Image super-resolution optimization method based on similar image retrieval, medium and equipment

Abstract

The invention discloses an image super-resolution optimization method based on similar image retrieval, a medium and equipment, and belongs to the field of image data processing. The method comprises the steps that firstly, similar high-resolution images are retrieved from a high-resolution image material library; then extracting features of the similar high-resolution images, introducing an attention mechanism, taking a high-resolution intermediate image obtained by a bicubic interpolation algorithm as a query, extracting features related to the low-resolution images in the similar high-resolution images, and fusing the features related to the low-resolution images with the features of the low-resolution images after up-sampling to generate high-resolution images; and finally, taking the generated high-resolution image as a new query of the attention mechanism, and obtaining a final high-resolution image generation result through iteration. According to the method, related detail features of the high-resolution image similar to the input image are combined, feature expression of the low-resolution image is enhanced through an attention mechanism, and a high-resolution version of the low-resolution image is better generated.

Description

technical field [0001] The invention belongs to the field of image data processing, in particular to methods for image retrieval and image super-resolution in computer vision. Background technique [0002] Image super-resolution (SR) aims to reconstruct high-resolution images from low-resolution corresponding images. This task can restore high-resolution images from low-resolution images, greatly reduce the cost of manually drawing high-resolution image materials, and quickly provide high-resolution image materials. [0003] Unlike traditional model-based image super-resolution, current mainstream image super-resolution schemes are based on deep learning, especially some schemes based on convolutional neural network (CNN). Such models usually learn the mapping between low-resolution images and high-resolution images through large-scale datasets of high- and low-resolution image pairs. The current mainstream super-resolution direction is to improve by increasing the complex...

AI Technical Summary

Problems solved by technology

Since the input data is only a low-resolution image, the model does not intuitively receive the details of the high-resolution image during the process of upsampling the low-resolution image to generate a high-resolution image, and it is finally difficult to generate and real high-resolution image details comparable results

Images