Arbitrary-multiple video super-resolution method introducing scale information

Abstract

The invention belongs to the field of super-resolution of computer vision, and aims to solve the problem of random magnification of super-resolution of a video and realize decimal magnification. According to the arbitrary-multiple video super-resolution method introducing scale information, firstly, adjacent frames are aligned at a feature level by using variable convolution; the method also includes fusing the features extracted by each element vector in the reverse position projection matrix by using a class attention mechanism; and finally, inputting the offset representing the movement of the pixel points between the continuous adjacent frames and the target frame into a position weight prediction module to obtain a position weight, and inputting the alignment features output by the alignment module into a feature weight prediction module to obtain a feature weight; and finally, correspondingly multiplying the two weights to obtain an up-sampling filtering kernel weight, and respectively multiplying the fusion feature by the up-sampling filtering kernel weight to realize amplification of any times. The invention is mainly applied to video super-resolution arbitrary-time amplification occasions.

Description

technical field [0001] The invention belongs to the super-resolution field of computer vision and relates to a method for realizing arbitrary magnification. Specifically, it involves an arbitrary magnification video super-resolution method that introduces scale information. Background technique [0002] Video super-resolution algorithm: an algorithm that expands the resolution of a video by a certain multiple. Video super-resolution usually inputs 2N+1 (N=1, 2, 3, ...) frame images, and the frame to be restored is called the target frame, and the frame that provides information for the target frame is called the adjacent frame. Video super-resolution relies on The timing information between adjacent frames and their own spatial information recover high-resolution frames. The process of video super-resolution algorithm can be divided into four parts: alignment, fusion, reconstruction, and upsampling. [0003] The difference between video super-resolution and image super-re...

AI Technical Summary

Problems solved by technology

[0012] In order to overcome the deficiencies of the prior art, the present invention aims to solve the problem of arbitrary magnification of video super-resolution, and realize fractional magnification

Images