[go: up one dir, main page]

CN111178451A - A license plate detection method based on YOLOv3 network - Google Patents

A license plate detection method based on YOLOv3 network Download PDF

Info

Publication number
CN111178451A
CN111178451A CN202010002151.6A CN202010002151A CN111178451A CN 111178451 A CN111178451 A CN 111178451A CN 202010002151 A CN202010002151 A CN 202010002151A CN 111178451 A CN111178451 A CN 111178451A
Authority
CN
China
Prior art keywords
target
license plate
box
loss
predicted
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.)
Pending
Application number
CN202010002151.6A
Other languages
Chinese (zh)
Inventor
屈景怡
冯晓赛
杨俊�
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.)
Civil Aviation University of China
Original Assignee
Civil Aviation University of China
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 Civil Aviation University of China filed Critical Civil Aviation University of China
Priority to CN202010002151.6A priority Critical patent/CN111178451A/en
Publication of CN111178451A publication Critical patent/CN111178451A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/60—Type of objects
    • G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
    • G06V20/63—Scene text, e.g. street names
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/60—Type of objects
    • G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
    • G06V20/625—License plates

Landscapes

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

Abstract

本发明提供了一种基于YOLOv3网络的车牌检测方法，包括数据预处理：搜集包含车牌的图片，分类整理；通过数据均衡的方法扩大较少的新能源车牌数量、使馆领馆车牌数量和民航车牌数量，使之与蓝色车牌数量相当，将上述图片标记为数据集；特征提取：将数据预处理中得到的数据集编码后输入到Darknet‑53网络，取最后两层残差层的输出作为特征矩阵；分类预测：将特征提取中得到的两个不同维度的特征矩阵，拼接后送入逻辑回归分类器，输出车牌的位置和种类。本发明所述的基于YOLOv3网络的车牌检测方法，采用锚盒机制，只用一次特征提取，预测出车牌目标的位置和类别两种信息，减少了计算量，提高了计算速度。

Figure 202010002151

The invention provides a license plate detection method based on YOLOv3 network, including data preprocessing: collecting pictures containing license plates, sorting and sorting; expanding the number of fewer new energy license plates, the number of embassy and consulate license plates and civil aviation license plates by means of data balancing The number of blue license plates is equal to the number of blue license plates, and the above picture is marked as a dataset; Feature extraction: The dataset obtained in data preprocessing is encoded and input to the Darknet‑53 network, and the output of the last two residual layers is taken as Feature matrix; classification prediction: The two feature matrices of different dimensions obtained in feature extraction are spliced and sent to the logistic regression classifier to output the location and type of the license plate. The license plate detection method based on the YOLOv3 network of the present invention adopts the anchor box mechanism and uses only one feature extraction to predict the position and category of the license plate target, which reduces the amount of calculation and improves the calculation speed.

Figure 202010002151

Description

License plate detection method based on YOLOv3 network
Technical Field
The invention belongs to the technical field of big data and deep learning, and particularly relates to a license plate detection method based on a YOLOv3 network.
Background
With the rapid development of civil aviation in China, the airport scale is getting bigger and bigger. Various tools and vehicles in an airport present huge challenges to an automatic vehicle management system, and license plate detection is an important link of the automatic vehicle management system. At present, a plurality of license plate detection systems are put into commercial use in China. There are two main methods for detecting license plates at home and abroad: one is a traditional license plate detection method based on prior characteristics; the other is a license plate detection method based on deep learning. The traditional license plate detection method mainly utilizes the characteristics of the license plate, such as contour, texture, color and the like, to model the license plate. These conventional methods are characterized by small calculation amount, low accuracy and poor robustness. With the fact that the AlexNet network based on deep learning in 2012 obtains the ILSVRC champion in the current year, the convolutional neural network obtains rich results, and development of the target detection method based on deep learning is greatly promoted. At present, target detection networks based on deep learning are mainly divided into two-step detection networks and one-step detection networks. The two-step detection network mainly comprises R-CNN series, such as R-CNN, SPP-NET, Fast R-CNN and the like, and is characterized in that a target detection task is divided into two steps: the position and type of the target are detected by first detecting pre-selected frames (Region pro-potential) of suspected targets, then extracting the features of the candidate frames, and deducing the probability that the candidate frames belong to each target. The single-step detection network directly obtains the class probability and the position of the target through one-step feature extraction, and the method is faster than a two-step detection method. Typical single-step assays are available as YOLO, YOLO9000, YOLOv3, SSD. The YOLOv3 network outputs the category and the bounding box of the predicted target object directly in a regression mode, and is characterized by high detection speed but poor detection effect on small target groups. In the license plate detection, the license plate target is large, and no small target group exists, so that YOLOv3 is used as a basic network for license plate detection.
Disclosure of Invention
In view of this, the present invention is directed to providing a YOLOv3 network-based license plate detection method, so as to provide a YOLOv3 network-based license plate detection method that is capable of adapting to a complex environment and has high detection accuracy.
In order to achieve the purpose, the technical scheme of the invention is realized as follows:
a license plate detection method based on a YOLOv3 network comprises the following steps: data preprocessing: collecting pictures containing license plates, and sorting; expanding the number of new energy license plates, the number of embassy and guild license plates and the number of civil aviation license plates by a data balancing method to be equal to the number of blue license plates, and marking the pictures as a data set; feature extraction: coding a data set obtained in data preprocessing, inputting the coded data set into a Darknet-53 network, and taking the output of the last two residual error layers as a feature matrix; and (3) classification prediction: and splicing the two feature matrixes with different dimensions obtained in the feature extraction, sending the spliced feature matrixes into a logistic regression classifier, and outputting the position and the type of the license plate.
Further, the data equalization method includes: and expanding the number of picture samples by methods of distortion, rotation and random noise addition.
Further, the Darknet-53 network consists of a series of 1 × 1 and 3 × 3 convolutional layers, each followed by a BN layer and a LeakyReLU layer.
Further, in the feature extraction, the Darknet-53 network includes an activation function, a convolutional layer, a shortcut link layer, a routing layer, an upsampling layer and a YOLO layer, and the YOLO layer includes anchor box parameters, target categories and the number of preselected frames.
Further, the step of using the anchor box parameter to select the target candidate box by YOLOv3 is as follows: YOLOv3 divides the input image into s × s grids, each grid predicts the positions of n candidate frames and the confidence degrees of the target types corresponding to the suspected targets in the candidate frames according to the anchor box parameters and the multi-scale scaling of the feature map, and the method for obtaining the positions of the target candidate frames by the anchor box parameters is shown as the following formula:
Figure RE-GDA0002412973180000031
in the above formula, cx,cyRepresenting grid position, anchor box center as grid center, bh,bwRepresenting the length and width offset of the target candidate box with respect to the anchor box, bx,byDenotes the center offset, σ is a logistic regressionThe function of the function is that of the function,
Figure RE-GDA0002412973180000032
further, the specific method of classification prediction is as follows: in the feature extraction, the output of the last two residual error layers of the DarkNet-53 network is subjected to sampling and tensor splicing to obtain two feature maps with different sizes, license plate targets with different sizes are predicted on two scales with different sizes, five parameters of x, y, w, h and p are required to be predicted in each scale, wherein the x, y, w and h correspond to the abscissa of the upper left corner, the ordinate of the upper left corner, the width and the height of a license plate target boundary frame in a data set label, and the p represents the category and the confidence coefficient of the corresponding category of the license plate targets,
further, the method of using the logistic regression classifier in the classification prediction is that YOLOv3 mainly implements logistic regression from the feature map to the output parameters through three loss functions: loss of target location offset Lloc(l, g) for determining the position of the license plate target; target confidence loss Lconf(o, c) for determining the probability that the license plate target belongs to different license plate types; target classification loss Lcla(O, C) for indicating the kind of license plate to which the license plate object belongs, wherein λ1,λ2,λ3Is the equilibrium coefficient:
L(O,o,C,c,l,g)＝λ1Lloc(l,g)+λ2Lconf(o,c)+λ3Lcla(O,C)
target confidence loss: the target confidence coefficient represents the probability of the target existing in the target rectangular frame, and the target confidence coefficient loss adopts binary cross entropy loss, wherein oiE {0,1} represents whether the target really exists in the predicted target boundary box i, 0 represents the absence, and 1 represents the existence;
Figure RE-GDA0002412973180000033
the Sigmoid probability of whether the target exists in the predicted target rectangular frame i is shown, and the predicted value c is obtainediObtained by sigmoid function:
Figure RE-GDA0002412973180000034
loss of target class: target class penalty Lcla(O, C) also employs a binary cross-entropy penalty, where O isijE {0,1}, which represents whether the jth class target really exists in the prediction target boundary box i, 0 represents nonexistence, and 1 represents existence;
Figure RE-GDA0002412973180000041
the Sigmoid probability of the j-th class target in the boundary box i of the network prediction target is represented by a predicted value CijObtained by sigmoid function:
Figure RE-GDA0002412973180000042
loss of target location: loss of target location Lloc(l, g) using the sum of squares of the difference between the true deviation value and the predicted deviation value, wherein
Figure RE-GDA0002412973180000043
Indicating the predicted rectangular box coordinate offset, where the net predicts the offset, not the direct predicted coordinate,
Figure RE-GDA0002412973180000044
indicating the coordinate offset between the preselected frame and the default frame with which it matches,
Figure RE-GDA0002412973180000045
the middle superscript m is the { x, y, w, h }, bx、by、bw、bhRespectively the predicted upper left abscissa and upper left ordinate of the target rectangular frame, the width of the bounding box, the height of the bounding box, cx、cy、pw、phRespectively the horizontal coordinate of the upper left corner, the vertical coordinate of the upper left corner, the width of the boundary box, the height of the boundary box, gx、gy、gw、ghRespectively the upper left-hand abscissa, the upper left-hand ordinate and the edge of the real target rectangular frame matched with the default preselected frameThe width of the bounding box and the height of the bounding box, and the parameters are mapped on the prediction characteristic diagram
Figure RE-GDA0002412973180000046
Compared with the prior art, the license plate detection method based on the YOLOv3 network has the following advantages:
(1) the license plate detection method based on the YOLOv3 network adopts an anchor box mechanism, only uses one-time feature extraction to predict two kinds of information of the position and the category of a license plate target, reduces the calculated amount and improves the calculation speed.
(2) The license plate detection method based on the YOLOv3 network is suitable for the arrangement of the anchor boxes of the license plate targets, so that the extraction of the target candidate frames is more targeted, the multi-scale feature prediction quantity is reasonably arranged aiming at the characteristic that the license plate targets have larger occupation ratio in the whole picture, and the detection speed is improved on the premise of not reducing the detection accuracy.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate an embodiment of the invention and, together with the description, serve to explain the invention and not to limit the invention. In the drawings:
fig. 1 is a state diagram of a data equalization method according to an embodiment of the present invention;
FIG. 2 is a diagram illustrating the YOLOv3 anchor box mechanism according to an embodiment of the present invention;
fig. 3 is a structural diagram of a license plate detection network based on the YOLOv3 network according to an embodiment of the present invention;
FIG. 4 is a block diagram of Darknet-53 of a feature extraction network according to an embodiment of the present invention;
FIG. 5 shows reduced operands compared to YOLOv3 according to an embodiment of the present invention;
FIG. 6 is a network hyper-parameter set according to an embodiment of the present invention;
FIG. 7 is a comparison of the performance of the embodiments of the present invention and the YOLOv3 network for detecting license plates.
Detailed Description
It should be noted that the embodiments and features of the embodiments may be combined with each other without conflict.
In the description of the present invention, it is to be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", and the like, indicate orientations or positional relationships based on those shown in the drawings, and are used only for convenience in describing the present invention and for simplicity in description, and do not indicate or imply that the referenced devices or elements must have a particular orientation, be constructed and operated in a particular orientation, and thus, are not to be construed as limiting the present invention. Furthermore, the terms "first", "second", etc. are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first," "second," etc. may explicitly or implicitly include one or more of that feature. In the description of the present invention, "a plurality" means two or more unless otherwise specified.
In the description of the present invention, it should be noted that, unless otherwise explicitly specified or limited, the terms "mounted," "connected," and "connected" are to be construed broadly, e.g., as meaning either a fixed connection, a removable connection, or an integral connection; can be mechanically or electrically connected; they may be connected directly or indirectly through intervening media, or they may be interconnected between two elements. The specific meaning of the above terms in the present invention can be understood by those of ordinary skill in the art through specific situations.
The present invention will be described in detail below with reference to the embodiments with reference to the attached drawings.
The noun explains:
YOLOv3: YOLOv3 is An object detection network, produced by Josep Redmond et al in the literature Joseph Redmon, Ali faradai, yolovi 3: An incorporated improvement. arxiv: 1804.02767, respectively.
Darknet-53 network: YOLOv 3.
A license plate detection method based on YOLOv3 network, as shown in fig. 1 to 7, includes the following steps: data preprocessing: and collecting pictures containing the license plate, and sorting.
The data preprocessing mainly comprises five license plate types, namely a blue license plate, a yellow license plate, a new energy license plate, a license plate of a guild hall and a license plate special for civil aviation of a small civil vehicle.
The number of new energy license plates, the number of license plates of a messenger hall and the number of license plates of a civil aviation are reduced by a data balancing method, so that the number of various license plates is basically the same, and the purpose of balancing data is achieved.
The data equalization method is shown in fig. 1, and the number of picture samples is expanded by means of warping, rotating and adding random noise. The four pictures in fig. 1 represent from left to right in turn: the method comprises the steps of expanding picture original images, distorting the original images, rotating the original images and adding noise into the original images.
The pictures were labeled in the format of a Pascal VOC data set by manual labeling.
Feature extraction: and inputting the data codes obtained in the data preprocessing into a Darknet-53 network, and taking the output of the last two residual error layers as a feature matrix.
Feature extraction: and (3) outputting the last two residual error layers of the DarkNet-53 network, and performing up-sampling and tensor splicing to obtain two feature maps with different sizes for detecting license plate targets with different sizes.
The Darknet-53 network in the feature extraction comprises an activation function, a convolutional layer, a shortcut link layer (ShortcutConnections), a routing layer, an upsampling layer and a YOLO layer (the YOLO layer is used for realizing the functions of splicing and feature extraction after the feature graph is upsampled).
It should be noted that Darknet-53 does not have a pooling layer in the conventional sense, but rather achieves the effect of reducing the dimension of the feature map by adjusting the step size of the convolution.
Darknet-53 is composed of a series of 1X 1 and 3X 3 convolutional layers. Each convolutional layer is followed by a BN layer and a LeakyReLU layer.
The YOLO layer includes YOLOv 3-specific parameters such as anchor box parameters, object class, number of pre-selected frames, and the like. The step of using the anchor box parameter to select the target candidate box by the YOLOv3 is as follows: YOLOv3 divides the input image into s × s grids, and each grid predicts the positions of n candidate frames and the confidence of the target type corresponding to the suspected target in the candidate frame according to the multi-scale scaling of the anchor box parameters and the feature map (t)x,ty,tw,th,to) Wherein (t)x,ty) Represents the coordinates of the center of the candidate frame, (t)w,th) Width and height of the candidate box, toRepresenting the target class confidence.
The method for obtaining the target candidate frame position from the anchor box parameters is shown in the following formula (1):
Figure RE-GDA0002412973180000081
in the above formula (1), cx,cyRepresenting the grid location. As shown in fig. 2 below, the grid position is represented by the upper left coordinate of the grid, the dashed box represents the anchor box, and the gray box represents the offset of the target candidate box from the anchor box. The target candidate box is located in the grid (c)x,cy) In, px,pyHeight and width of anchor box, center of anchor box being center of grid, bh,bwRepresenting the length and width offset of the target candidate box with respect to the anchor box, bx,byRepresenting the center offset, sigma is a logistic regression function,
Figure RE-GDA0002412973180000082
and (3) classification prediction: and splicing the two feature matrixes with different dimensions obtained in the feature extraction, sending the spliced feature matrixes into a logistic regression classifier, and outputting the position and the type of the license plate.
And (3) splicing two feature matrixes with different sizes obtained in feature extraction through sampling and tensor, and predicting license plate targets with different sizes on two scales with different sizes. In each scale, five parameters of x, y, w, h and p need to be predicted. Wherein x, y, w and h correspond to the abscissa of the upper left corner, the ordinate of the upper left corner, the width and the height of the bounding box of the license plate target in the data set label, and p represents the category and the confidence coefficient of the corresponding type of the license plate target.
In classification prediction, YOLOv3 implements logistic regression from feature maps to output parameters mainly through three loss functions: loss of target location offset Lloc(l, g) for determining the position of the license plate target; target confidence loss Lconf(o, c) for determining the probability that the license plate target belongs to different license plate types; target classification loss Lcla(O, C) for representing the kind of the license plate to which the license plate object belongs. Wherein λ1,λ2,λ3Is the equilibrium coefficient:
L(O,o,C,c,l,g)＝λ1Lloc(l,g)+λ2Lconf(o,c)+λ3Lcla(O,C) (2)
(1) target confidence loss: the target confidence degree represents the probability of the target existing in the target rectangular frame, and the target confidence degree loss adopts Binary Cross Entropy loss (Binary Cross Entropy). Wherein o isiE {0,1} represents whether the target really exists in the predicted target bounding box i, 0 represents not existing, and 1 represents existing.
Figure RE-GDA0002412973180000091
And (4) the Sigmoid probability of whether the target exists in the predicted target rectangular box i or not is shown. Will predict the value ciObtained by sigmoid function:
Figure RE-GDA0002412973180000092
(2) loss of target class: target class penalty Lcla(O, C) also employs a binary cross entropy penalty. Wherein, OijE {0,1}, which indicates whether the jth class target really exists in the predicted target bounding box i, 0 indicates that the jth class target does not exist, and 1 indicates that the jth class target exists.
Figure RE-GDA0002412973180000093
Indicating the existence of the jth category in the network prediction target bounding box iTarget Sigmoid probability, from predicted value CijObtained by sigmoid function.
Figure RE-GDA0002412973180000094
(3) Loss of target location: loss of target location Lloc(l, g) the sum of the squares of the difference between the true deviation value and the predicted deviation value is used. Wherein
Figure RE-GDA0002412973180000095
And the coordinate offset of the predicted rectangular box is shown, wherein the network predicts the offset and does not directly predict the coordinate.
Figure RE-GDA0002412973180000096
Indicating the coordinate offset between the preselected frame and the default frame with which it matches,
Figure RE-GDA0002412973180000097
the middle superscript m is e { x, y, w, h }. bx、by、bw、bhThe horizontal coordinate of the upper left corner of the predicted target rectangular frame, the vertical coordinate of the upper left corner, the width of the boundary frame and the height of the boundary frame are respectively. c. Cx、cy、 pw、phRespectively the abscissa of the upper left corner of the default preselected frame, the ordinate of the upper left corner, the width of the bounding box and the height of the bounding box. gx、gy、gw、ghThe horizontal coordinate of the upper left corner of the real target rectangular frame, the vertical coordinate of the upper left corner, the width of the boundary frame and the height of the boundary frame which are matched with the default preselected frame are respectively set. These parameters are mapped on the predicted feature map.
Figure RE-GDA0002412973180000101
The specific method for classified prediction comprises the following steps: the process of processing the signature graph into the target class replaces Softmax with logistic regression, which has the advantage that the license plate labels that are classified may not be independent of each other. In addition, logistic regression is used to score the portion of the anchor box that surrounds a target, i.e., how likely this location is to be a license plate. This step is performed prior to prediction, and unnecessary anchor boxes may be eliminated to reduce the amount of computation.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims (7)

1.一种基于YOLOv3网络的车牌检测方法，其特征在于：包括如下步骤：1. a license plate detection method based on YOLOv3 network, is characterized in that: comprise the steps: 数据预处理：搜集包含车牌的图片，分类整理；通过数据均衡的方法扩大较少的新能源车牌数量、使馆领馆车牌数量和民航车牌数量，使之与蓝色车牌数量相当，将上述图片标记为数据集；Data preprocessing: collect pictures containing license plates and sort them out; expand the number of new energy license plates, the number of embassy and consulate license plates, and the number of civil aviation license plates by means of data balance, so that they are equal to the number of blue license plates, and mark the above pictures. is the dataset; 特征提取：将数据预处理中得到的数据集编码后输入到Darknet-53网络，取最后两层残差层的输出作为特征矩阵；Feature extraction: encode the dataset obtained in data preprocessing and input it to the Darknet-53 network, and take the output of the last two residual layers as the feature matrix; 分类预测：将特征提取中得到的两个不同维度的特征矩阵，拼接后送入逻辑回归分类器，输出车牌的位置和种类。Classification prediction: The two feature matrices of different dimensions obtained in the feature extraction are spliced and sent to the logistic regression classifier to output the location and type of the license plate. 2.根据权利要求1所述的一种基于YOLOv3网络的车牌检测方法，其特征在于：所述数据均衡的方法为：通过扭曲、旋转、加入随机噪声的方法，扩充图片样本数量。2. A kind of license plate detection method based on YOLOv3 network according to claim 1, is characterized in that: the method of described data equalization is: by the method of twisting, rotating, adding random noise, expand picture sample quantity. 3.根据权利要求1所述的一种基于YOLOv3网络的车牌检测方法，其特征在于：Darknet-53网络由一系列的1×1和3×3的卷积层组成，每个卷积层后都会跟一个BN层和一个LeakyReLU层。3. A kind of license plate detection method based on YOLOv3 network according to claim 1, it is characterized in that: Darknet-53 network is made up of a series of 1×1 and 3×3 convolutional layers, after each convolutional layer Will be followed by a BN layer and a LeakyReLU layer. 4.根据权利要求1所述的一种基于YOLOv3网络的车牌检测方法，其特征在于：所述特征提取中Darknet-53网络包括激活函数、卷积层、快捷链路层、路由层、上采样层和YOLO层，YOLO层包括了锚盒参数、目标类别和预选框个数。4. a kind of license plate detection method based on YOLOv3 network according to claim 1, is characterized in that: in described feature extraction, Darknet-53 network comprises activation function, convolution layer, shortcut link layer, routing layer, upsampling layer and YOLO layer, the YOLO layer includes anchor box parameters, target category and the number of preselected boxes. 5.根据权利要求4所述的一种基于YOLOv3网络的车牌检测方法，其特征在于：所述YOLOv3用锚盒参数选取目标候选框的步骤为：YOLOv3将输入图像分成s×s的网格，每个网格根据锚盒参数及特征图多尺度缩放预测出n个候选框的位置及候选框内疑似目标对应的目标种类置信度，由锚盒参数获得目标候选框位置的方法如下式所示：5. a kind of license plate detection method based on YOLOv3 network according to claim 4, is characterized in that: the step that described YOLOv3 selects target candidate frame with anchor box parameter is: YOLOv3 divides input image into the grid of s × s, Each grid predicts the position of n candidate frames and the confidence of the target type corresponding to the suspected target in the candidate frame according to the anchor box parameters and multi-scale scaling of the feature map. The method of obtaining the target candidate frame position from the anchor box parameters is shown in the following formula :
Figure FDA0002353870070000021
Figure FDA0002353870070000021
 上式中，cx,cy表示网格位置，锚盒中心为网格中心，bh,bw表示目标候选框相对锚盒的长宽偏移量，bx,by表示中心偏移量，σ为逻辑回归函数，
Figure FDA0002353870070000022
In the above formula, c x , c y represent the grid position, the center of the anchor box is the grid center, b h , b w represent the length and width offset of the target candidate frame relative to the anchor box, b x , b y represent the center offset quantity, σ is the logistic regression function,
Figure FDA0002353870070000022
 6.根据权利要求1所述的一种基于YOLOv3网络的车牌检测方法，其特征在于：分类预测的具体方法为：特征提取中DarkNet-53网络最后两个残差层的输出，经过采样和张量拼接后，得到两个大小不同的特征图，在两个不同大小的尺度上预测不同大小的车牌目标，每个尺度中，需要预测出x、y、w、h、p这五个参数，其中x、y、w、h对应数据集标签中的车牌目标边界框的左上角横坐标、左上角纵坐标、宽度、高度，p表示车牌目标对应种类的类别及置信度，6. a kind of license plate detection method based on YOLOv3 network according to claim 1 is characterized in that: the concrete method of classification prediction is: the output of the last two residual layers of DarkNet-53 network in feature extraction, after sampling and Zhang After splicing, two feature maps of different sizes are obtained, and license plate targets of different sizes are predicted on two scales of different sizes. In each scale, five parameters of x, y, w, h, and p need to be predicted. Where x, y, w, and h correspond to the upper left abscissa, upper left ordinate, width and height of the bounding box of the license plate target in the label of the dataset, and p represents the category and confidence of the corresponding type of the license plate target, 7.根据权利要求6所述的一种基于YOLOv3网络的车牌检测方法，其特征在于：在分类预测中使用逻辑回归分类器的方法为，YOLOv3主要通过三个损失函数实现由特征图到输出参数的逻辑回归：目标定位偏移量损失Lloc(l,g)，用来确定车牌目标的位置；目标置信度损失Lconf(o,c)，用来确定车牌目标属于不同车牌种类的概率；目标分类损失Lcla(O,C),用来表示车牌目标所属车牌的种类，其中λ1,λ2,λ3是平衡系数：7. a kind of license plate detection method based on YOLOv3 network according to claim 6, is characterized in that: the method that uses logistic regression classifier in classification prediction is, YOLOv3 mainly realizes from feature map to output parameter through three loss functions The logistic regression of : target positioning offset loss L loc (l, g), used to determine the location of the license plate target; target confidence loss L conf (o, c), used to determine the probability that the license plate target belongs to different license plate types; The target classification loss L cla (O, C) is used to represent the type of license plate to which the license plate target belongs, where λ 1 , λ 2 , λ 3 are balance coefficients: L(O,o,C,c,l,g)＝λ1Lloc(l,g)+λ2Lconf(o,c)+λ3Lcla(O,C)L(O,o,C,c,l,g)=λ 1 L loc (l,g)+λ 2 L conf (o,c)+λ 3 L cla (O,C) 目标置信度损失：目标置信度表示目标矩形框内存在目标的概率，目标置信度损失采用的是二值交叉熵损失，其中，oi∈{0,1}表示预测目标边界框i中是否真实存在目标，0表示不存在，1表示存在；
Figure FDA0002353870070000023
表示预测目标矩形框i内是否存在目标的Sigmoid概率，将预测值ci通过sigmoid函数得到：Target confidence loss: The target confidence represents the probability of the existence of the target in the target rectangular box, and the target confidence loss adopts the binary cross-entropy loss, where o i ∈ {0,1} indicates whether the predicted target bounding box i is true or not There is a target, 0 means it does not exist, and 1 means it exists;
Figure FDA0002353870070000023
Represents the sigmoid probability of predicting whether there is a target in the target rectangle i, and the predicted value c i is obtained by the sigmoid function:
Figure FDA0002353870070000031
Figure FDA0002353870070000031
 目标类别损失：目标类别损失Lcla(O,C)同样采用的是二值交叉熵损失，其中，Oij∈{0,1}，表示预测目标边界框i中是否真实存在第j类目标，0表示不存在，1表示存在；
Figure FDA0002353870070000034
表示网络预测目标边界框i内存在第j类目标的Sigmoid概率，由预测值Cij通过sigmoid函数得到：Target category loss: The target category loss L cla (O,C) also uses the binary cross-entropy loss, where O ij ∈ {0,1}, which indicates whether there is a real j-th target in the predicted target bounding box i, 0 means does not exist, 1 means exists;
Figure FDA0002353870070000034
Represents the sigmoid probability of the j-th type of target in the target bounding box i predicted by the network, which is obtained from the predicted value C ij through the sigmoid function:
Figure FDA0002353870070000032
Figure FDA0002353870070000032
 目标定位损失：目标定位损失Lloc(l,g)采用的是真实偏差值与预测偏差值差的平方和，其中
Figure FDA0002353870070000035
表示预测矩形框坐标偏移量，这里网络预测的是偏移量，不是直接预测坐标，
Figure FDA0002353870070000036
表示与之匹配的预选框与默认框之间的坐标偏移量，
Figure FDA0002353870070000037
中上标m∈{x,y,w,h}，bx、by、bw、bh分别为预测的目标矩形框的左上角横坐标、左上角纵坐标、边界框的宽、边界框的高，cx、cy、pw、ph分别为默认预选框的左上角横坐标、左上角纵坐标、边界框的宽、边界框的高，gx、gy、gw、gh分别为与默认预选框匹配的真实目标矩形框的左上角横坐标、左上角纵坐标、边界框的宽、边界框的高，这些参数都是映射在预测特征图上的，Target localization loss: The target localization loss L loc (l, g) is the sum of the squares of the difference between the true deviation value and the predicted deviation value, where
Figure FDA0002353870070000035
Indicates the coordinate offset of the predicted rectangular frame, where the network predicts the offset, not the direct predicted coordinate.
Figure FDA0002353870070000036
Indicates the coordinate offset between the matching preselected box and the default box,
Figure FDA0002353870070000037
The superscript m∈{x,y,w,h}, b x , b y , b w , and b h are the abscissa of the upper left corner, the ordinate of the upper left corner, the width of the bounding box, the boundary of the predicted target rectangle, respectively The height of the box, c x , c y , p w , and ph are the abscissa of the upper left corner, the ordinate of the upper left corner, the width of the bounding box, the height of the bounding box, g x , g y , g w , g h are the abscissa of the upper left corner, the ordinate of the upper left corner, the width of the bounding box, and the height of the bounding box of the real target rectangle matching the default pre-selection box. These parameters are mapped on the prediction feature map,
Figure FDA0002353870070000033
Figure FDA0002353870070000033
CN202010002151.6A 2020-01-02 2020-01-02 A license plate detection method based on YOLOv3 network Pending CN111178451A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010002151.6A CN111178451A (en) 2020-01-02 2020-01-02 A license plate detection method based on YOLOv3 network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010002151.6A CN111178451A (en) 2020-01-02 2020-01-02 A license plate detection method based on YOLOv3 network

Publications (1)

Publication Number Publication Date
CN111178451A true CN111178451A (en) 2020-05-19

Family

ID=70652593

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010002151.6A Pending CN111178451A (en) 2020-01-02 2020-01-02 A license plate detection method based on YOLOv3 network

Country Status (1)

Country Link
CN (1) CN111178451A (en)

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111709381A (en) * 2020-06-19 2020-09-25 桂林电子科技大学 Road environment target detection method based on YOLOv3-SPP
CN111723697A (en) * 2020-06-05 2020-09-29 广东海洋大学 An Improved Driver Background Segmentation Method Based on Mask-RCNN
CN111814783A (en) * 2020-06-08 2020-10-23 三峡大学 Accurate license plate location method based on license plate vertex offset estimation
CN112149514A (en) * 2020-08-28 2020-12-29 中国地质大学（武汉） Method and system for detecting safety dressing of construction worker
CN112560852A (en) * 2020-12-15 2021-03-26 三峡大学 Single-stage target detection method with rotation adaptive capacity based on YOLOv3 network
CN112711996A (en) * 2020-12-22 2021-04-27 中通服咨询设计研究院有限公司 System for detecting occupancy of fire fighting access
CN112818964A (en) * 2021-03-31 2021-05-18 中国民航大学 Unmanned aerial vehicle detection method based on FoveaBox anchor-free neural network
CN114974579A (en) * 2022-04-20 2022-08-30 山东大学齐鲁医院 Auxiliary judgment system and equipment for endoscopic treatment prognosis of gastrointestinal submucosal tumors
CN114998902A (en) * 2022-04-07 2022-09-02 讯飞智元信息科技有限公司 Area detection method and related device, equipment and medium
CN119478356A (en) * 2024-10-23 2025-02-18 南京理工大学 A post-processing optimization technique for a single-stage target recognition algorithm

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108009526A (en) * 2017-12-25 2018-05-08 西北工业大学 A kind of vehicle identification and detection method based on convolutional neural networks
CN109815886A (en) * 2019-01-21 2019-05-28 南京邮电大学 A pedestrian and vehicle detection method and system based on improved YOLOv3
CN110059554A (en) * 2019-03-13 2019-07-26 重庆邮电大学 A kind of multiple branch circuit object detection method based on traffic scene
CN110147807A (en) * 2019-01-04 2019-08-20 上海海事大学 A kind of ship intelligent recognition tracking
CN110245689A (en) * 2019-05-23 2019-09-17 杭州有容智控科技有限公司 Shield cutter identification and position finding and detection method based on machine vision
CN110399884A (en) * 2019-07-10 2019-11-01 浙江理工大学 A Feature Fusion Adaptive Anchor Frame Model Vehicle Detection Method

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108009526A (en) * 2017-12-25 2018-05-08 西北工业大学 A kind of vehicle identification and detection method based on convolutional neural networks
CN110147807A (en) * 2019-01-04 2019-08-20 上海海事大学 A kind of ship intelligent recognition tracking
CN109815886A (en) * 2019-01-21 2019-05-28 南京邮电大学 A pedestrian and vehicle detection method and system based on improved YOLOv3
CN110059554A (en) * 2019-03-13 2019-07-26 重庆邮电大学 A kind of multiple branch circuit object detection method based on traffic scene
CN110245689A (en) * 2019-05-23 2019-09-17 杭州有容智控科技有限公司 Shield cutter identification and position finding and detection method based on machine vision
CN110399884A (en) * 2019-07-10 2019-11-01 浙江理工大学 A Feature Fusion Adaptive Anchor Frame Model Vehicle Detection Method

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111723697A (en) * 2020-06-05 2020-09-29 广东海洋大学 An Improved Driver Background Segmentation Method Based on Mask-RCNN
CN111814783A (en) * 2020-06-08 2020-10-23 三峡大学 Accurate license plate location method based on license plate vertex offset estimation
CN111814783B (en) * 2020-06-08 2024-05-24 深圳市富浩鹏电子有限公司 Accurate license plate location method based on license plate vertex offset estimation
CN111709381A (en) * 2020-06-19 2020-09-25 桂林电子科技大学 Road environment target detection method based on YOLOv3-SPP
CN112149514B (en) * 2020-08-28 2023-04-07 中国地质大学（武汉） Method and system for detecting safety dressing of construction worker
CN112149514A (en) * 2020-08-28 2020-12-29 中国地质大学（武汉） Method and system for detecting safety dressing of construction worker
CN112560852A (en) * 2020-12-15 2021-03-26 三峡大学 Single-stage target detection method with rotation adaptive capacity based on YOLOv3 network
CN112711996A (en) * 2020-12-22 2021-04-27 中通服咨询设计研究院有限公司 System for detecting occupancy of fire fighting access
CN112818964A (en) * 2021-03-31 2021-05-18 中国民航大学 Unmanned aerial vehicle detection method based on FoveaBox anchor-free neural network
CN114998902A (en) * 2022-04-07 2022-09-02 讯飞智元信息科技有限公司 Area detection method and related device, equipment and medium
CN114974579A (en) * 2022-04-20 2022-08-30 山东大学齐鲁医院 Auxiliary judgment system and equipment for endoscopic treatment prognosis of gastrointestinal submucosal tumors
CN114974579B (en) * 2022-04-20 2024-02-27 山东大学齐鲁医院 Auxiliary judgment system and equipment for prognosis of endoscopic treatment of submucosal tumors of the digestive tract
CN119478356A (en) * 2024-10-23 2025-02-18 南京理工大学 A post-processing optimization technique for a single-stage target recognition algorithm

Similar Documents

Publication Publication Date Title
CN111178451A (en) A license plate detection method based on YOLOv3 network
CN109829399B (en) Vehicle-mounted road scene point cloud automatic classification method based on deep learning
CN108288088B (en) Scene text detection method based on end-to-end full convolution neural network
CN101989302B (en) Multilayer bitmap color feature-based image retrieval method
CN111461083A (en) A fast vehicle detection method based on deep learning
CN114913498B (en) A parallel multi-scale feature aggregation lane detection method based on key point estimation
CN114898327B (en) A vehicle detection method based on lightweight deep learning network
CN112200186A (en) Vehicle logo recognition method based on improved YOLO_V3 model
CN108846835A (en) The image change detection method of convolutional network is separated based on depth
CN110717493B (en) License plate recognition method containing stacked characters based on deep learning
CN111797920B (en) Remote sensing extraction method and system for depth network impervious surface with gate control feature fusion
CN116524189A (en) High-resolution remote sensing image semantic segmentation method based on coding and decoding indexing edge characterization
CN114998866A (en) Traffic sign identification method based on improved YOLOv4
CN113034511A (en) Rural Building Recognition Algorithm Based on High Resolution Remote Sensing Image and Deep Learning
CN117876401B (en) Cervical liquid-based thin layer cell image segmentation method based on SAM segmentation model
CN112712102A (en) Recognizer capable of simultaneously recognizing known radar radiation source individuals and unknown radar radiation source individuals
CN114998688A (en) A large field of view target detection method based on improved YOLOv4 algorithm
CN116012709A (en) High-resolution remote sensing image building extraction method and system
CN115272826A (en) An image recognition method, device and system based on convolutional neural network
CN115082781A (en) Ship image detection method and device and storage medium
CN119229416A (en) Deep learning model for small-size vehicle detection based on multi-scale module feature fusion
CN115937807A (en) Method and system for identifying and detecting road defects
CN117809293B (en) A small sample image object counting method based on deep neural network
CN108711150B (en) End-to-end pavement crack detection and identification method based on PCA
CN117274175A (en) Insulator defect detection method and storage medium based on improved neural network model

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
RJ01 Rejection of invention patent application after publication

Application publication date: 20200519

RJ01 Rejection of invention patent application after publication