. 第一级解码器包括依次连接的1个步长为1输出通道数为32的3×3卷积层、第五Relu激活函数层、第一级联层、1个步长为1输出通道数为16的3×3卷积层、第六Relu激活函数层、1个步长为2的第一上采样层、第二级联层、1个步长为1输出通道数为8的3×3卷积层、第七Relu激活函数层、1个步长为2的第二上采样层、1个步长为1输出通道为3的3×3卷积层以及第八Relu激活函数层；The first-stage decoder includes a 3×3 convolution layer with a step size of 1 and an output channel number of 32, a fifth Relu activation function layer, a first cascade layer, a 3×3 convolution layer with a step size of 1 and an output channel number of 16, a sixth Relu activation function layer, a first upsampling layer with a step size of 2, a second cascade layer, a 3×3 convolution layer with a step size of 1 and an output channel number of 8, a seventh Relu activation function layer, a second upsampling layer with a step size of 2, a 3×3 convolution layer with a step size of 1 and an output channel number of 3, and an eighth Relu activation function layer. 其中，第一级编码器中第四Relu激活函数层的输出即为第一级解码器中步长为1输出通道数为32的3×3卷积层的输入；第一级解码器中的第一级联层用于将第一级编码器中第二最大池化层的输出与第一级解码器中第五Relu激活函数层的输出进行级联，第一级解码器中的第二级联层用于将第一级编码器中第一最大池化层的输出与第一级解码器中第一上采样层的输出进行级联；Among them, the output of the fourth Relu activation function layer in the first-stage encoder is the input of the 3×3 convolution layer with a step size of 1 and an output channel number of 32 in the first-stage decoder; the first cascade layer in the first-stage decoder is used to cascade the output of the second maximum pooling layer in the first-stage encoder with the output of the fifth Relu activation function layer in the first-stage decoder, and the second cascade layer in the first-stage decoder is used to cascade the output of the first maximum pooling layer in the first-stage encoder with the output of the first upsampling layer in the first-stage decoder; 第二级编码器包括依次连接的1个步长为1输出通道数为8的1×1卷积层、第九Relu激活函数层、2个步长为1输出通道数为8的3×3卷积层、第十Relu激活函数层、1个步长为2窗口大小为3×3的第三最大池化层、第三级联层、2个步长为1输出通道数为16的3×3卷积层、第十一Relu激活函数层、1个步长为2窗口大小为3×3的第四最大池化层、第四级联层、1个步长为1输出通道数为32的3×3卷积层以及第十二Relu激活函数层；The second-stage encoder includes a 1×1 convolution layer with a stride of 1 and an output channel number of 8, a ninth Relu activation function layer, two 3×3 convolution layers with a stride of 1 and an output channel number of 8, a tenth Relu activation function layer, a third maximum pooling layer with a stride of 2 and a window size of 3×3, a third cascade layer, two 3×3 convolution layers with a stride of 1 and an output channel number of 16, an eleventh Relu activation function layer, a fourth maximum pooling layer with a stride of 2 and a window size of 3×3, a fourth cascade layer, a 3×3 convolution layer with a stride of 1 and an output channel number of 32, and a twelfth Relu activation function layer. 其中，第二级编码器中的第三级联层用于将第一级编码器中第一最大池化层的输出与第二级编码器中的第三最大池化层的输出进行级联，第二级编码器中的第四级联层用于将第一级编码器中第二最大池化层的输出与第二级编码器中的第四最大池化层的输出进行级联；The third cascade layer in the second-stage encoder is used to cascade the output of the first maximum pooling layer in the first-stage encoder with the output of the third maximum pooling layer in the second-stage encoder, and the fourth cascade layer in the second-stage encoder is used to cascade the output of the second maximum pooling layer in the first-stage encoder with the output of the fourth maximum pooling layer in the second-stage encoder; 第二级解码器包括依次连接的1个步长为1输出通道数为32的3×3卷积层、第十三Relu激活函数层、第五级联层、1个步长为1输出通道数为16的3×3卷积层、第十四Relu激活函数层、1个步长为2的第三上采样层、第六级联层、1个步长为1输出通道数为8的3×3卷积层、第十五Relu激活函数层、1个步长为2的第四上采样层、2个步长为1输出通道为3的3×3卷积层以及第十六Relu激活函数层；The second stage decoder includes a 3×3 convolution layer with a step size of 1 and an output channel number of 32, a thirteenth Relu activation function layer, a fifth cascade layer, a 3×3 convolution layer with a step size of 1 and an output channel number of 16, a fourteenth Relu activation function layer, a third upsampling layer with a step size of 2, a sixth cascade layer, a 3×3 convolution layer with a step size of 1 and an output channel number of 8, a fifteenth Relu activation function layer, a fourth upsampling layer with a step size of 2, two 3×3 convolution layers with a step size of 1 and an output channel number of 3, and a sixteenth Relu activation function layer. 其中，第二级编码器中第十二Relu激活函数层的输出即为第二级解码器中步长为1输出通道数为32的3×3卷积层的输入；第二级解码器中的第五级联层用于将第一级解码器中第五Relu激活函数层的输出、第二级编码器中第四最大池化层的输出、第二级解码器中第十三Relu激活函数层的输出进行级联，第二级解码器中的第六级联层用于将第一级解码器中第一上采样层的输出、第二级编码器中第三最大池化层的输出、第二级解码器中第三上采样层的输出进行级联。Among them, the output of the twelfth Relu activation function layer in the second-level encoder is the input of the 3×3 convolutional layer with a step size of 1 and an output channel number of 32 in the second-level decoder; the fifth cascade layer in the second-level decoder is used to cascade the output of the fifth Relu activation function layer in the first-level decoder, the output of the fourth maximum pooling layer in the second-level encoder, and the output of the thirteenth Relu activation function layer in the second-level decoder; the sixth cascade layer in the second-level decoder is used to cascade the output of the first upsampling layer in the first-level decoder, the output of the third maximum pooling layer in the second-level encoder, and the output of the third upsampling layer in the second-level decoder. 4.根据权利要求3所述的自适应目标特征增强的实时雾天道路车辆目标检测方法，其特征在于，步骤2中的轻量级自适应目标特征增强模块包括超参数预测模块，可导图像处理模块以及门控模块。4. The real-time foggy road vehicle target detection method with adaptive target feature enhancement according to claim 3 is characterized in that the lightweight adaptive target feature enhancement module in step 2 includes a hyperparameter prediction module, a guideable image processing module and a gating module. 5.根据权利要求4所述的自适应目标特征增强的实时雾天道路车辆目标检测方法，其特征在于，所述超参数预测模块包括依次连接的1个步长为1输出通道数为16的3×3卷积层、第十七Relu激活函数层、1个步长为2窗口大小为3×3的第五最大池化层、1个步长为1输出通道数为32的3×3卷积层、第十八Relu激活函数层、1个步长为2窗口大小为3×3的第六最大池化层、1个步长为1输出通道数为32的3×3卷积层、第十九Relu激活函数层、1个步长为2窗口大小为3×3的第七最大池化层、1个步长为1输出通道数为64的3×3卷积层、第二十Relu激活函数层、1个步长为2窗口大小为3×3的第八最大池化层、1个步长为1输出通道数为64的3×3卷积层、第二十一Relu激活函数层、1个步长为2窗口大小为3×3的第九最大池化层、1个步长为1输出通道数为128的3×3卷积层、第二十二Relu激活函数层、1个全局平均池化层、1个输出神经元个数为256的第一全连接层以及2个并行的第二全连接层与第三全连接层；所述第二全连接层的输出神经元个数为15，第三全连接层的输出神经元个数为6；其中，步长为1输出通道数为16的3×3卷积层的输入为轻量级去雾模块输出的精去雾图像，第二全连接层的输出为可导图像处理模块中的15个超参数，第三全连接层的输出为门控模块中的6个权重参数。5. The real-time foggy road vehicle target detection method with adaptive target feature enhancement according to claim 4 is characterized in that the hyperparameter prediction module includes a 3×3 convolution layer with a step size of 1 and an output channel number of 16, a seventeenth Relu activation function layer, a fifth maximum pooling layer with a step size of 2 and a window size of 3×3, a 3×3 convolution layer with a step size of 1 and an output channel number of 32, an eighteenth Relu activation function layer, a sixth maximum pooling layer with a step size of 2 and a window size of 3×3, a 3×3 convolution layer with a step size of 1 and an output channel number of 32, a nineteenth Relu activation function layer, a seventh maximum pooling layer with a step size of 2 and a window size of 3×3, a 3×3 convolution layer with a step size of 1 and an output channel number of 64, a twentieth Relu activation function layer, a step size of 2 and a window size of 3×3 The eighth maximum pooling layer, a 3×3 convolution layer with a stride of 1 and an output channel number of 64, the twenty-first Relu activation function layer, a ninth maximum pooling layer with a stride of 2 and a window size of 3×3, a 3×3 convolution layer with a stride of 1 and an output channel number of 128, the twenty-second Relu activation function layer, a global average pooling layer, a first fully connected layer with an output number of 256 neurons, and two parallel second fully connected layers and a third fully connected layer; the number of output neurons of the second fully connected layer is 15, and the number of output neurons of the third fully connected layer is 6; wherein, the input of the 3×3 convolution layer with a stride of 1 and an output channel number of 16 is the refined dehazed image output by the lightweight dehazing module, the output of the second fully connected layer is the 15 hyperparameters in the derivable image processing module, and the output of the third fully connected layer is the 6 weight parameters in the gating module. 6.根据权利要求5所述的自适应目标特征增强的实时雾天道路车辆目标检测方法，其特征在于，所述可导图像处理模块包括曝光度调节、白平衡、伽马变换、对比度增强、锐化、颜色调节六个无序可导图像处理器；可导图像处理模块接收轻量级去雾模块输出的精去雾图像以及超参数预测模块输出的15个超参数，输出对应处理后的图像；6. The real-time foggy road vehicle target detection method with adaptive target feature enhancement according to claim 5 is characterized in that the guideable image processing module includes six disordered guideable image processors for exposure adjustment, white balance, gamma transformation, contrast enhancement, sharpening, and color adjustment; the guideable image processing module receives the refined defogging image output by the lightweight defogging module and 15 hyperparameters output by the hyperparameter prediction module, and outputs the corresponding processed image; 所述曝光度调节用以下映射函数表示：The exposure adjustment is represented by the following mapping function: Jexposure(x)＝2E×J精(x)； Jexposure (x) = 2 E × Jfine (x); 其中，J精(x)表示轻量级去雾模块输出的精去雾图像，Jexposure(x)为曝光度调节后的精去雾图像，E为曝光度调节超参数；Where Jfine (x) represents the refined dehazed image output by the lightweight dehazing module, Jexposure (x) represents the refined dehazed image after exposure adjustment, and E represents the exposure adjustment hyperparameter. 所述白平衡用以下映射函数表示：The white balance is represented by the following mapping function: Jwb(x)＝(Wrr(x),Wgg(x),Wbb(x))；J wb (x) = (W r r (x), W g g (x), W b b (x)); 其中，Jwb(x)＝为白平衡后的精去雾图像，Wr，Wg，Wb分别表示r，g，b颜色通道的白平衡超参数，r(x)，g(x)，b(x)分别表示精去雾图像J精(x)在r，g，b颜色通道的像素值；Wherein, Jwb (x)= is the refined dehazed image after white balance, Wr , Wg , Wb represent the white balance hyperparameters of the r, g, b color channels respectively, r(x), g(x), b(x) represent the pixel values of the refined dehazed image Jprecision (x) in the r, g, b color channels respectively; 所述伽马变换用以下映射函数表示：The gamma transform is represented by the following mapping function: 其中，Jgamma(x)为伽马变换后的精去雾图像，γ表示伽马变换超参数；Where J gamma (x) is the refined dehazed image after gamma transformation, and γ represents the gamma transformation hyperparameter; 所述对比度增强用以下映射函数表示：The contrast enhancement is represented by the following mapping function: Jcontrast(x)＝α×En(J精(x))+(1-α)×J精(x)；J contrast (x) = α × En (J fine (x)) + (1-α) × J fine (x); 其中，Jcontrast(x)为对比度增强后的精去雾图像，α表示对比度增强超参数；Where J contrast (x) is the refined dehazed image after contrast enhancement, and α represents the contrast enhancement hyperparameter; Lum(J精(x))＝0.3r(x)+0.59g(x)+0.11b(x)；Lum( J (x))=0.3r(x)+0.59g(x)+0.11b(x); Sigmoid()为非线性激活函数，Lum()表示像素亮度函数，En()表示图像亮度增强函数；Sigmoid() is a nonlinear activation function, Lum() represents the pixel brightness function, and En() represents the image brightness enhancement function; 所述锐化用以下映射函数表示：The sharpening is represented by the following mapping function: Jsharpen(x)＝J精(x)+λ1(J精(x)-Gaussian(J精(x)))；J sharpen (x) = J fine (x) + λ 1 (J fine (x) - Gaussian (J fine (x))); 其中，Jsharpen(x)为锐化后的精去雾图像，λ1为锐化超参数，Gaussian()为高斯滤波函数；Wherein, J sharpen (x) is the sharpened dehazed image, λ 1 is the sharpening hyperparameter, and Gaussian() is the Gaussian filter function; 所述颜色调节用以下映射函数表示：The color adjustment is represented by the following mapping function: 其中，Jtone(x)为颜色调节后的精去雾图像，ti为第i+1个颜色调节超参数，i＝0,1,2,......,7，clip(y,0,1)代表截断函数； Wherein, J tone (x) is the refined dehazed image after color adjustment, ti is the i+1th color adjustment hyperparameter, i=0,1,2,...,7, clip(y,0,1) represents the truncation function; 7.根据权利要求6所述的自适应目标特征增强的实时雾天道路车辆目标检测方法，其特征在于，所述门控模块接收超参数预测模块输出的6个权重参数，用于对六个无序可导图像处理器处理后的图像进行自适应加权融合，得到自适应增强图像J强(x)：7. The real-time foggy road vehicle target detection method with adaptive target feature enhancement according to claim 6 is characterized in that the gating module receives the six weight parameters output by the hyperparameter prediction module, and is used to perform adaptive weighted fusion on the images processed by the six disordered derivable image processors to obtain an adaptive enhanced image J strength (x): J强(x)＝w1Jexposure(x)+w2Jwb(x)+w3Jgamma(x)+w4Jcontrast(x)+w5Jsharpen(x)+w6Jtone(x)；J strong (x)＝w 1 J exposure (x)+w 2 J wb (x)+w 3 J gamma (x)+w 4 J contrast (x)+w 5 J sharpen (x)+w 6 J tone (x); 其中，w1为曝光度调节权重参数、w2为白平衡权重参数、w3为伽马变换权重参数、w4为对比度增强权重参数、w5为锐化权重参数、w6为颜色调节权重参数。Among them, w1 is the exposure adjustment weight parameter, w2 is the white balance weight parameter, w3 is the gamma transformation weight parameter, w4 is the contrast enhancement weight parameter, w5 is the sharpening weight parameter, and w6 is the color adjustment weight parameter. 8.根据权利要求7所述的自适应目标特征增强的实时雾天道路车辆目标检测方法，其特征在于，步骤3的具体方式为：8. The real-time foggy road vehicle target detection method with adaptive target feature enhancement according to claim 7, characterized in that the specific manner of step 3 is: 步骤301，将去雾网络训练集中的图像输入至轻量级去雾模块中第一级编码器的步长为1输出通道数为8的1×1卷积层中，轻量级去雾模块的第一级解码器的第八Relu激活函数层输出通道数为3的粗映射图K粗(x)，根据粗映射图K粗(x)计算去雾网络训练集中的图像所对应的粗去雾图像：Step 301, input the image in the defogging network training set into the 1×1 convolution layer with a step size of 1 and an output channel number of 8 of the first-level encoder in the lightweight defogging module, and the eighth Relu activation function layer of the first-level decoder of the lightweight defogging module outputs a coarse mapping map Kcoarse (x) with a channel number of 3. The coarse defogging image corresponding to the image in the defogging network training set is calculated according to the coarse mapping map Kcoarse (x): J粗(x)＝K粗(x)×I'(x)-K粗(x)+1； Jcoarse (x)= Kcoarse (x)×I'(x) -Kcoarse (x)+1; 其中，I'(x)表示去雾网络训练集中的图像，I'(x)＝I(x)或J(x)，且I'(x)为通道数为3的RGB图像；J粗(x)即为粗去雾图像；Where I'(x) represents the image in the dehazing network training set, I'(x) = I(x) or J(x), and I'(x) is an RGB image with 3 channels; Jcoarse (x) is the coarse dehazed image; 根据轻量级去雾模块的图像级联层将去雾网络训练集中的图像与第一级解码器输出的对应粗去雾图像进行级联，得到通道数为6的级联图像，所述级联图像输入至轻量级去雾模块中第二级编码器的步长为1输出通道数为8的1×1卷积层中，轻量级去雾模块的第二级解码器的第十六Relu激活函数层输出通道数为3的精映射图K精(x)；根据精映射图K精(x)计算去雾网络训练集中的图像所对应的精去雾图像：According to the image cascade layer of the lightweight defogging module, the image in the defogging network training set is cascaded with the corresponding coarse defogging image output by the first-level decoder to obtain a cascaded image with 6 channels. The cascaded image is input into the 1×1 convolution layer with a step size of 1 and an output channel number of 8 of the second-level encoder in the lightweight defogging module. The sixteenth Relu activation function layer of the second-level decoder of the lightweight defogging module outputs a fine map Kfine (x) with a channel number of 3; the fine defogging image corresponding to the image in the defogging network training set is calculated according to the fine map Kfine (x): J精(x)＝K精(x)×I'(x)-K精(x)+1；J精(x)＝K精(x)×I'(x)-K精(x)+1; 步骤302，将开源的无雾道路车辆目标检测数据集作为去雾真值标签，采用L1损失函数计算粗去雾图像以及对应的去雾真值标签之间的损失值，记为loss1，采用L2损失函数计算精去雾图像以及对应的去雾真值标签之间的损失值，记为loss2，计算轻量级去雾模块对应的总损失值Loss：Step 302: Use the open-source fog-free road vehicle target detection dataset as the defogging true value label, use the L1 loss function to calculate the loss value between the coarse defogging image and the corresponding defogging true value label, recorded as loss 1 , use the L2 loss function to calculate the loss value between the fine defogging image and the corresponding defogging true value label, recorded as loss 2 , and calculate the total loss value Loss corresponding to the lightweight defogging module: Loss＝0.4loss1+0.6loss2；根据总损失值Loss，采用梯度下降法对轻量级去雾模块进行反向传播训练，直至总损失值收敛，结束训练，并保存当前轻量级去雾模块。Loss = 0.4loss 1 + 0.6loss 2 ; According to the total loss value Loss, the gradient descent method is used to perform back propagation training on the lightweight defogging module until the total loss value converges, the training is terminated, and the current lightweight defogging module is saved. 9.根据权利要求8所述的自适应目标特征增强的实时雾天道路车辆目标检测方法，其特征在于，步骤4的具体方式为：9. The real-time foggy road vehicle target detection method with adaptive target feature enhancement according to claim 8, characterized in that the specific manner of step 4 is: 步骤401，冻结轻量级去雾模块参数，将整体网络训练集中的图像输入至实时雾天道路车辆目标检测模型中，轻量级去雾模块输出精去雾图像至超参数预测模块以及可导图像处理模块，超参数预测模块输出可导图像处理模块中的15个超参数以及门控模块中的6个权重参数；Step 401, freeze the parameters of the lightweight defogging module, input the images in the overall network training set into the real-time foggy road vehicle target detection model, the lightweight defogging module outputs the refined defogging image to the hyperparameter prediction module and the derivable image processing module, and the hyperparameter prediction module outputs 15 hyperparameters in the derivable image processing module and 6 weight parameters in the gating module; 可导图像处理模块根据超参数预测模块输出的15个超参数以及轻量级去雾模块输出的精去雾图像，得到曝光度调节、白平衡、伽马变换、对比度增强、锐化、颜色调节后的精去雾图像；The guideable image processing module obtains the refined defogging image after exposure adjustment, white balance, gamma transformation, contrast enhancement, sharpening and color adjustment according to the 15 hyperparameters output by the hyperparameter prediction module and the refined defogging image output by the lightweight defogging module; 门控模块根据超参数预测模块输出的6个权重参数，对可导图像处理模块输出的曝光度调节、白平衡、伽马变换、对比度增强、锐化、颜色调节后的精去雾图像进行加权融合，得到对应的自适应增强图像；The gating module performs weighted fusion on the refined dehazed image after exposure adjustment, white balance, gamma transformation, contrast enhancement, sharpening, and color adjustment output by the guideable image processing module according to the six weight parameters output by the hyperparameter prediction module to obtain the corresponding adaptive enhanced image; 自适应增强图像输入至目标检测网络YOLOv5中，输出整体网络训练集中的图像所对应的车辆目标的类别和坐标，将整体网络训练集中的图像包含的车辆目标的类别和坐标作为目标检测真值标签，根据目标检测网络YOLOv5自身的损失函数，计算输出车辆目标的类别和坐标与目标检测真值标签之间的损失值，根据损失值，采用梯度下降法对超参数预测模块以及目标检测网络YOLOv5进行反向传播训练，直至损失值收敛，结束训练，并保存当前实时雾天道路车辆目标检测模型。The adaptive enhanced image is input into the target detection network YOLOv5, and the category and coordinates of the vehicle target corresponding to the image in the overall network training set are output. The category and coordinates of the vehicle target contained in the image in the overall network training set are used as the target detection true value label. According to the loss function of the target detection network YOLOv5 itself, the loss value between the output vehicle target category and coordinates and the target detection true value label is calculated. According to the loss value, the gradient descent method is used to perform back propagation training on the hyperparameter prediction module and the target detection network YOLOv5 until the loss value converges, the training is terminated, and the current real-time foggy road vehicle target detection model is saved.
CN202411732235.XA 2024-11-29 2024-11-29 Real-time foggy-day road vehicle target detection method with enhanced self-adaptive target characteristics Active CN119851242B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202411732235.XA CN119851242B (en) 2024-11-29 2024-11-29 Real-time foggy-day road vehicle target detection method with enhanced self-adaptive target characteristics

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202411732235.XA CN119851242B (en) 2024-11-29 2024-11-29 Real-time foggy-day road vehicle target detection method with enhanced self-adaptive target characteristics

Publications (2)

Publication Number Publication Date
CN119851242A true CN119851242A (en) 2025-04-18
CN119851242B CN119851242B (en) 2025-10-14

Family

ID=95358795

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202411732235.XA Active CN119851242B (en) 2024-11-29 2024-11-29 Real-time foggy-day road vehicle target detection method with enhanced self-adaptive target characteristics

Country Status (1)

Country Link
CN (1) CN119851242B (en)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5502245A (en) * 1995-06-07 1996-03-26 Twenty-First Century Research Corporation Methods of making intermediate oxidation products by controlling transient conversion in an atomized liquid
CN115661777A (en) * 2022-11-03 2023-01-31 西安邮电大学 Semantic-combined foggy road target detection algorithm
CN116681603A (en) * 2023-04-21 2023-09-01 中国科学技术大学 Image defogging method, device, electronic equipment and computer storage medium
CN116824542A (en) * 2023-06-13 2023-09-29 重庆市荣冠科技有限公司 Light-weight foggy-day vehicle detection method based on deep learning
CN116912569A (en) * 2023-07-11 2023-10-20 南京理工大学 An improved Yolov5 dead wood detection method combined with defogging algorithm
CN117115617A (en) * 2023-08-04 2023-11-24 三峡大学 A method for identifying small target faults in foggy weather scenes of transmission lines
CN118864325A (en) * 2024-07-29 2024-10-29 华南师范大学 An end-to-end rapid defogging detection joint method and device

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5502245A (en) * 1995-06-07 1996-03-26 Twenty-First Century Research Corporation Methods of making intermediate oxidation products by controlling transient conversion in an atomized liquid
CN115661777A (en) * 2022-11-03 2023-01-31 西安邮电大学 Semantic-combined foggy road target detection algorithm
CN116681603A (en) * 2023-04-21 2023-09-01 中国科学技术大学 Image defogging method, device, electronic equipment and computer storage medium
CN116824542A (en) * 2023-06-13 2023-09-29 重庆市荣冠科技有限公司 Light-weight foggy-day vehicle detection method based on deep learning
CN116912569A (en) * 2023-07-11 2023-10-20 南京理工大学 An improved Yolov5 dead wood detection method combined with defogging algorithm
CN117115617A (en) * 2023-08-04 2023-11-24 三峡大学 A method for identifying small target faults in foggy weather scenes of transmission lines
CN118864325A (en) * 2024-07-29 2024-10-29 华南师范大学 An end-to-end rapid defogging detection joint method and device

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
殷旭平，钟平，薛伟，肖子轩，李广: "通过风格迁移的浓雾天气条件下无人机图像目标检测方法", 《航空兵器》, vol. 28, no. 3, 14 January 2021 (2021-01-14), pages 22 *

Also Published As

Publication number Publication date
CN119851242B (en) 2025-10-14

Similar Documents

Publication Publication Date Title
CN113313657B (en) An unsupervised learning method and system for low-light image enhancement
Li et al. Single image dehazing via conditional generative adversarial network
Yin et al. A novel image-dehazing network with a parallel attention block
CN114693924B (en) A road scene semantic segmentation method based on multi-model fusion
CN108765336B (en) Image dehazing method based on dark and bright primary color prior and adaptive parameter optimization
Cheng et al. A highway traffic image enhancement algorithm based on improved GAN in complex weather conditions
CN111539246B (en) Cross-spectrum face recognition method and device, electronic equipment and storage medium thereof
CN106157267A (en) A kind of image mist elimination absorbance optimization method based on dark channel prior
CN113947538B (en) A multi-scale efficient convolutional self-attention single image rain removal method
CN116188307A (en) All-day image low-light enhancement and semantic segmentation joint method and system
CN113420794A (en) Binaryzation Faster R-CNN citrus disease and pest identification method based on deep learning
CN115409738B (en) Method, system, device and storage medium for defogging video images of patrol river
CN108921887B (en) Underwater scene depth map estimation method based on underwater light attenuation priori
CN113362375A (en) Moving object detection method for vehicle
CN118505581A (en) Image defogging method based on multi-scale dense feature fusion and gate control jump connection
Li et al. Progressive encoding-decoding image dehazing network
CN103778605A (en) Greasy weather image enhancement method
CN116704284A (en) An Enhanced Refinement Network and Method for Target Detection in Bad Weather
CN119851242B (en) Real-time foggy-day road vehicle target detection method with enhanced self-adaptive target characteristics
CN106683055A (en) Degradation model and group sparse representation-based foggy day image restoration method
CN117808707B (en) Multi-scale image defogging method, system, device and storage medium
Lin et al. WeaGAN: Generative adversarial network for weather translation of image among multi-domain
Jing et al. An efficient dehazing method using pixel unshuffle and color correction
CN119888656A (en) Self-supervision blind image decomposition method oriented to automatic driving scene
CN117253230B (en) An end-to-end domain-adaptive semantic segmentation method for infrared images

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant