20220525 RCNN---＞Faster RCNN

1、RCNN (region CNN). A pioneer in the field of object detection .CNN The Internet

1> Candidate area generation

Use selective search The traditional way , First split the picture , Merge areas with a high probability of containing the same object , Normalize , Get a fixed size image .

2>cnn feature extraction

Characteristic drawing reeler

3> svm classifier .

Linear two classifiers for classification , Difficult sample mining to balance the imbalance of positive and negative samples .

Difficult sample mining (Online Hard Sample Mining)_AI Visual netqi's blog -CSDN Blog _ Difficult sample mining

4> Location refinement
The regressor returns to the target area

2、Fast RCNN      End to end 、 be based on VGG16、 Fast
improvement ：
    1> Shared winder

    Put the whole picture into the winder network , Or use it selective search The way , But the amount of calculation is reduced .
    2>Roi Pooling

    Feature pooling , Arbitrary scale transformation , Any size picture input .
    3> Multitasking loss
        Classification and regression are trained together , Use softmax The function classification
3、faster rcnn.    Put forward rpn Extract candidate box network , utilize anchor

    function ：

   1> Feature extraction network

vggnet

2>RPN

   1>>anchor Generate
        Each point of the feature map corresponds to 9 individual anchor, Corresponding to the original image, it can basically cover all objects
   2>>RPN Winder network
        Use 1*1 The winder gets each... In the characteristic graph anchor Prediction score and prediction offset value of
   3>>RPN loss
      Only during training , Will all anchors Match the label , A good match anchors Positive sample , The opposite is a negative sample , Get the classification and offset true values , And the predicted score in part II 、 Offset values do loss Calculation
    4>> Generate proposal
        The value predicted in the second part after using the loss calculation , Select the better proposal, Into the network
    5>> Screening ROI ( Region of interest ) 
3> ROI Pooling
    The essential , Accept the characteristic diagram and ROI, Output to RCNN. because ROI Different feature sizes , Different dimensions , Unable to send to the fully connected network , So use feature pooling , Fixed dimension .

4>RCNN 
  1>> take roi Connect to the fully connected network , Input rcnn Prediction score and prediction offset
   2>> Calculation rcnn Truth value
  3>>Rcnn Loss   
Classification and regression input dimensions 21 and 84
 

Fasterrcnn It is a two order algorithm , namely RPN\RCNN , It is necessary to calculate the loss , The former needs to provide the latter with regions of interest .

RPN Output anchor It's the forecast , anchor With the label iou\ Offset For the truth .

RPN Loss calculation ：

Predicted value and true value , Calculate the loss . Including classification and regression .

classification ： Just distinguish the background 、 prospects , Two classification , Cross entropy loss . Incoming score .
Return to ： The offset and truth values are large , Use 1 Order loss function , Easy to converge .

nms:
  stay RPN Step four , You will get more than 10000 points anchors  , But there will be multiple overlaps anchors, use nms Remove the overlapping box ( Just remove the overlapping boxes ), According to the score, select the front 2000 As the final proposal
 

Screening proposal obtain roi：
utilize proposal And label iou Calculation , elect 256 individual roi.