Mask wearing detection based on yolov3

pick want

In order to solve the problem of low recognition accuracy caused by many small-size targets in the detection of objects worn by citizens' masks , This paper proposes a method based on ＹＯＬＯｖ３ Improved algorithm Ｍ＿ＹＯＬＯｖ３. Reconstruct the feature pyramid mechanism , The original ３＊３ The pyramid like structure of is expanded to ４＊４ Size , Change the number of prior frames from ９ Add to １６, adopt The above methods reduce the receptive field of neural network , enhance Ｍ＿ＹＯＬＯｖ３ Sensitivity to small targets . Put the original loss function ＩｏＵ On behalf of in by ＤＩｏＵ , Solve the problem that it is difficult to confirm the descending direction of the gradient when the border is regressed . Based on network disclosure ４０６５ The experimental results of a mask dataset show that ,Ｍ＿ＹＯＬＯｖ３ Of ｍＡＰ（ Average precision Degree mean ） by ８８．４, a Ｔｉｎｙ＿ＹＯＬＯｖ３ and ＹＯＬＯｖ３ Of ｍＡＰ They have been promoted separately １５．９ and ７．２.
key word object detection ;ＹＯＬＯｖ３; Mask wearing test ; Characteristic pyramid ; Convolutional neural networks

introduction

2020 year , Novel coronavirus pneumonia rages the world , Wearing masks correctly is an important measure to prevent the virus from spreading between people . Almost all general algorithms can be applied to mask wearing detection [1], But like People block 、 The target object is small 、 Dense crowds and other issues , The effect detected by the general algorithm is not very ideal . To solve these problems , Researchers have done a lot of research work and made remarkable progress :Pang etc. [2] A spatial attention based on mask is designed Force mechanism module , Let the model pay more attention to the characteristics of the unobstructed part of the pedestrian ; Liu etc. [3] Combine the full convolution network with the idea of variable convolution , In order to increase the flexibility of model feature coding , Position sensitive DCN [4] Pooling , Let the model learn the corresponding features from the parts visible to pedestrians as much as possible , To prevent the interference of other objects . With the continuous development of computer vision technology , Target detection technology is also constantly optimized , It can be roughly divided into two categories : Two steps (Two-Stage) Dharma and single step (One-Stage) Law . Compared with , These two kinds of algorithms are also opposite , The precision of two-step detection is high , The single step method is fast . The most classic one-step method is SSD[5] (singleshot multibox detector)、YOLO[6-9] (youonlylookonce) Series algorithm ; Two step method has R-CNN(region conventionalneuralnetwork) Series algorithm . Considering the need to achieve a real-time detection state in the personnel detection task under actual monitoring , and YOLO Series network can ensure real-time performance in terms of detection speed , Among them the first 3 Substitute version YOLOV3 At the same time, the time and accuracy of detection are taken into account , Compared with its It's a higher version ,YOLOV3 Have more mature 、 More stable technology , So this paper focuses on the general target detection algorithm YOLOV3 On the basis of , Improve the algorithm , Hope to get better detection effect .

1 YOLOV3 Algorithm principle

YOLOV3 yes Redmon And so on , Mainly by backbone network Darknet-53 and YOLO Detection layer composition ,Darknet-53 The structure is mainly It is used to extract the characteristic information of the image ,YOLO Layers are used to predict their categories and Location information . The backbone network structure of the algorithm is shown in Figure 1 Shown .YOLOV3 The algorithm has two obvious advantages : The first advantage is the adoption of Darknet53 [10] Network as the backbone feature extraction network , And combined ResNet residual The idea of network structure [11]. Convolution layer mainly includes two types of filters , Namely 1×1 and 3×3, The former filter is used to compress features , The latter filter The function of the device is mainly through reducing the width and height , To expand the number of channels . One of the biggest features of this structure is that it can be raised by increasing the corresponding depth High accuracy , However, gradient explosion and gradient will also occur in the training model Vanishing question topic ,YOLOV3 The residual block inside the algorithm uses jump connection , It promotes the fusion learning of multiple different features .

chart 1 YOLOV3 Main network structure
For detection images with different sizes ,YOLOV3 The algorithm uses 13×13, 26×26,52×52 The characteristic map of the scale is tested [12], Because each scale Their feelings are different , The size of the detection image is also differentiated , scale The smaller the image, the larger the image will be detected , namely 13×13 Detect large image , and 52×52 Detect small size images ,26×26 Detect medium image . For each For the scale Branch , In each grid, it will detect 3 results , This is because For each ruler degree Next Meeting Yes 3 individual First Examination box (anchorbox), yes root According to the K-Means Generated by clustering . The final will be 3 The results of this test are integrated using non great value Suppression system (non-maximumsuppression,NMS), a have to junction fruit . for instance , Input an image to be detected , Just started to draw it It is divided into S×S The grid of , Need to predict C Categories , In the end 3 The scale was The tensor obtained is S×S× [3× (5+C)], Which contains the target border 4 Offset coordinates and confidence scores , Therefore, it enhances the recognition of small-scale objects Detection capability . This is also YOLOV3 Another big difference between the algorithm and other algorithms advantage . However, the task of wearing masks directly used in natural occasions is still There are some deficiencies . firstly ,YOLOV3 Although the detection accuracy of small targets There is a certain improvement in , But at the same time, there are also problems of insufficient shallow feature extraction problem ; second ,YOLOV3 The accuracy of prediction is to use IoU (intersectionoverunion) Loss function to determine the quality of the prediction box , But when IoU When the value of increases , The accuracy of detection will decrease ; third , For nature The scene is covered 、 Dense crowds and small-scale target detection etc. ask topic , YOLOV3 There are still deficiencies . For the above problems , This article is to improve Accuracy of target detection algorithm for masks in natural environment , With YOLOV3 The algorithm is improved and optimized .

2 The improved YOLOv3 Algorithm

This paper deals with YOLOV3 The improvement of mainly includes 3 In terms of , They are for the main network structure 、 Feature enhanced networks and IoU Improvement of loss function .

2.1 Feature extraction network Darknet_D

Darknet_３ As YOLOv3 Feature extraction network , stay primary count Law After feature extraction, the image information is transferred to the detector for frame regression , And by the original do person ＲｅｄｍｏｎＪ real Examination Prove bright ,Ｄａｒｋｎｅｔ５３ stay sex can And efficiency Good morning! period Ｄａｒｋｎｅｔ The network of the series has been significantly improved ［１］. from On Ｍ＿ＹＯＬＯｖ３ A prediction channel will be added , If you add directly in the shallow network Add this channel to pre measuring , because volume product layer a Less , At this time, the characteristic diagram does not Yes Effective semantic information , There is hardly any reference price for forecasting on this basis value . so Ｍ＿ＹＯＬＯｖ３ carry Out 了 Ｄａｒｋｎｅｔ＿Ｄ network Collateral , it yes On Ｄａｒｋ－ ｎｅｔ５３ On the basis of , A residual structure feature extraction is added at the tail The Internet . The improved class feature pyramid uses residual structure ３、４、５、６ It's about The processed characteristic map is used for multi-scale prediction , Pictured ３ Shown .

chart ３ Ｍ＿ＹＯＬＯｖ３ Feature extraction network

2.2 Class feature pyramid transformation

YOLOv３ The original algorithm used ３＊３ Class characteristic pyramid knot of structure structure , Corresponding ９ individual ruler " Of edge box . primary The reason is yes through too Ｄａｒｋｎｅｔ５３ Of ３ layer Ｒｅｓ＿ｂｏｄｙ junction structure （ Its whole call by Ｒｅｓｂｌｏｃｋ＿ｂｏｄｙ） class like remnant Bad network Collateral The addition principle of superimposes each feature map , and And because this have to Out ３ individual ruler Inch prediction loses Out Ｙ. Change Into the count Law Ｍ＿ＹＯＬＯｖ３ The algorithm transforms it into ４＊４ Class characteristic pyramid structure , Not only is the hierarchy more clear , More sensitive to different sizes , And in each class feature pyramid output layer, an additional Prediction candidate frames are used , Then the output prediction value Ｙ It will be more accurate . Besides , In the residual knot structure Ｒｅｓｂｌｏｃｋ＿ｂｏｄｙ in ,Ｍ＿ＹＯＬＯｖ３ Protect Stay ＹＯＬＯｖ３ Of ＤＢＬ Structure, but modify its parameters , This collection product 、 normalization 、 Activate the network module integrating functions , can enough very good Protect leave Image information and feature extraction . Data processed by feature pyramid , The model packs it into ４ individual Ｙ value , each individual Ｙ value from ４ individual ’ａｎｃｈｏｒ junction structure ’ Group become , Every time One individual ’ａｎｃｈｏｒ structure ’ And the central coordinate containing this predicted value x、y, Offset value and prediction edge The width and height of the frame ｗ、ｈ, And Its pre measuring class other ｃｌａｓｓｅｓ and Set up Letter degree ｓｃｏｒｅ, Such as chart ４ Shown .

chart ４ Improved pyramid structure with similar characteristics

2.3 The number of prior frames increases

primary YOLOv３ There are algorithms ９ individual First Examination edge box , and And use this ９ individual Different sizes Of edge box branch become ３ Group , Yes No Same as ruler degree Of ３ individual , sign chart Into the Line forecast .
The target object size of the experimental data varies greatly , some some mouth cover Of Pixels less than ６４＊６４ｐｘ, Some masks have pixels larger than ５１２＊５１２ｐｘ. this Some excessive size differences lead to the original ＹＯＬＯｖ３ The algorithm is listed on the mask data set Not good now . Improve the algorithm Ｍ＿ＹＯＬＯｖ３ send use 了 １６ Forecast frames as Candidate box . Each of them ４ Candidate boxes correspond to a feature map , Every time individual , sign chart Use ４ Candidate edges box Into the That's ok pre measuring . Therefore, increasing the number of prior frames Law , can more good optimum Should be Count According to the , carry high count Law knowledge other fine degree , and increase strong Its extensive Chemical ability .

2.4 Replacement of the loss function of the intersection and union ratio

ＩｏＵ （ｉｎｔｅｒｓｅｃｔｉｏｎｏｖｅｒｕｎｉｏｎ） Of in writing name call yes “ hand over and Than ”, That is, the intersection of the prediction frame and the real frame , And the union of two boxes Than , Among many target detection algorithms based on deep learning ,ＩｏＵ all yes One An important regional plan count finger mark , Such as type （１） the in , Its in Ｂ and Ｂｇｔ branch Do not represent the area of the prediction box and the real box .
ＹＯＬＯｖ３ Use １ and ＩｏＵ Gradient regression is carried out for the difference of , send use The The loss function can simply and effectively converge the loss value of the training network , Such as type （２） Shown . However, this loss function cannot be used normally under certain circumstances , example Such as really There is no intersection between the real frame and the prediction frame , nothing On edge box far near , this when ＬＩｏＵ Are equal to １, Then there is no distance between the predicted frame and the real frame significance . In this case, the algorithm cannot judge the quality and damage of the two frames The direction of valueless convergence , Pictured ５ Shown

chart 5 The border has no intersection
In response to this question ,Rezatofighi H Put forward 〖G_IoU〗^([10]) idea , The penalty term is added to the original loss function , When there is no intersection of borders , The farther the distance between the two frames is, the greater the loss value , Such as the type （３） Shown .

In style :Ｃ Is the smallest circumscribed rectangle of two borders , In this case , When there is no intersection between borders , Can also judge the distance , And therefore find the descending gradient .
G_IoU Nor can it be applied to all situations , If there is a containment relationship between two borders , Such as one big one small two borders Ａ and Ｂ, Small border Ａ Area and Ｂ The union of areas equals Ａ The area of , namely B^A ⋃▒〖B^B=BA 〗. Then in any case Loss Values are always equal , This is the sum of On Statement L_IoU The problems are similar , The algorithm cannot find the decreasing direction of the loss value , Pictured ６ Shown .

chart ６ The border has a containment relationship
D_IoU yes zhengzhaohui etc. ^([12]) stay G_IoU On the basis of Another loss function of intersection and union ratio deduced from this

In style ：ｄ Represents the Euclidean distance between the center points of two frames .ｃ Is the diagonal distance of the smallest circumscribed rectangle . The way it's calculated and G_IoU class like , Also in the original IoU The penalty term is added to the loss function . The difference is G_IoU The punishment of is more reasonable , It can solve the problem that it is difficult to find the regression gradient when the border contains , Pictured ７ Shown .

chart 7 DIoU Frame
D_IoU It can also provide the moving direction of the prediction frame . But if two boxes contain ,G_IoU Unable to adapt , and D_IoU Not only can we find the direction of the prediction box , And the loss function converges very quickly .
As mentioned above IoU Loss function , Its role is to assist the final Loss convergence , More reasonable IoU The loss function enables the algorithm to better fit the data .
stay M_YOLOv3 The algorithm uses D_IoU To replace the original L_IoU As a loss function , On this basis, the intersection and union ratio of the prediction frame and the real frame is calculated , The experiment proves that it has better detection effect .

2.5 The increase of the proportionality coefficient

stay YOLOv3 In the algorithm, , The loss function of the training network is actually the loss of the central point 、 Loss of width and height 、 Classification loss and confidence loss , This loss value is summed up to , Such as the type （５） Shown
Loss=L_xy+L_wh+L_c+L_s (5)
2.4 As mentioned in section D_IoU The loss function will improve L_xy and L_wh Of meter count , Make the calculation of these two loss values more reasonable . because ＹＯＬＯｖ３ It's multiscale check measuring , When the loss value converges, each frame will be calculated once . So wide and high Big Of Frame Ｌｘｙ and Ｌｗｈ The value of will be higher , On the whole Ｌｏｓｓ Value has the wrong effect .
To solve this problem , The original author proposed the proportional coefficient ｂｏｘ＿ｌｏｓｓ＿ ｓｃａｌｅ This indicator , It is made up of numbers ２ And the offset value ｗ and ｈ The product of It is calculated that , among ｗ and ｈ yes ０ To １ Between ｇｒｏｕｎｄ＿ｔｒｕｔｈ wide High offset value , Such as the type （６） Shown . stay Ｌｘｙ and Ｌｗｈ The last step of the calculation will be the proportional system Count discharge shrink , Increase the influence weight of the small size frame , And then reduce the meaningless influence caused by different prior frame sizes

2.6 Weighted correction of loss function

Such as the type （５） Shown , total Ｌｏｓｓ Values are determined by ４ Independent Of Ｌｏｓｓ Add up and get . because Ｍ＿ＹＯＬＯｖ３ Used ＤＩｏＵ And the scale coefficient is increased , In the type （５） Medium right ４ Add items and divide Ｌｃ The weight calculation has been carried out outside . For balance ４ Influence weight of loss value ,Ｍ＿ＹＯＬＯｖ３ stay most end Ｌｏｓｓ When calculating , hold branch class damage loss value Ｌｃ Weighted calculation , Such as the type （８） Shown

３ Experimental results and Analysis

3.1 Experimental data preparation

3.1.1 Data set introduction
In this experiment, the public mask data set is selected １, This data set comes from the network , from ｌａｂｅｌｉｍｇ Software annotation . according to ３００ｐｘ Divide the input image into , Small size 、 Medium ruler " 、 Big ruler " 、 super Big ruler " （ Any attribute of length and width is greater than ９００ｐｘ）, , respectively, Yes ６５８ Zhang 、１５５７ Zhang 、５６６ Zhang 、１３４ Zhang , not notes Ming Xiaoyou １１５０ Zhang , total ４０６５ A picture . This data set is related to ＶＯＣ２００７ The data set format is consistent , Each piece of data uses a picture corresponding to it ｘｍｌ phase junction close . Each picture corresponds to a ｘｍｌ file , Record various attributes of the target object on the picture , For example, the target object is at the center of the picture 、 The width and height of the marking box 、 Information such as the category of the target object . There are only two classes in this dataset ,”ｆａｃｅ＿ｍａｓｋ” and ”ｆａｃｅ”. In the above properties of the target object , The length and width of the target object marking box appear especially by heavy want , They can represent the size of the target object . because ＹＯＬＯｖ３ Original algorithm Of real Examination yes stay ＶＯＣ２００７ On the dataset , The data of this experiment is to detect faces and masks , And ＶＯＣ２００７ Compared with the data set, the target size is generally smaller . Therefore, this data set and ＶＯＣ２００７ Compare the target size of the data set , According to unilateral ６４、１２８、２５６、５１２ Size the boundaries , See table １.

thus it can be seen , The distribution of this data set is more extensive and detailed , And the size of small targets is too large . Improved algorithm proposed ４ Three layer pyramid structure , And will ａｎｃｈｏｒｓ The number is from ９ increase Add To １６ How to do it , Theoretically support Ｍ＿ＹＯＬＯｖ３ Can get better recognition effect .
3.1.2 Data to enhance
In deep learning theory , The more training data you have , The better the effect of the training model . Because this experimental data only ４０００ The remaining pictures are not enough , It is necessary to use data enhancement methods to simulate more data samples . It can not only make the algorithm model have more learning samples , At the same time, the noise is increased due to data enhancement , Reduce the phenomenon of model over fitting .
In this experiment , Zoom image used 、 Translation transformation 、 Flip 、 Color jitter （ＲＧＢ－＞ＨＳＶ－＞ＲＧＢ） as well as ｂｏｘｅｓ Relocation and other data enhancement means , Ensure that the experimental data is sufficient , Enhance the generalization ability of the model , Improve the robustness of the algorithm .

3.2 The process of training the model

To get the appropriate a priori box size , The most common method is to cluster the labeled data . Among many clustering methods , Under the condition that the number of clusters is set in advance ,Ｋ－Ｍｅａｎｓ Undoubtedly, it has become the preferred solution .
Ｋ－Ｍｅａｎｓ The algorithm is very intuitive , The initial artificial constant Ｋ,Ｋ Represents the number of categories finally divided by the algorithm . The algorithm will randomly select the initial point as the centroid , And by calculating the similarity between each sample and the centroid （ Euclidean distance is generally used ）, Classify the sample points into the most similar cluster , next , Recalculate the centroid of each cluster （ Cluster heart ）, Repeat the process , Until the cluster heart no longer changes , Finally, the cluster and cluster center of each sample are determined . Ｍ＿ＹＯＬＯｖ３ Used The number of clusters is １６ Of Ｋ－Ｍｅａｎｓ The algorithm clustering . Got １６ Group ａｎｃｈｏｒｓ After that, a similar normalization process was carried out , Final have to Out ａｎｃｈｏｒｓ ruler " Such as Next ：（２３,２３）,（３５,３４）,（４４,４４）, （５３,５４）,（６５,６３）,（７６,７５）,（８９,９１）,（１０８,１０５）,（１２１,１２３）,（１４２,１４２）,（１６４,１７５）,（１６７,１５８）,（１９０,１９０）,（２１５,２１２）, （２５３,２５４）,（３７２,３６２）. Two values in brackets （ｗｉｄｔｈ,ｈｉｇｈｔ） branch Don't mean ａｎｃｈｏｒ Width and height .
Considering the hardware conditions of this experiment , Input the picture size uniformly ｒｅｓｉｚｅ by ３５２＊３５２, The initial learning rate is set to ０．００１, Use monitoring during callback ｖａｌ＿ｌｏｓｓ Change the way to reduce the learning rate , Its in The parameter of is set to ｐａｔｉｅｎｃｅ＝６,ｆａｃｔｏｒ＝０．４. total training Practice batch Time ｅｐｏｃｈ＝ ８００,ｂａｔｃｈ＿ｓｉｚｅ＝１０, training Practice network Collateral Ｌｏｓｓ The convergence of values is shown in Figure ８ Shown .

chart 8 Loss convergence

3.3 Evaluation index and experimental results

3.3.1 The recognition effect of small targets
When detecting the input picture , A picture often includes many goals , It is easy to miss inspection , In particular, the missing detection of small-size objects is more common , Pictured ９ Shown （ A rectangular box with a long text box representative “ masks ”, The shorter rectangle of the text box represents “ Face ”）. chart ９ in ＹＯＬＯｖ３ Algorithm and Ｔｉｎｙ＿ＹＯＬＯｖ３ The algorithm has missed detection , and Ｍ＿ＹＯＬＯｖ３ With a variety of improvement measures , It is more important for small targets sensitive , It is not easy to miss inspection . Experimental results show that , The improved algorithm , To some extent, it improves the problem of difficult recognition of small target objects , Pictured ９ Shown .

chart 9 3 Comparison of detection effects of two algorithms
3.3.2 Recognition effect under difficult conditions
In the experiment ,３ This detection algorithm has a good detection effect for most pictures , Errors mainly occur On confused and occluded data . Such as chart 10 in , The mask in the first row of pictures is very similar to the face color , Only Ｍ＿ ＹＯＬＯｖ３ Algorithm sentence set just indeed （ The longer rectangle of the text box represents “ masks ”, The shorter rectangle of the text box represents “ Face ”）; The second row of pictures are paper folding fans covering the face , Its shape, color and size are very similar to masks , The recognition results of all algorithms are wrong .
3.3.3 ｍＡＰ The evaluation index
The goal is The evaluation indicators used in the detection field are mainly ｍＡＰ（ｍｅａｎ ａｖｅｒａｇｅｐｒｅｃｉｓｉｏｎ）, Average accuracy mean , That is to say ＡＰ Average value . and ＡＰ Is through （９） obtain , In essence, it is to choose 0.1 As γ The interval of , Put each γ Yes Should be Of Ｐｉｎｔｅｒｐ Value addition , Finally, take the mean . among Ｐｉｎｔｅｒｐ yes ＰＲ In the curve , With a recall rate （ｒｅｃａｌｌ） As the abscissa point , The ordinate value corresponding to the curve , Such as the type （１０） Shown （ In the calculation process in , If abscissa ｘ１ There is no corresponding ordinate value , Then use the closest two points ｘ２、ｘ３ Points on the corresponding curve , Make extension cord and ｘ＝ｘ１ Straight lines intersect , The intersection point is ｘ１ Corresponding ordinate value ）

chart 11 It's a category ”ｆａｃｅ” Corresponding Ｐ－Ｒ curve , In the figure, we can see Ｍ＿ＹＯＬＯｖ３ Of ＰＲ The curve is slightly better than the unchanged calculation Law . There are two in this experiment individual class ,”ｆａｃｅ＿ｍａｓｋ” class and ”ｆａｃｅ” class .ｍＡＰ Value is for two classes ＡＰ Sum the values and average .

chart 11 “face” Category Ｐ－Ｒ curve

3.3.4 experimental result

The experiment shows that ,Ｍ ＿ＹＯＬＯｖ３ Algorithm comparison with the original Ｔｉｎｙ＿ ＹＯＬＯｖ３ and ＹＯＬＯｖ３ Algorithm in ｍＡＰ The value has increased １５．９ and ７．２, The detection ability is better than the original algorithm .Ｍ＿ＹＯＬＯｖ３ Of ｆｐｓ The value is ２４, To a certain extent, it can detect the wearing of masks in real time. See the table ２.

4 Conclusion

Improve the algorithm Ｍ＿ＹＯＬＯｖ３ Algorithm in ＹＯＬＯｖ３ Deepen the feature extraction network and add prediction channels , And a more reasonable loss function is used , To improve the ＹＯＬＯｖ３ The algorithm is difficult to recognize small-scale targets , It is applicable to the detection of small target objects . The conclusion is drawn in the experiment of mask data set based on more small-size targets ,ｍＡＰ Index compared with the original ＹＯＬＯｖ３ Promoted ７．２. In the experiment , Some data are extremely confusing , For example, figure １０ Chinese women use paper folding fans to cover their mouths , Several algorithms in this paper all incorrectly determine that it is a mask . In response to this question , temporary Unable to resolve , It's worth thinking about and finding the answer in the next step , Maybe use something similar ＤＯＴＡ The data format of the four point coordinates of the dataset will Improve this problem . except ＹＯＬＯｖ３ Outside the algorithm , Other target detection algorithms are also contending . As a scholar, we should follow the times Step , Continue to learn and understand other target detection algorithms and their network structures , Such as single-stage ＲｅｔｉｎａＮｅｔ［１３］、
ＹＯＬＯｖ４, Two stage ＤＣＲ、ＳＮＩＰ［１４］、ＳＮＩＰＥＲ etc. .

reference ：

[1]NIUZuodong,QINTao,LIHandong,etal.Improvedalgo-
rithmofretinafacefornaturalscenemaskweardetection [J].
ComputerEngineeringandApplications,2020,56 (12):1-7
(inChinese). [ Niu Zuodong , Qin Tao , Li Handong , etc. . improvement RetinaFace
Natural field view mouth cover palin wear check measuring count Law [J]. meter count machine work cheng And Should be use ,
2020,56 (12):1-7.]
[2]PangY,XieJ,HarisKhan M,etal.Mask-guidedattention
networkforoccludedpedestriandetection [C]//IEEE/CVF
InternationalConferenceon Computer Vision.arXiv,2019:
4967-4975.
[3]Liu T,Luo W, Ma L,etal.Couplednetworkforrobust
pedestriandetection with gated multi-layerfeatureextraction
anddeformableocclusionhandling [J].IEEE Transactionson
ImageProcessing,2020,30:1.
[4]ZhuX,HuH,LinS,etal.DeformableConvNetsV2:More
deformable,betterresults [C]//IEEE/CVF Conferenceon
ComputerVisionandPatternRecognition.IEEE,2019:9308-
9316.
[5]LiuW,AnguelovD,ErhanD,etal.SSD:Singleshotmulti-
boxdetector [C]//EuropeanConferenceonComputerVision.
SpringerInternationalPublishing,2016:21-37.
[6]RedmonJ,DivvalaS,GirshickR,etal.Youonlylookonce:
Unified,real-timeobjectdetection [C]//ComputerVision &
PatternRecognition.IEEE,2016:779-788.
[7]RedmonJ,FarhadiA.YOLO9000:Better,faster,stronger
[C]//IEEEConferenceonComputerVision& PatternRecog-
nition.IEEE,2017:6517-6525.
[8]REDMON J, FARHADI A. YOLOV3: An incremental
improvement [C]//ProceedingsofIEEEConferenceonCom-
puter Vision and Pattern Recognition.Washington:IEEE
Press,2018:1-6.
[9]BochkovskiyA, Wang CY,Liao HYM.YOLOv4:Optimal
speedandaccuracyofobjectdetection [J].arXiv:2004.10934,
2020.
[10]WangY,ZhengJC.Real-timefacedetectionbasedonYOLO
[C]//IEEEInternationalConferenceon KnowledgeInnova-
tionandInvention,2018:221-224.
Welcome to join me for wechat exchange and discussion （ Please note csdn Add ）