Article Directory
1. Preface
This project is for face recognition by mtcnn+arcsoftmax
2 Face detection
2.1 Overall process
Overview ideas: the use of cascaded thought, on a network as input the output of the next network
2.2 Network structure
2.2.1 Pnet
Function : Determine whether there is a face, and give a face frame and key points, and provide a suggestion frame for the R network
Input : 12x12x3 picture
Output :
1. Probability of human face. I used the two-classification method to make the loss function, so I only output 1x1x1 1 channel (if the multi-class loss function is used, 2 channels can be output)
The reason for outputting a channel: One is that the loss function uses two classifications as the loss function, and the other is when performing coordinate inverse calculations.
2. The coordinates of the offset of the face detection frame (upper left and lower right). Output 1x1x4 4 channels
3, face key point coordinates. The output is 1x1x10 (not implemented in the code)
2.2.2 Rnet
Function : To further determine the suggestion box of Pnet output, and to further improve the accuracy of the presence or absence of the face
Input : 24x24x3
Output : Fully connected
1, 1 == "face/no face
2, 4 == "recommended frame coordinate offset
3, 10 == "key point coordinate offset (code not implemented)
3) Onet
Effect : the output of Rnet further refine and further improve the face accuracy of the presence or absence of
difference : compared with other networks, network layers totaled O 6 layers
Input : 48x48x3
Output : Fully connected
1, 1== "face/no face
2, 4==" suggested frame coordinate offset
3, 10 == "key point coordinate offset (code not implemented)
2.3 Data processing
Use data set: CelebA
download link: https://pan.baidu.com/s/1_e8pSnfeMT0fCFEtvm8pSA
extraction code: wpav
2.3.1 Data generation
Project folder composition
Idea:
2.3.2 dataset
Note: When I do, in the return value, the image data, confidence, and offset are returned separately (the confidence and offset may not be returned separately)
Question:
Training