A multi-scale feature fusion method for super-resolution binocular image reconstruction

Abstract

The invention provides a binocular image super-resolution method for multi-scale feature fusion. First of all, in the feature extraction module, a hybrid skipping residual connection is used to form a loop structure with the improved empty space pyramid pooling block, which is used to extract multi-scale features of the image, and then the dilated residual and the loop structure are alternately cascaded to extract the The features are fused; then, the disparity attention module is introduced to obtain the corresponding relationship in the binocular image, and the useful information of the image pair is integrated. Among them, the network ability of the stereo matching feature is learned by using the transition expansion residual block, and the disparity map of the binocular image is obtained. ;Finally, four dilated residual blocks are used to map the low-dimensional space to the high-dimensional space, the super-resolution left (right) image is reconstructed by sub-pixel convolution, and FReLU is used in the entire network to improve the efficiency of capturing spatial correlation . The invention uses the expanded residual loop structure to extract multi-scale features of the image, realizes excellent super-resolution performance, and has wide applicability.

Description

technical field [0001] The invention relates to image processing technology, in particular, to a binocular image super-resolution reconstruction method using an expanded residual loop structure to extract and fuse multi-scale features. Background technique [0002] Visual information is the main source of information for human beings to obtain all things. The process of improving the original image resolution through software is called image super-resolution (SR) reconstruction. Image super-resolution technology meets the needs of people's perception. The visual field is booming. The research of image super-resolution has experienced leaps from the interpolation-based method to the learning-based method, from manual processing to artificial intelligence, from shallow network to deep network, and from deep network to lightweight network. Image super-resolution technology has made outstanding contributions to the development of human science and technology. Single-image supe...

AI Technical Summary

Problems solved by technology

[0004] The purpose of the present invention is to solve the problem of image super-resolution reconstruction, provide two low-resolution images of the left and right viewpoints in the same scene, and directly restore a high-resolution image through a deep learning network, which can be reconstructed Super-resolution left (right) image

Images