A Brief Comment on Fast RCNN (Test) Algorithm for Target Detection

The fast RCNN detection part can be mainly divided into four modules:

1. Feature extraction: used to extract image features, usually the trunk; Such as vgg, resnet, mobilenet;

2.RPN (Regional Suggestion Network): used to generate candidate frames, mainly used for some rough classification and regression operations;

3.RoI Pooling: It mainly solves the problem that the full connection layer needs fixed input size, but the actual input size is different;

4. Classification and regression: fine classification and regression.

The general flow of the fast rcnn algorithm is as follows:

The trunk extracts color images and outputs the feature map of the last layer. Then these feature maps are further extracted based on 3×3 convolution kernel to enhance the robustness of the model. Send the output to two branches. The first branch is related to category, which is mainly used to simply classify and distinguish background and objects, and is aimed at anchor points; The second branch is used to preliminarily predict the offset of the candidate frame, which is also based on the anchor; Then send the results of the first two branches to the suggestion in the diagram. First, the first 6000 candidate frames will be screened according to the score of the positive category, and then the coordinates of the anchor point will be integrated with the obtained offset to get the preliminary candidate frame coordinates. Then, NMS will be carried out to remove the boxes with serious overlap, and then the top 300 boxes will be scored according to the category. Then the results are sent to roi pooing layer to generate a feature area with a fixed size, which is convenient for the following fully connected layer to accept information; Finally, the fully connected layer is used to extract features, and fine categories and frame offsets are obtained.