Mask wearing detection based on yolov1

pick want

In recent years , The target detection algorithm has made a great breakthrough . The more popular algorithms can be divided into two categories , One is based on Region Proposal Of R-CNN System algorithm （R-CNN,Fast R-CNN, Faster R-CNN）, They are two-stage Of , You need to use heuristics first （selective search） perhaps CNN The Internet （RPN） produce Region Proposal, And then in Region Proposal Do classification and regression . And the other is Yolo、SSD This kind of one-stage Algorithm , It uses only one CNN The network directly predicts the categories and locations of different targets . The first method has higher accuracy , But it's slow ; The second kind of algorithm is fast , But less accurate . This paper introduces Yolo Algorithm , Its full name is You Only Look Once: Unified, Real-Time Object Detection. This topic basically Yolo The characteristics of the algorithm summarize ：You Only Look Once It means only once CNN operation ;Unified It means that this is a unified framework , Provide end-to-end The forecast ; and Real-Time reflect Yolo The algorithm is fast .
key word Image classification ; Mask detection ;YOLOv1; Feature learning ; Preliminary training

introduction

2019 The outbreak of novel coronavirus pneumonia , Great changes have taken place in people's travel —— since 1 month 24 After Wuhan was declared closed on the th , All provinces and cities have successively launched the first level response to major public health emergencies to control population mobility . Many cities have stipulated that masks must be worn 、 Take your temperature before you can take public transportation .2 month 10 Before the rework day on the th , Shanghai 、 Beijing and other key cities have also successively released new regulations ： Access to the airport 、 Rail transit 、 Inter-City Bus Station 、 Health care institutions 、 Shopping malls, supermarkets and other public places , Those who do not wear masks will be discouraged .
2 month 13 Japan , Baidu PaddlePaddle announces open source industry's first mask face detection and classification model . Based on this model , It can detect a large number of faces in the public scene at the same time , Mark the face with and without mask , Fast recognition of various scenes without attention 、 Don't pay attention to virus protection , Even people with fluke mentality , Reduce potential safety hazards in public . At the same time, build more public welfare applications for epidemic prevention .

Design concept

On the whole ,Yolo The algorithm uses a separate CNN Model implementation end-to-end Target detection for , The whole system is shown in the figure above ： First, input the picture resize To $448\times 448$, And then send in CNN The Internet , Finally, the network prediction results are processed to obtain the detected target . comparison R-CNN The algorithm is a unified framework , It's faster , and Yolo The training process is also end-to-end Of .
The system divides the input image into $S\times S$ grid . If the center of a target falls into a grid cell , Then the grid unit is responsible for detecting the target . Each grid cell predicts $B$ individual bounding box and box The confidence score of . These confidence scores reflect the model's confidence that the box contains an object , And it thinks the accuracy of box prediction . Formally , We define confidence as $Pr（Object）\times IOU$,$IOU$ finger intersection over union between the predicted box and the ground truth. If there is no prediction object in this cell , Then the confidence should be zero ; Otherwise, the confidence is equal to $IOU$. Every bounding box from $5$ A prediction consists of ：$x$,$y$,$w$,$h$ And confidence .$x,y$ Express bounding box Relative to the center of the grid cell boundary ;$w,h$ Predict width and height relative to the entire image ; Last , The confidence level represents the difference between the prediction frame and the real frame $IOU$.
Each grid cell also predicts $ C $ Conditional class probabilities $Pr（Clas{s}_i|Object）$. These probabilities are conditional on the grid cell containing the detection object . No matter what B How many , Predict only one set of class probabilities for each grid cell . At testing time , Multiply the probability of the condition category by the confidence prediction of each box :
$$Pr(Clas{s}_i|Object)\times Pr(Object)\times IO{U}^{truth}pred=Pr(Clas{s}_i)\times IO{U}^{truth}pred$$
Thus, the confidence of a certain class of each box is obtained . These results represent the probability of the category appearing in the box , And predicted bounding box The degree of suitability for the tested object .

Network design

According to the above structure diagram , The input image size is $448\times 448$, After several convolution layers and pooling layers , Turn into $7\times 7\times 1024$ tensor （ The penultimate cube in Figure 1 ）, Finally, it goes through two full connection layers , The output tensor dimension is $7\times 7\times 30$, This is it. Yolo v1 The whole neural network structure , It is not much different from the general convolution object classification network , The big difference is this ： The last full connection layer of the classified network , Usually connected to a one-dimensional vector , Different bits of a vector represent different categories , The output vector here is a three-dimensional tensor （ $7\times 7\times 30$）. Above picture Yolo Network structure , Inspired by GoogLeNet, It's also v2、v3 in Darknet The vanguard of . Essentially nothing special , Not used BN layer , Used a layer Dropout. Except that the output of the last layer uses a linear activation function , All other layers are used Leaky Relu Activation function .
A fast version of YOLO, It aims to break through the boundary of rapid object detection . Fast YOLO The neural network used has fewer convolution layers （ 9 A rather than 24 individual ）, And there are fewer filters in these layers . In addition to the size of the network ,YOLO and Fast YOLO All training and test parameters are the same .

LOSS function

After the neural network structure is determined , The training effect is good or bad , from Loss Functions and optimizers decide .Yolo v1 Use the ordinary gradient descent method as the optimizer . Here is a key interpretation Yolo v1 The use of Loss function ：

Loss function , The paper gives a more detailed explanation . All losses are based on the sum of squares error formula , Let's not look at the formula for the moment ${\lambda }_{coord}$ And ${\lambda }_{noobj}$ , The predicted value of output and the loss caused are :
The center point of the prediction box $(x,y)$ . The loss caused is the first line . among $1_{ij}^{obj}$ Is the control function , At the grid points where the label contains the object , The value is 1 ; If the lattice does not contain an object , The value is 0. That is to say, we only calculate the loss of those grid points that the real object belongs to , If the lattice does not contain an object , So the predicted value does not affect the loss function .(x, y) Simple sum of squares error between values and labels .
The width and height of the prediction box $(w,h)$ . The loss caused is the second line of the figure .$1_{ij}^{obj}$ It means the same thing , It's also that only the grid point to which the real object belongs will cause losses . Here to （w, h） The treatment in the loss function takes the root sign respectively , The reason lies in , If you don't take the root , The loss function tends to adjust the prediction box with larger size . for example ,$20$ The deviation of one pixel , about $800\times 600$ There is little effect on the prediction box of , At this time IOU The value is still very large , But for $30\times 40$ The prediction box has a great impact . The root sign is to eliminate the difference between the large size box and the small size box as much as possible .
Confidence of prediction frame C. The third and fourth lines . When the lattice does not contain an object , The label of the confidence level is 0; If there are objects , The label of the confidence is the difference between the prediction box and the real object box IOU The number .
Probability of object class P. The fifth row . Corresponding category position , The tag value is 1, The rest are 0, The same as the classification network .
Now let's look at ${\lambda }_{coord}$ And ${\lambda }_{noobj}$ ,Yolo The problem of object detection , It's a typical problem of unbalanced number of categories . among 49 Grid points , A lattice containing an object usually has only 3、4 individual , The rest are grid points without objects . At this point, if no measures are taken , So object detection is mAP It won't be too high , Because models tend to have no grid points of objects .${\lambda }_{coord}$ And ${\lambda }_{noobj}$ The role of , That is, let the lattice containing the object , More weight in the loss function , Make the model more “ Attaches great importance to ” The loss caused by a grid containing objects . In the paper , ${\lambda }_{coord}$ And ${\lambda }_{noobj}$ The values of are 5 And 0.5.

ginseng Examination writing offer

[] Li Xingchen . The fusion YOLO Detection of multi-target tracking algorithm [J]. Computer engineering and science ,2020(4):665-672.
[2] Li Zhangwei . Overview of vision based object detection methods [J]. Computer engineering and Application ,2020(3):1-9.
[3] Liu Lei . A kind of YOLO Identification and Mean shift Tracking traffic flow statistical method [J. Manufacturing automation ,2020(2):16-20.
[4] Wang Feng . Application of artificial intelligence in drawing recognition in door and window inspection [J]. Sichuan cement ,2019(9):137-138.
Welcome to join me for wechat exchange and discussion （ Please note csdn Add ）