Target detection notes fast r-cnn

Faster R-CNN

1.R-CNN

• 2014 Year by year Ross Girshick In the paper Rich feature hierarchies for accurate object detection and semantic segmentation It is proposed that .

  

1）RCNN Algorithm flow ：

​ 1. A picture is generated 1K~2K Candidate areas （ Use Selective Search Method ）

​ 2. For each candidate area , Use deep networks to extract features

​ 3. Features are fed into each type of SVM classifier , Judge whether it belongs to this category

​ 4. Use regression to fine tune candidate box position

2）R-CNN The problem is ：

​ 1. The test speed is slow

​ 2. Training is slow

​ 3. Training needs a lot of space

2.Fast R-CNN

​ When selecting eigenvalues ,R-CNN The eigenvalues of candidate regions need to be calculated repeatedly ,fast R-CNN, Just calculate the eigenvalue of the whole graph , Then the candidate regions are mapped to the feature map , Thus greatly reducing the amount of calculation .

​ Not all candidate boxes are selected during training , Only a part will be taken , The sampling of training data should have positive samples and negative samples , In the author's paper , One time use 64 Candidate box , There are positive samples and negative samples , The author believes that as long as the candidate box and the real sample box IOU Greater than 0.5 Then it is recognized as a positive sample , Less than 0.5 Of is certified as a negative sample .

​ Select samples and pass ROI Pooling Layer shrink to 7*7 Size

1） classifier

​ Output N+1 The probability of categories ,N Types of detection targets , A background , adopt softmax The sum of all the probabilities after the function is 1.

2） Bounding box regressors

​ Output the corresponding N+1 Candidate bounding box regression parameters for categories ($d_x,d_y,d_w,d_h$), common (N+1)*4 Nodes .

3） Loss function

​ The loss function is also divided into classification loss and boundary box regression loss .

• Classified loss ：$L_{cls}(p,u)=-log{p}_u$​​​ That is to say softmax Cross entropy loss ​

  ​ Cross entropy loss (Cross Entropy Loss)

1. Many classification problem (softmax Output , all The sum of the output probabilities is 1)

$H=-{\sum }_i{o}_i^{\ast }log(o_i)$​​

2. Two classification problem (sigmoid Output , Every Output nodes do not affect each other )

$H=-\frac{1}{N}{\sum }_{i=1}^N[o_i^{\ast }log{o}_i+(1-o_i^{\ast })log(1-o_i)]$​

ps: among $o_i^{\ast }$​​ Is the true label value ,$o_i$​ For the predicted value , Default $log$​ With $e$​ The bottom is equal to $ln$

• Bounding box regression loss ：$[u>=1]$​​​​​​ yes Iverson brackets ​, When u>=1 When is 1, Others are 0.
  Iverson brackets represent , When the detection target is the background, the bounding box loss does not exist .
  

  ​ Bounding box regression parameters $v_i$ Use $G_i$​ The inverse solution of the formula is sufficient .

4）Fast R-CNN frame

3.Faster R-CNN（RPN+Fast R-CNN）

​ contrast Fast R-CNN,Faster R-CNN It's just Use one RPN The Internet has replaced SS Candidate box generation algorithm .

1）RPN structure

• Acquisition of feature map , Use ZF Convolution network obtained 256 Characteristic diagram of the channel , Use VGG16 Convolution network obtained 512 Characteristic diagram of the channel

• according to The original image corresponding to the characteristic image Location ( Here we use the feature graph and the original graph Equal proportion corresponds to that will do , Error allowed , That is, if the proportion is not integer ), frame K Specified size ( Here is Author use Size ) Of anchor boxes.

  

• Through the above steps, we can get nearly 2w individual anchor boxes, Remove crossing edges Is left 6k, There are many boxes with overlapping parts , Use Non maximum suppression ,IOU Set to 0.7, So each picture is left with 2k individual Candidate box .

• adopt cls layer Generate 2K A category score ( Background and objects Probability , here Use softmax Many classification , if Use sigmoid The second category is K individual Category score ), adopt reg layer Generate 4K A bounding box regression parameter .

• About positive and negative samples ： Sample and calibration box IOU Greater than 0.7 Or the biggest One is called Positive sample , Less than 0.3 Of is called Negative sample .

2）RPN Loss function

​ Cross entropy used for classification loss Calculation ( Select according to the classification using the algorithm Many classification still Two classification ), Border regression loss Calculation Same as Fast R-CNN Boundary box loss calculation Agreement .$\lambda$​ The value in the article is 10, Therefore, the parameters can be approximately merged . Back Fast R-CNN The loss function of the network is the same as before . In the article Use dual networks to train alone , It is usually used directly RPN and Fast R-CNN Lose the method of joint training .

3）Faster R-CNN frame

4） The problem is

• The detection effect of small targets is poor High level abstraction , Resulting in the loss of features .
• The model is big , The detection speed is slow , The main reason is because of two predictions .