CN118505974A - Sonar image target detection method and system based on deep learning - Google Patents

Abstract

The invention discloses a sonar image target detection method and system based on deep learning, and belongs to the technical field of underwater recognition. The method comprises the following steps: acquiring a sonar image, and screening and arranging the sonar image to form a sonar image data set; constructing a sonar image target detection model YOLO_EC, introducing an ELAN module to replace a part of C3 modules on the basis of YOLOv networks for feature extraction, and introducing an MP module to replace a convolution layer for downsampling; introducing a C2f module into the feature fusion network to perform feature fusion; pre-training the model by using an optical image data set through a mode of style conversion and migration learning, and training by using a sonar image data set; and inputting the sonar image to be detected into a trained detection model to obtain a sonar image target detection result. According to the invention, the feature extraction capability and the parallel computing capability of the network structure enhancement model are improved, and the detection precision of the sonar image target by the migration learning enhancement model is utilized.

Description

A sonar image target detection method and system based on deep learning

Technical Field

The present invention relates to the technical field of underwater target recognition, and in particular to a sonar image target detection method and system based on deep learning.

Background Art

Underwater target detection technology can be divided into detection technology based on sonar echo and detection technology based on sonar image according to the working principle. Detection technology based on sonar echo processes echo signals and analyzes and detects seabed information by analyzing the characteristics of echo signals. Target detection based on sonar images has the characteristics of strong expression ability, high precision, high resolution, wider application, multi-beam and strong real-time performance. Although target detection technology based on sonar echo still has its importance in underwater detection, target detection technology based on sonar images has more advantages and development potential, and will be more widely used in underwater detection in the future.

The imaging sonars currently used for underwater detection mainly include side scan sonar, synthetic aperture sonar, multi-beam sonar and forward-looking sonar. Multi-beam sonar imaging is mostly three-dimensional sound images. The generated images are generally layered color topographic maps after processing. The imaging of the terrain is relatively intuitive. This sonar is often used in tasks involving terrain detection and buried object detection. Synthetic aperture sonar (SAS) is a high-resolution imaging system. Its principle is that when the transducer sails in a given direction, it interferes and synthesizes the echo data sampled in a continuous time series in an orderly manner, which is equivalent to using a smaller physical aperture to move to form a larger virtual aperture, thereby improving the azimuth resolution of the sound wave. The resolution can theoretically reach 1/2 of the physical size, so that the azimuth resolution is independent of the propagation distance. Side scan sonar and synthetic aperture sonar obtain side-view sound images, while forward-looking sonar obtains forward-view sound images. Both are two-dimensional sound images. The two-dimensional sound image is composed of pixels, and the grayscale intensity of the pixels changes with the intensity of the received sound wave signal. Side scan sonar is mainly composed of transmitting array elements and receiving array elements. When working, the transmitting array element emits sound wave pulses and propagates outward in a spherical shape. The sound waves will be scattered when encountering underwater targets or the bottom of the water, and the backscattered waves will be received by the receiving array elements on the sonar device. According to the different intensities of the received sound waves, different color areas are drawn on the sound map. There are three types of areas in the sonar image: target highlight area, shadow area and seabed reverberation area. The reflection of sound waves from an object will produce an object highlight area. The shadow area is caused by insufficient acoustic backscattering behind the object. The rest of the information is contained in an area called the seabed reverberation area.

Due to the complex and changeable characteristics of seawater, the obtained sonar image sequences are characterized by high noise, low contrast and poor quality. At present, a lot of efforts have been made to achieve underwater sonar target detection. They are mainly divided into traditional methods and deep learning methods. Traditional sonar image target detection methods are usually based on image processing techniques such as threshold segmentation, edge detection, morphology, etc., and mainly use artificial methods such as physical modeling to distinguish foreground targets and background pixels to achieve target location. Traditional methods are complex because they usually require manual determination of the type of features to be extracted, and deep features of sonar images cannot be extracted by these methods.

With the development of computer hardware technology, Convolutional Neural Networks (CNN) have been performing well in the field of image processing with the support of high-computing devices. It has become a standard practice to use deep neural networks to improve the performance of various image classification problems. These attempts have produced state-of-the-art results in many applications, such as vehicle license plate recognition, medical diagnostic imaging, lip reading, fruit and vegetable recognition, etc. Naturally, deep learning has also been introduced into the field of sonar images by scholars, and has become one of the hot topics in active sonar image classification research in recent years. Therefore, the research on deep learning in sonar image target recognition has important practical significance.

In recent years, researchers have introduced deep learning target detection technology, which has performed very well in the field of optics, into the field of sonar images and have achieved certain results. References Galusha A, Dale J, Keller J M, et al. Deep Convolutional Neural Network Target Classification for Underwater Synthetic Aperture Sonar Imagery [J]. Detection and Sensing of Mines, Explosive Objects, and Obscured Targets Xxiv, 2019, 11012. A method of using a 4-layer CNN to classify synthetic sonar aperture images is proposed. The complexity of this model is extremely high, which is unacceptable in engineering. References Williams D P, Dugelay S. Multi-view SAS Image Classification Using Deep Learning [J]. Oceans 2016 Mts/Ieee Monterey, 2016. The image classification problem was first proposed using DCNN active sonar. Literature Park J, Jung DJ. Identifying Tonal Frequencies in a Lofargram with Convolutional Neural Networks [J]. 2019 19th International Conference on Control, Automation and Systems (Iccas 2019), 2019: 338-341. A Lofargram recognition system using CNN was proposed. Although its precision and recall are acceptable, it takes time.

Summary of the invention

Purpose of the invention: In order to overcome the defects in the above-mentioned prior art, the purpose of the present invention is to provide a sonar image target detection method and system based on deep learning, so as to effectively improve the accuracy of sonar image target detection.

Technical solution: In order to achieve the above invention objectives, the present invention adopts the following technical solution:

A sonar image target detection method based on deep learning includes the following steps:

S1, obtain sonar images and screen and sort them, and after classification, perform median filtering and data enhancement on the images to form a sonar image dataset;

S2. Construct a sonar image target detection model YOLO_EC. The YOLO_EC model introduces an ELAN module in the feature extraction network to replace a part of the C3 module for feature extraction and an MP module to replace the convolutional layer for downsampling based on the YOLOv5 network; and introduces a C2f module in the feature fusion network for feature fusion.

S3. Based on the established YOLO_EC sonar image target detection model, the model is pre-trained using the optical image dataset through style conversion and transfer learning to obtain the basic parameter set of the model, and then the sonar image dataset is input into the model for training to obtain a trained detection model;

S4. Input the sonar image to be detected into the trained detection model to obtain the sonar image target detection result.

Furthermore, in step S1, performing median filtering to reduce noise on the image includes: for a pixel in the image, obtaining the grayscale value of its neighboring pixels and arranging them in order into a group, and storing the median value of the group to replace the original pixel value;

After the sonar image dataset is annotated in YOLO format, data enhancement is performed on the dataset, including one or more of random cropping, rotation/affine transformation, random rotation, scaling, translation, perspective transformation, and Mosaic enhancement.

Furthermore, the ELAN module divides the parameters passed from the upper layer into two branches for processing. The first branch passes through a 1×1 convolution layer to reduce the number of channels to half of the original number. The second branch first passes through a 1×1 convolution layer to reduce the number of channels to half of the original number, and then passes through four 3×3 convolution layers for feature extraction. The number of channels remains unchanged, and the features passed through two convolution layers are extracted by introducing a skip-layer link method. Finally, the four feature maps are superimposed together, and the feature extraction result with the number of channels doubled is obtained, so that a feature map with the same size and twice the number of channels is obtained.

Furthermore, the first branch of the MP module first passes through a maximum pooling layer for downsampling, and then passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number; the second branch first passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number, and then passes through convolution layers with a convolution kernel of 3*3 and a step size of 2 for downsampling; finally, the results of the first branch and the second branch are combined to obtain the downsampling result.

Furthermore, the main branch of the C2f module consists of a convolutional layer with k=1, s=1, c=c, a splitting module, and several bottleneck layer modules. k is the convolution kernel size, s is the stride size, c is the number of channels, and c=c means that the convolution layer keeps the number of channels c of the incoming feature map unchanged. The splitting module divides the number of channels of a feature map into two parts, namely c/2, through a splitting operation. One part is directly connected to the Concat part, and the other part continues to be input into the bottleneck layer module. Each time it passes through a bottleneck layer module, the output will be additionally transmitted to the Concat part. Finally, the fusion module fuses X+1 parameters to obtain a feature map with a channel number of (X+1)*(c/2).

Furthermore, the image processing process of the YOLO_EC model is as follows: after the image is input, it passes through two convolutional layers, a C2f module and a convolutional layer, and then passes through the ELAN module to obtain the first feature map P1, and then passes through the MP module and the ELAN module to obtain the second feature map P2, and finally passes through the MP and ELAN modules and inputs into the SPPF module to obtain the feature map P3; the feature map P3 passes through the convolutional layer to obtain the feature map P4, and after P4 is upsampled, it is fused with P2 to obtain the feature map T1, and then After that, T1 passes through the C2f module and the convolution layer to obtain the feature map P5. After P5 is upsampled, it is fused with P1 to obtain the feature map T2. Then T2 passes through the C2f module to obtain the output feature map D1. After D1 passes through the convolution layer and is fused with P5, it obtains the feature map T3. After T3 passes through the C2f module, it obtains the output feature map D2. After D2 passes through the convolution layer and is fused with P4, it obtains the feature map T4. After T4 passes through the C2f module, it obtains the output feature map D3. After the output inference of the feature maps D1, D2, and D3, the final inference result of the model is obtained.

Furthermore, in step S3, adaptive instance regularization AdaIN is used to perform style conversion, which is expressed as follows:

Where x is the optical image, y is the sonar image, μ and σ are the mean and variance of the pixel values of the image, μ(y) refers to the mean of the sonar image, and σ(y) refers to the variance of the sonar image.

The present invention also provides a sonar image target detection system based on deep learning, comprising:

The sonar image preprocessing module obtains sonar images and screens and organizes them. After classification, the images are subjected to median filtering, noise reduction and data enhancement to form a sonar image data set.

The target detection model construction module constructs the sonar image target detection model YOLO_EC. The YOLO_EC model introduces the ELAN module in the feature extraction network to replace part of the C3 module for feature extraction and the MP module to replace the convolution layer for downsampling based on the YOLOv5 network; and introduces the C2f module in the feature fusion network for feature fusion;

The model training module uses the established YOLO_EC sonar image target detection model and uses the optical image dataset to pre-train the model through style conversion and transfer learning to obtain the basic parameter set of the model. The sonar image dataset is then input into the model for training to obtain a trained detection model.

The target detection module inputs the sonar image to be detected into the trained detection model to obtain the sonar image target detection result.

The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the sonar image target detection method based on deep learning as described above are implemented.

The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the sonar image target detection method based on deep learning as described above are implemented.

Beneficial effects: The present invention applies deep learning to the field of sonar images and reasonably improves the network model. By improving the feature extraction network, it effectively fuses features at different levels to optimize the information flow, improves the accuracy and robustness of the model when dealing with targets of different sizes and complexities, and achieves more efficient feature transfer in the network; by improving the feature fusion network, a lower resolution is used in the early stage to reduce the computational burden, and then the resolution is gradually increased to enhance detail recognition, thereby improving accuracy while maintaining efficiency. Transfer learning of style conversion is also applied. By performing style conversion on the source domain data set in a targeted manner, the dependence of transfer learning on the scale of the data set can be effectively alleviated. The present invention effectively improves the accuracy and efficiency of sonar image target detection.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a sonar image target detection method based on deep learning according to the present invention.

FIG2 is a network structure diagram of the YOLO-EC model of the present invention;

FIG3 is a diagram of an ELAN module network structure of the present invention;

FIG4 is a diagram of the network structure of the MP module of the present invention;

FIG5 is a diagram of a C2f module network structure of the present invention;

FIG6 is a schematic diagram of the transfer learning principle of the present invention;

FIG7 is a schematic diagram of the style conversion principle of the present invention;

FIG. 8 is a diagram of target detection results according to an embodiment of the present invention.

DETAILED DESCRIPTION

In order to have a clearer understanding of the features and advantages of the technical solution of the present invention, the composition and implementation of the specific solution are explained below in conjunction with the accompanying drawings.

1 , the present invention proposes a sonar image target detection method based on deep learning, comprising the following steps:

Step S1, obtaining sonar images and screening and sorting them to obtain a sonar image dataset.

Sonar images are obtained from the Internet, experiments, and open source data sets and screened and sorted, including: cropping, segmenting, and classifying the collected images, eliminating images that are too large and have poor imaging effects, and finally dividing the images into four parts: shipwrecks, crashed planes, human bodies, and unknown targets. The images are then processed by median filtering. Median filtering is a nonlinear signal processing technology based on sorting statistics theory that can effectively suppress noise. Its core is to replace the pixel grayscale value of the image with the median value of the grayscale value of the neighboring pixels. Arrange the neighboring pixels in grayscale order, and store the median value of the group to replace the original pixel value of the point. The mathematical expression of median filtering is shown in formula (1):

g(x,y)＝median{f(x-i,y-j)},(i,j)∈S (1)

Where f(x,y) is the pixel value in the original image, g(x,y) is the pixel value in the processed image, S is the window, and i,j are the length and width of the window.

After annotating the sonar image dataset into YOLO format, data enhancement is performed on the dataset, such as random cropping, rotation/affine transformation, random rotation, scaling, translation, perspective transformation, and Mosaic enhancement.

Step S2, constructing a sonar image target detection model YOLO_EC based on the YOLOv5 network. The ELAN module, MP module and C2f module are introduced into the YOLO_EC model to improve the YOLOv5 network and realize accurate detection of underwater targets.

The network structure improvement of the YOLO_EC model in the present invention includes the improvement of the feature extraction network and the improvement of the feature fusion network, and adopts the FReLU activation function.

In the feature extraction network, ELAN modules are used to replace a certain number of C3 modules, and MP modules are used for downsampling instead of the convolutional layer in the original network. This can enhance the network's ability to learn sonar image targets, avoid using too many transition layers while performing multi-branch calculations, and make the shortest gradient path of the entire network quickly become longer, effectively improving the problem in the ResNet structure that stacking residual layers only increases the longest gradient path but does not change the shortest gradient path. The ELAN module optimizes the information flow by effectively fusing features at different levels, improves the accuracy and robustness of the model when dealing with targets of different sizes and complexities, and achieves more efficient feature transfer in the network. The structure of the ELAN module is also more conducive to reasoning and training under high-concurrency computing devices (such as GPUs). The network structure diagram of the ELAN module is shown in Figure 3. The ELAN module divides the parameters passed from the upper layer into two branches for processing. The first branch passes through a 1×1 convolution layer to reduce the number of channels to half of the original. The second branch first passes through a 1×1 convolution layer to reduce the number of channels to half of the original number, and then passes through four 3×3 convolution layers for feature extraction, the number of channels remains unchanged, and the features that pass through two convolution layers are extracted by introducing skip-layer links. Finally, the four feature maps are superimposed together, and the number of channels is doubled. In this way, a feature map with the same size and twice the number of channels is obtained. The network structure diagram of the MP module is shown in Figure 4. The MP module is a downsampling module. The first branch first passes through a maximum pooling layer. The function of maximum pooling is downsampling, and then passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number. The second branch first passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number, and then passes through convolution layers with a convolution kernel of 3*3 and a step size of 2 to perform downsampling. Finally, the results of the first branch and the second branch are combined to obtain the downsampling result. This module fuses the two downsampling modes and finally obtains a feature map with the same number of channels and a quarter of the original image size.

The neck part is used as a feature fusion network. The present invention uses a C2f module for feature fusion. The C2f network structure diagram is shown in Figure 5. The main branch of the C2f module consists of a convolution layer with k=1, s=1, c=c, a split module, and several bottleneck layer modules. k is the convolution kernel size, s is the stride size, c is the number of channels, and c=c means that the convolution layer keeps the number of channels c of the incoming feature map unchanged. The split module divides the number of channels of a feature map into two parts, namely c=c/2, by a split operation. One part is directly connected to the Concat part, and the other part continues to be input to the bottleneck layer module. After each bottleneck layer module, the output is additionally transmitted to the Concat module. Finally, the fusion module fuses X+1 parameters to obtain a feature map with a channel number of c=(X+1)*(c/2). Using the C2f module, more abundant gradient flow information can be obtained. On the basis of using the bottleneck layer, the C2f module adds more skip layer connections and additional split operations. The split operation refers to dividing a larger feature map into two smaller feature maps to reduce the amount of calculation and the number of parameters. The C2f structure can use a lower resolution in the early stage to reduce the computational burden, and then gradually increase the resolution to enhance detail recognition, thereby improving accuracy while maintaining efficiency.

The overall structure of the improved YOLO_EC sonar image target detection model is shown in Figure 2. After the image is input, it passes through two Conv layers, a C2f module and a convolution layer (Conv), and then passes through the ELAN module to obtain the first feature map P1, and then passes through the MP module and the ELAN module to obtain the second feature map P2. Finally, after passing through the MP and ELAN modules, it is input into the SPPF module to obtain the feature map P3. Feature map P3 passes through the Conv layer to obtain feature map P4. After P4 is upsampled (Upsample), it is fused with P2 to obtain feature map T1. Then T1 passes through the C2f module and the Conv layer to obtain feature map P5. After P5 is upsampled, it is fused with P1 to obtain feature map T2. Then T2 passes through the C2f module to obtain the output feature map D1. D1 passes through the Conv layer and P5 to obtain feature map T3. T3 passes through the C2f module to obtain the output feature map D2. D2 passes through the Conv layer and is fused with P4 to obtain feature map T4. After T4 passes through the C2f module, it obtains the output feature map D3. After output inference, the feature graphs D1, D2, and D3 are used to obtain the final inference result of the model.

Step S3: Based on the established YOLO_EC sonar image target detection model, the model is pre-trained by style conversion plus transfer learning.

The principle of transfer learning is shown in Figure 6. The model is first pre-trained in the source domain, i.e., the optical image dataset, and the pre-trained parameters in the obtained pre-trained model can play a role in target detection in sonar images. The optical image dataset adopts the mainstream ImageNet dataset. The dataset obtained by the present invention contains 1000 categories of common optical image targets, with 1300 images in each category. In order to enhance the connection between the optical dataset and the sonar image dataset and reduce the pressure of training, 16 categories of images are selected, including ships, airplanes, and other optical image types similar to targets in sonar images, as well as pictures related to sonar usage scenarios such as beaches and coasts. The images with official annotations in the 16 categories of datasets are selected to form a new training dataset, which is named IN-16. IN-16 contains a total of 9728 pictures, including 3 categories of aircraft, 11 categories of various ships, and 2 categories of beaches and coasts, totaling 16 categories of images. All images in the dataset are annotated. Before pre-training of transfer learning, the present invention performs style conversion on the optical data set to improve the similarity between the optical image and the sonar image, that is, the optical image data set pays more attention to the shape features of the target in the image, so that in the pre-training process, the model can better learn and retain the ability to extract shape features, and show better feature extraction ability when the parameters are migrated to the sonar image field. The schematic diagram of style conversion is shown in Figure 7. The encoder in the style conversion network is a trained VGG convolutional neural network. The optical image is used as the content image and the sonar image is used as the optical image input encoder to extract the image features of both. Then, after passing through the AdaIN layer, the sonar image style is mapped to the optical image through calculation, and finally the image size is restored through the upsampling layer in the decoder. The present invention uses AdaIN style conversion, and the method proposes adaptive instance regularization AdaIN (Adaptive Instance Normalization). The affine parameters in AdaIN do not need to be obtained through learning. They are calculated in real time through the input style image and adapt to different style images. Any style conversion can be achieved through AdaIN. The specific formula is shown in formula (2):

Where x is the content image, which is an optical image in this invention, and y is the style image, which is a sonar image in this invention. μ and σ are the mean and variance of the pixel values of the image, respectively. μ(y) refers to the mean of the sonar image, which is achieved by adding all the pixel values of the sonar image and averaging them. σ(y) refers to the variance of the sonar image, which is achieved by subtracting the mean from the pixel values of each point in the sonar image and squaring them, then summing and dividing by the number of pixels. The regularized input is simply scaled by σ(y) and offset by μ(y).

Step S4, input the sonar image data set into the YOLO_EC model for training to obtain a detection model, and input the image to be detected into the detection model to obtain the sonar image target detection result.

The dataset after using sonar images to perform style conversion on optical images shows that the accuracy of the model after transfer learning is improved to a certain extent compared to the original one. This shows that style conversion plays a certain role in migrating the features of two datasets with large differences. By performing targeted style conversion on the source domain dataset, the dependence of transfer learning on large-scale datasets can be effectively alleviated, and the effect of the model in the sonar image detection task can be further improved. The final sonar image target detection result is shown in Figure 8.

In an embodiment of the present invention, the training set used for training is a self-built data set, which is processed through screening, cropping, data enhancement, and filtering and noise reduction. Research is conducted on the application of deep learning methods in the field of sonar image target detection based on this data set. In terms of model establishment, considering the application needs of actual projects, YOLOv5 is selected as the reference model for target detection, and the network structure is improved on this basis. In order to effectively fuse features at different levels to optimize information flow and improve the accuracy and robustness of the model when processing targets of different sizes and complexities, the ELAN module is used to optimize the feature extraction network. In order to reduce the computational burden and enhance the detail recognition capability at the same time, the C2f module is used to modify the feature fusion network. It has been verified that the improved method can effectively improve the detection performance of the sonar image target detection model.

For the overfitting problem, pre-training the model in a field with a larger amount of data through transfer learning to improve feature extraction capabilities is a common method. Given the large difference between the optical dataset and the sonar dataset, the present invention converts the optical dataset into a dataset with sonar image style through AdaIN style conversion, making the features between the optical dataset and the sonar image closer. This greatly improves the help of the transfer parameters for sonar image target detection.

Based on the same inventive concept as the method, the present invention also provides a sonar image target detection system based on deep learning, comprising:

The sonar image preprocessing module obtains sonar images and screens and organizes them. After classification, the images are subjected to median filtering, noise reduction and data enhancement to form a sonar image data set.

The target detection model construction module constructs the sonar image target detection model YOLO_EC. The YOLO_EC model introduces the ELAN module in the feature extraction network to replace part of the C3 module for feature extraction and the MP module to replace the convolution layer for downsampling based on the YOLOv5 network; and introduces the C2f module in the feature fusion network for feature fusion;

The model training module uses the established YOLO_EC sonar image target detection model and uses the optical image dataset to pre-train the model through style conversion and transfer learning to obtain the basic parameter set of the model. The sonar image dataset is then input into the model for training to obtain a trained detection model.

The target detection module inputs the sonar image to be detected into the trained detection model to obtain the sonar image target detection result.

It should be understood that the deep learning-based sonar image target detection system in the embodiment of the present invention can implement all the technical solutions in the above method embodiment, and the functions of its various functional modules can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can refer to the relevant description in the above embodiment, which will not be repeated here.

The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the sonar image target detection method based on deep learning as described above are implemented.

The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the sonar image target detection method based on deep learning as described above are implemented.

It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices (systems), computer equipment or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

The present invention is described with reference to a flow chart of a method according to an embodiment of the present invention. It should be understood that each process in the flow chart and a combination of processes in the flow chart can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in one or more processes in the flow chart.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in one or more processes of the flowchart.

These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart.

Claims (10)

1. A sonar image target detection method based on deep learning, characterized in that it includes the following steps: S1, obtain sonar images and screen and sort them, and after classification, perform median filtering and data enhancement on the images to form a sonar image dataset; S2. Construct a sonar image target detection model YOLO_EC. The YOLO_EC model introduces an ELAN module in the feature extraction network to replace a part of the C3 module for feature extraction and an MP module to replace the convolutional layer for downsampling based on the YOLOv5 network; and introduces a C2f module in the feature fusion network for feature fusion. S3. Based on the established YOLO_EC sonar image target detection model, the model is pre-trained using the optical image dataset through style conversion and transfer learning to obtain the basic parameter set of the model, and then the sonar image dataset is input into the model for training to obtain a trained detection model; S4. Input the sonar image to be detected into the trained detection model to obtain the sonar image target detection result. 2. The method according to claim 1, characterized in that in step S1, performing median filtering to reduce noise on the image comprises: for a pixel in the image, obtaining the grayscale value of its neighboring pixels and arranging them in order into a group, and storing the median value of the group to replace the original pixel value; After the sonar image dataset is annotated in YOLO format, data enhancement is performed on the dataset, including one or more of random cropping, rotation/affine transformation, random rotation, scaling, translation, perspective transformation, and Mosaic enhancement. 3. The method according to claim 1 is characterized in that the ELAN module divides the parameters passed from the upper layer into two branches for processing, the first branch passes through a 1×1 convolution layer to reduce the number of channels to half of the original number; the second branch first passes through a 1×1 convolution layer to reduce the number of channels to half of the original number, and then passes through four 3×3 convolution layers for feature extraction, the number of channels remains unchanged, and by introducing a skip-layer link method, the features passed through two convolution layers are extracted; finally, the four feature maps are superimposed together, and the number of channels is twice the original number of feature extraction results, so that a feature map with the same size and twice the number of channels is obtained. 4. The method according to claim 1 is characterized in that the first branch of the MP module first passes through a maximum pooling layer for downsampling, and then passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number; the second branch first passes through a convolution layer with a convolution kernel of 1×1 to reduce the number of channels to half of the original number, and then passes through convolution layers with a convolution kernel of 3*3 and a step size of 2 for downsampling; finally, the results of the first branch and the second branch are combined to obtain the downsampling result. 5. The method according to claim 1 is characterized in that the main branch of the C2f module consists of a convolutional layer with k=1, s=1, c=c, a splitting module, and several bottleneck layer modules, k is the convolution kernel size, s is the stride size, c is the number of channels, c=c means that the convolution layer keeps the number of channels c of the incoming feature map unchanged, the splitting module divides the number of channels of a feature map into two parts, namely c/2, through a splitting operation, one part is directly connected to the Concat part, and the other part continues to be input into the bottleneck layer module, and each time it passes through a bottleneck layer module, the output will be additionally transmitted to the Concat part, and finally the fusion module fuses X+1 parameters to obtain a feature map with a channel number of (X+1)*(c/2). 6. The method according to claim 5 is characterized in that the YOLO_EC model processes the image as follows: after the image is input, it passes through two convolutional layers, a C2f module and a convolutional layer, and then passes through the ELAN module to obtain the first feature map P1, and then passes through the MP module and the ELAN module to obtain the second feature map P2, and finally after passing through the MP and ELAN modules, it is input into the SPPF module to obtain the feature map P3; the feature map P3 passes through the convolutional layer to obtain the feature map P4, and after P4 is upsampled, it is fused with P2 to obtain To feature map T1, then T1 passes through the C2f module and the convolution layer to obtain feature map P5, P5 is upsampled and fused with P1 to obtain feature map T2, then T2 passes through the C2f module to obtain output feature map D1; D1 passes through the convolution layer and is fused with P5 to obtain feature map T3, T3 passes through the C2f module to obtain output feature map D2; D2 passes through the convolution layer and is fused with P4 to obtain feature map T4, T4 passes through the C2f module to obtain output feature map D3; feature maps D1, D2, and D3 are output inference to obtain the final inference result of the model. 7. The method according to claim 1, characterized in that, in step S3, adaptive instance regularization AdaIN is used to perform style conversion, which is expressed as follows: Where x is the optical image, y is the sonar image, μ and σ are the mean and variance of the pixel values of the image, μ(y) refers to the mean of the sonar image, and σ(y) refers to the variance of the sonar image. 8. A sonar image target detection system based on deep learning, characterized by comprising: The sonar image preprocessing module obtains sonar images and screens and organizes them. After classification, the images are subjected to median filtering, noise reduction and data enhancement to form a sonar image data set. The target detection model construction module constructs the sonar image target detection model YOLO_EC. The YOLO_EC model introduces the ELAN module in the feature extraction network to replace part of the C3 module for feature extraction and the MP module to replace the convolution layer for downsampling based on the YOLOv5 network; and introduces the C2f module in the feature fusion network for feature fusion; The model training module uses the established YOLO_EC sonar image target detection model and uses the optical image dataset to pre-train the model through style conversion and transfer learning to obtain the basic parameter set of the model. The sonar image dataset is then input into the model for training to obtain a trained detection model. The target detection module inputs the sonar image to be detected into the trained detection model to obtain the sonar image target detection result. 9. A computer device, comprising: one or more processors; Memory; and One or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the sonar image target detection method based on deep learning as described in any one of claims 1 to 7 are implemented. 10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the sonar image target detection method based on deep learning as described in any one of claims 1 to 7 are implemented.