Target detection model compression method based on knowledge distillation

Abstract

The invention provides a target detection model compression method based on knowledge distillation, and the method comprises the steps: extracting feature maps of a teacher network model and a student network model through an FPN, calculating the difference between Gram matrixes of the corresponding feature maps of the teacher network model and the student network model, and carrying out the reverse propagation. According to the method, the student network model can learn the phase degree between different channels from the teacher network model, so that the detection precision of the student network model is improved, the target detection model can be effectively compressed, and the detection precision is ensured while compression is performed. Wherein the teacher network model is a Faster RCNN taking ResNet101 as a backbone network, and the student network model is a Faster RCNN taking ResNet50 as a backbone network, so that about half of the number of middle layers is reduced, effective model compression is realized, and the detection precision of the compressed model is ensured by applying a Gram matrix.

Description

technical field [0001] The invention belongs to the technical field of image recognition, and in particular relates to a method for compressing a target detection model based on knowledge distillation. Background technique [0002] Target detection technology is mainly used to detect targets in images or videos, and the detection content includes the category of the target and the coordinates of the target. In recent years, the target detection algorithm based on deep learning has made great progress. The more popular algorithms can be divided into two categories, one is the R-CNN algorithm based on Region Proposal (including R-CNN, FastR-CNN, FasterR -CNN, etc.), they are Two-stage algorithms, which need to generate target candidate boxes in advance, and then classify and return the candidate boxes. The other type is One-stage algorithms such as YOLO and SSD, which can use a convolutional neural network (CNN) to directly predict the categories and positions of different ta...

AI Technical Summary

Problems solved by technology

However, there is still a lack of corresponding methods in the prior art. For the relevant technicians, there are still many technical difficulties in how to use knowledge distillation to compress the target detection model.

Images