Ultra-light image super-resolution reconstruction method

Abstract

The invention discloses an extremely light image super-resolution reconstruction method, which effectively alleviates the problem of shallow information loss by constructing a dense feature fusion neural network skeleton. According to the invention, based on a distillation network mechanism, a multi-scale receptive field feature fusion architecture is designed, and richer and more diversified features can be extracted; for the feature extraction subunits, a feature extraction module which is more efficient and lighter is adopted, so that the efficiency of the whole network is greatly improved; according to the method, a very simple non-local module based on Hash mapping is designed, and the correlation between points is deeply mined at very low cost; in addition, spatial information, channel information and second-order information are ingeniously fused, and an attention module with better performance is obtained. The method provided by the invention is small in parameter quantity, small in calculation amount and high in precision, and exceeds all image super-resolution reconstruction methods below 800K at present.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to an extremely lightweight image super-resolution reconstruction method. Background technique [0002] Single image super-resolution aims to reconstruct a high-resolution image from a low-resolution input, but since a certain low-resolution image can be degenerated from multiple high-resolution images, it leads to uncertainty in the task and high difficulty. In order to better solve this problem, many methods based on convolutional neural networks appeared and replaced the previous traditional methods, and achieved good results. However, these methods often have a large number of parameters and expensive calculation consumption, so the lightweight super-resolution network is further proposed and studied. [0003] The early lightweight super-resolution network only used 20 layers for feature extraction, but its effect was only a little better than the traditional method. Then the l...

AI Technical Summary

Problems solved by technology

[0007] Aiming at the technical problems existing in the prior art, the present invention provides an extremely lightweight image super-resolution reconstruction method, the purpose of which is to solve the problem that the prior art is not light enough and efficient

Images