Target detection series fast r-cnn

 

Author's brief introduction ： Bald Sue , Committed to describing problems in the most popular language

Looking back ： Target detection series —— The work of the mountain RCNN The principle,

Near term goals ： Have 10000 fans
Support Xiao Su ： give the thumbs-up 、 Collection 、 Leaving a message.

 

Target detection series ——Fast R-CNN The principle,

Write it at the front

   In the last article , We introduced RCNN Principle , Detail stamp *** Here's a brief overview RCNN The algorithm steps of ：

1. Candidate area generation
2. Feature extraction by neural network
3. SVM Classifier classification
4. The regressor corrects the position of the candidate box

---

   Following RCNN After the release of ,RGB The great God published again Fast R-CNN Let's take a look at the expression in the paper Fast R-CNN Pictures of the results , as follows ：

   About this classic picture , Now you just need another intuitive feeling , In depth analysis will follow . This article is more RCNN With a big hint ,Fast R-CNN The main steps are as follows ：

1. Candidate area generation
2. Full image input network , The candidate frame is projected onto the feature graph to obtain the feature matrix
3. The characteristic matrix is ROI pooling Layers are scaled to uniform size , Then flatten the characteristic map to get the prediction result

   You can see , Only RCNN and Fast R-CNN From the steps of , There are still some differences between them , The details of these steps will be discussed below , Let's see

 
 

Candidate area generation

   Candidate region generation and R-CNN There is no difference between , The same is true of SS Algorithm , I won't repeat it here , What is not clear can be referred to Part 1 R-CCN post

 

 

Full image input network , The candidate frame is projected onto the feature graph to obtain the feature matrix

   Remember when R-CNN What do we input into the network ？ No selling here , stay R-CNN The loser is Jing SS The algorithm gets 2000 Candidate box , This obviously requires a huge amount of computation ; And in the Fast R-CNN in , We only need to input the original image into the feature extraction network to get the feature map of the original image . Here you may have such a problem ： Since the input network is the original image , How to use the candidate boxes generated in the original image in the first step ？ In fact, this part is based on hekaiming's SPP-Net—— A candidate frame in the original image will be mapped to the corresponding position of the resulting feature map after passing through the neural network , This position is computable . For your convenience , I drew the following figure for your reference ：

   About the above mapping rules , You can refer to this blog ：SPP-net

 
 

The characteristic matrix is ROI pooling Layers are scaled to uniform size , Then flatten the characteristic map to get the prediction result

   stay Fast R-CNN in , We're not like R-CNN Forced scaling of the picture as in , But after we get the mapping on the characteristic graph （ That is, the candidate box ）, Compare these candidate boxes ROI pooling Operation to uniformly scale candidate boxes of different sizes to a uniform size ,ROI pooling The operation is shown in the figure below ： That is, regardless of the size of the original feature map , We all divide the feature map into 7*7=49 Equal parts , Then use maximum pooling or average pooling for each serving , Down sampling the original feature map into 7*7 Uniform size .

   After the feature map becomes a unified size , It can be flattened and sent to the full connection layer , Then connect softmax Layer and the regressor Layer can output .

 
 

Fast R-CNN Complete process

   Through the above description , Now I believe that if you look at this picture again, it will be more profound . The general process is consistent with the above , There is no narration here

 
 

Loss function

   The loss function consists of two parts , Part is classified loss , One part is the boundary box regression loss .

 
 

Summary

  Fast R-CNN That's all for the principle of , I hope it can help you . It will be updated later Faster_RCNN And related code explanation , Come on ！！！

 
 

Reference link

RCNN A collection of theories

Deep learning _FasterRCNN Detailed explanation of the paper

 
 
If the article is helpful to you , It would be

Whew, whew, whew ~~duang~~ A great bai

Those interested in artificial intelligence can scan the code to join C Station artificial intelligence official Community , Welcome to join the exchange study , Progress together ！！