Super-resolution model compression and acceleration method based on self-distillation contrast learning

Abstract

The invention discloses a super-resolution model compression and acceleration method based on self-distillation comparative learning, which is characterized in that a loss function based on comparative learning is introduced by adopting a self-distillation framework, a teacher model with a relatively large parameter quantity is trained, the first 1 / r channels of each layer are selected to form a student model, the parameters of the student model are shared with the teacher model; the reconstruction loss and the comparison loss are calculated by using the prediction output of the teacher model, the negative sample and the high-resolution image. Compared with the prior art, the method has plug-and-play performance, the authenticity of the generated image can still be ensured while compression and acceleration are carried out, the parameter quantity and the calculation quantity of the model are greatly reduced, all existing super-resolution models can be compressed and accelerated at the same time, the upper bound and the lower bound of a solution space are restrained by using comparison loss. The performance of the model is ensured while the parameter quantity of the model is reduced, so that the restored picture has higher authenticity.

Description

technical field [0001] The invention relates to the technical field of super-resolution of digital images, in particular to a super-resolution model compression and acceleration method based on self-distillation contrastive learning. Background technique [0002] With the advancement of digital image imaging equipment technology, the number of images acquired through cameras, tablets, mobile phones, surveillance and other equipment is increasing exponentially. Thanks to the rapid development of deep learning, people use these images to perform tasks such as image classification, semantic segmentation, and pedestrian re-identification, which greatly facilitates people's daily life. The performance of these computer vision systems is usually affected by the quality of the acquired images. However, due to the imperfection of imaging systems, atmospheric environments and processing methods, digital images will be lost or damaged to a certain extent during the process of formatio...

AI Technical Summary

Problems solved by technology

More and more scholars have begun to study the construction of a more sophisticated neural network structure to make full use of the information in the image, which has also led to a significant increase in the amount of memory and calculation required for the super-resolution algorithm based on deep learning.

[0004] In the existing technology, there are too many parameters and calculations in the super-resolution model, which has caused certain difficulties in the deployment of the super-resolution model.

Images