A Computational Method for Single Image Super-resolution with Dual-Channel Convolutional Neural Networks

Abstract

The invention discloses a single image super-resolution calculation method for a dual-channel input convolutional neural network, which includes (1) subjecting a known high-resolution image to fuzzy degradation processing to obtain a low-resolution image of the same size; (2) The blurred low-resolution image obtained in step (1) is decomposed into the texture part and the smooth structure part of the low-resolution image; the texture part and the smooth structure part of the high-resolution image are obtained; (3) the texture part obtained in step (2) is The low-resolution texture part and the original low-resolution image are combined to obtain a dual-channel input, and the output of the high-resolution texture part is obtained; (4) The high-resolution texture part output obtained in step (3) is combined with the original low-resolution image to obtain the final The image super-resolution reconstruction result is obtained to complete the super-resolution reconstruction; (5) Calculate the difference between the high-resolution texture parts obtained in step (4) and step (2) to obtain the texture part loss; minimize the texture loss and image The sum of losses is used to optimize the network structure parameters.

Description

technical field [0001] The invention belongs to the technical field of image processing, and in particular relates to a single-image super-resolution calculation method of a dual-channel convolutional neural network, which is applicable to various computer vision tasks, such as face recognition, target tracking, license plate recognition and the like. Background technique [0002] Single image super-resolution (SR, Super-Resolution) technology refers to the process of restoring a low-resolution (LR, Low-Resolution) image into a high-resolution (HR, High-Resolution) image by software. This technology has a wide range of applications, such as video surveillance, medical imaging, remote sensing satellite imaging, etc. Existing SR rate algorithms can be divided into three categories: interpolation-based [1], reconstruction-based [2] and learning-based methods [3-10]. Among them, the image SR algorithm based on learning learns the functional mapping relationship between LR and H...

AI Technical Summary

Problems solved by technology

The SRCNN model proves that it is simple and feasible to directly learn the end-to-end mapping between LR and HR, and the effect is also very good, but the reconstructed high-frequency features are still not satisfactory

Images