Convolutional neural network training method and device, image reconstruction method and device and medium

Abstract

The invention discloses a convolutional neural network training method and device, an image reconstruction method and device and a storage medium. The convolutional neural network comprises a featureextraction module, a nonlinear mapping module and an image reconstruction module, and the nonlinear mapping module comprises a plurality of connection groups and a first weighted channel cascade unit;the first weighted channel cascade unit carries out weighted integration on the information output by each connection group and then outputs the information, each connection group comprises a plurality of connection units and a second weighted channel cascade unit, and the second weighted channel cascade unit carries out weighted integration on the information input into the first connection unitand the information output by the last connection unit and then outputs the information. According to the method, all jump connections in a network level are in a weighted channel cascading mode, different weights are given to feature map channels, the difference between the channels is increased, integration of local features of a latter convolutional layer is realized, and the representation capability of the convolutional neural network is improved. The method is widely applied to the technical field of image processing.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a convolutional neural network training method, an image reconstruction method, a device and a storage medium. Background technique [0002] In the process of image acquisition and transmission, etc., due to interference such as noise, the quality of the image is impaired, resulting in a reduction in the resolution of the image. Using methods such as interpolation and reconstruction, super-resolution reconstruction of images can be achieved, thereby converting low-resolution images into high-resolution images. However, it is difficult for existing technologies to reconstruct details such as high-frequency information in images, and some newly emerging image super-resolution technologies based on deep learning face the shortcomings of large model representation redundancy and low representation ability. In order to obtain better reconstruction effects, Need to consume mo...

AI Technical Summary

Problems solved by technology

However, it is difficult for existing technologies to reconstruct details such as high-frequency information in images, and some newly emerging image super-resolution technologies based on deep learning face the shortcomings of large model representation redundancy and low representation ability. In order to obtain better reconstruction effects, Need to consume more resources

Images