Image super-resolution reconstruction method based on cascade residual convolutional neural network

Abstract

The invention relates to the field of video and image processing. The objective of the invention is to effectively reduce the reconstruction difficulty of a high-resolution image, the good feature extraction capability of a convolutional neural network and the fitting capability of complex mapping. According to the image super-resolution reconstruction method based on the cascade residual convolutional neural network, the basic residual networks with the same structure are cascaded to form the cascade residual convolutional neural network, so that end-to-end mapping from inputting the low-resolution image to outputting the high-resolution image is realized; the basic residual network comprises a global residual channel and a feature extraction channel; down-sampling processing is carried out on the original high-resolution color image, so as to obtain a corresponding low-resolution image according to a basic residual network, performing bicubic interpolation upsampling on the low-resolution image to obtain an interpolation image, sending the interpolation image into a global residual channel of the basic residual network, and finally realizing information transmission between different levels of residual networks and forming final output. The image super-resolution reconstruction method is mainly applied to video and image processing occasions.

Description

technical field [0001] The invention belongs to the field of video and image processing, and relates to the improvement of image super-resolution reconstruction methods, the fusion of deep learning theory and image super-resolution reconstruction, and the implementation and application of residual convolutional neural networks in the field of high-resolution image reconstruction . Background technique [0002] Image super-resolution refers to the process of obtaining corresponding high-resolution images by using single or multiple low-resolution degraded image sequences. In many practical applications in the field of image processing, people often hope to obtain high-resolution original images, because high-resolution images mean higher pixel density, which can provide richer high-frequency detail information, so as to provide more information for the later stages of the image. Accurate extraction and utilization of processing and image information create a good foundation....

AI Technical Summary

Problems solved by technology

The loss function used in the above CNN network training is the minimum mean square error loss. Although it is beneficial to obtain a higher peak signal-to-noise ratio (PSNR), the subjective perception effect of the reconstructed image is poor. In order to solve such problems, from Ledig et al. of Twitter proposed the SRGAN [20] network, which completed the reconstruction of high-resolution images by constructing two sub-networks: the generation network and the discrimination network, and defined a new network loss function in it, making it include both traditional The minimum mean square error loss also includes a content loss function, thereby improving the human perception quality of the network reconstructed image, and correspondingly, the PSNR value of the reconstructed image has also decreased

Images