Bit-level data enhancement method for image super-resolution

Abstract

The invention discloses bit-level data enhancement for image super-resolution. The invention aims to solve the problem that the existing data enhancement method for super-resolution is insufficient in picture information utilization. The bit-level data enhancement method comprises the steps of bit transformation data enhancement in image RGB channels, bit exchange data enhancement between the image RGB channels, and bit exchange data enhancement between high-resolution and low-resolution image channels. According to the method, hidden light sensation information is mined by changing an image bit plane, picture texture details and color information are enriched, and the correlation of high-resolution and low-resolution images is linked. The method mainly comprises the following steps: 1) acquiring an image bit plane; 2) carrying out image bit plane transformation; 3) carrying out image remodeling; and 4) carrying out model training. According to the method, the training set can be extended in an even manner under the condition that the data set is limited, and the super-resolution recovery effect is enhanced in the optimal bit plane transformation combination.

Description

technical field [0001] The invention belongs to the field of image processing and relates to a bit-level data enhancement method for image super-resolution. Background technique [0002] Image super-resolution (referred to as SR) refers to recovering high-resolution images (referred to as HR) from low-resolution images (referred to as LR). Since convolutional neural networks are applied to SR problems, SR based on deep learning has emerged. Different network architectures and training strategy methods have achieved amazing progress in SR tasks. But correspondingly, these models often need to train thousands of high-quality pictures to exert the excellent performance of the network structure. In previous studies, a large amount of data sets can be artificially synthesized through a systematic degradation method to meet the model requirements. However, in practical applications, the models trained by synthetic datasets often perform poorly, which is often caused by the fact t...

AI Technical Summary

Problems solved by technology

Some have the same shortcomings as DA for most high-dimensional vision tasks, that is, the processing process only stays on the pixels or features of the image data set, and the hidden information of many images is not fully utilized, and most of the images themselves are lost. Information

Images