A convolution neural network generation method and a super-resolution method of an image

Abstract

The invention discloses a convolution neural network generation method for performing super-resolution processing on an image, and a super-resolution method of an image, a computing device and a mobile terminal. The convolution neural network generation method includes constructing a first processing block, a second processing block and a third processing block respectively, wherein the first processing block comprises a first convolutional layer, the second processing block comprises a second convolutional layer and the third processing block comprises a third convolutional layer; constructing a fourth processing block, the fourth processing block comprising a fourth convolution layer; connecting one or more first processing blocks, a second processing block, a third processing block anda fourth processing block according to a preset connection rule to generate a convolution neural network with the first processing block as an input and the third processing block as an output; according to the pre-acquired image data set, training the convolution neural network so that the output of the convolution neural network indicates the super-resolution image corresponding to the input image.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to a method for generating a convolutional neural network for performing super-resolution processing on images, a method for super-resolution of images, a computing device, and a mobile terminal. Background technique [0002] Super-resolution is a technology to improve the resolution of the original image. The process of obtaining a high-resolution image through a series of low-resolution images is super-resolution reconstruction. Super-resolution technology based on deep neural network has attracted widespread attention because it has achieved far more effects than traditional super-resolution technology. [0003] The existing super-resolution technology based on deep neural network can be roughly divided into two types. The first one is to enlarge the original image to the target size with traditional methods, and then use deep neural network to process the enlarg...

AI Technical Summary

Problems solved by technology

The first way can get the target image of any size, but because the image has been enlarged to the target size before it is sent to the deep neural network, it will significantly increase the calculation amount of the deep neural network

The second method extracts image features on the size of the original image, and the calculation amount of the deep neural network will be relatively small. However, due to the limitation of Sub-Pixel technology, this method can only obtain an integer multiple magnified image. If the target image size is not the same as the original image Integer multiples, additional processing is required in the subsequent processing part, which is more cumbersome

Images