Video super-resolution reconstruction method

Abstract

The invention discloses a video super-resolution reconstruction method, including a non-local space-time network and a progressive fusion network. In the non-local space-time network, input multiple frames are fused together to form a whole high-dimensional feature tensor graph, and the whole high-dimensional feature tensor graph is deformed, separated, calculated and extracted to obtain a multi-frame video mixed with the non-local space-time correlation. And in the progressive fusion network, multiple frames output by the non-local space-time network are sent into the progressive fusion residual block, and the space-time correlation among the multiple frames is gradually fused. And finally, the fused low-resolution feature tensor graph is amplified to obtain a final high-resolution videoframe. According to the video super-resolution reconstruction method, the spatial-temporal correlation among multiple frames is effectively fused, and rich texture details can be recovered while the video resolution is enhanced.

Description

technical field [0001] The invention belongs to the technical field of digital image processing, and relates to a video super-resolution reconstruction method, in particular to a video super-resolution reconstruction method based on non-local spatiotemporal correlation and progressive fusion network. Background technique [0002] With the development of science and technology, people's demand for high-definition video is also increasing, from 720P (1280×720) to 1080P (1920×1080) high-definition, to 4K (3840×2160) or even 8K (7680×4320) Ultra HD. Therefore, as a technique that can generate high-resolution video from a given low-resolution video, video super-resolution has an important use. At present, video super-resolution is widely used in fields such as video surveillance, digital television, and satellite remote sensing. [0003] Traditional super-resolution methods have interpolation-based methods, such as bicubic interpolation methods. These methods compute from a gi...

AI Technical Summary

Problems solved by technology

This method only needs to be processed once per frame, which is faster, but it cannot use the information of the next delayed frame to help restore the current frame

Jo et al. proposed a video super-resolution network based on 3D convolution. The network uses 3D convolution, which can make better use of the spatiotemporal correlation between frames, but has the disadvantages of huge calculation and low efficiency.

Images