Knowledge distillation method based on semantic segmentation intra-class feature difference

Abstract

The invention discloses a knowledge distillation method based on semantic segmentation intra-class feature difference, and aims to migrate dark knowledge learned by a complex model (teacher model) toa simplified model (student model), thereby maintaining the speed of a semantic segmentation model while improving the accuracy of the semantic segmentation model. The method includes: firstly, obtaining convolution features through a teacher model and a student model respectively; then, obtaining a feature map of each category center through average pooling operation guided by a mask, and calculating feature similarity between each pixel point and the corresponding category center to obtain an intra-category feature difference map; and finally, aligning the intra-class feature difference graph of the student model with the teacher model so as to achieve the purpose of improving the accuracy of the student model. Compared with the prior art, the distillation method provided by the invention is novel in thought, the obtained semantic segmentation model achieves a good effect in the aspects of accuracy and speed, and meanwhile, the method can be conveniently combined with other related technologies and has a very high practical application value.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and more specifically relates to a knowledge distillation method based on feature differences within semantic segmentation classes. Background technique [0002] Semantic segmentation refers to the pixel-level understanding of image scenes through computers, and is a basic research direction in the field of computer vision. In recent years, with the rapid development of fully convolutional neural networks, the accuracy of semantic segmentation models has been continuously improved, but most of these models require a large amount of computing resources, which also limits their applications in real life, such as autonomous driving, virtual reality , robots, and more. [0003] In order to solve this problem, model compression is a common idea, which can usually be divided into the following three categories: quantization, pruning and knowledge distillation. Among them, the idea of ​​knowle...

AI Technical Summary

Problems solved by technology

However, the student model obtained by these existing methods is often difficult to align with the teacher model in terms of intra-class feature differences, which also limits the improvement of the accuracy of the student model.

Images