Image super-resolution reconstruction method based on wavelet transformation and convolutional neural network

Abstract

The invention discloses an image super-resolution reconstruction method based on wavelet transformation and a convolutional neural network. The method comprises the following steps: at a training stage, carrying out Gaussian filtering and downsampling processing on a high-definition image Ih in a training data set to generate a low-definition image Il; carrying out single-scale two-dimensional discrete wavelet transform on the Ih to extract four frequency components thereof, such as a low frequency component FLL, a horizontal low-frequency vertical high-frequency component FLH, a horizontal high-frequency vertical low-frequency component FHL and a diagonal direction high-frequency component FHH; and with the Il serving as input data and the four frequency components of the Ih serving as labels, training four convolution neural network models; and in a super-resolution reconstruction stage, inputting the low-definition image Il into the four trained convolution neural network models to generate four frequency components of the high-definition image and carrying out single-scale two-dimensional discrete wavelet inverse transformation to generate a high-definition image Ih. The method carries out super-resolution reconstruction on the image at different frequencies, can make full use of learning capacity of the convolutional neural network, and substantially enhances the super-resolution reconstruction effect.

Description

technical field [0001] The invention relates to image super-resolution reconstruction technology, in particular to an image super-resolution reconstruction method based on wavelet transform and convolutional neural network. Background technique [0002] The process of reconstructing high-resolution images from one or more frames of low-resolution images using digital signal processing technology is called image super-resolution reconstruction. Image super-resolution reconstruction can solve the problem of low imaging resolution of hardware equipment due to the inherent resolution level limitation of digital image acquisition equipment. In addition, image super-resolution reconstruction can also solve the problems that affect image resolution due to various complex application environments, such as loss of image details, noise, and undersampling. Image super-resolution reconstruction can effectively improve the resolution of images while overcoming difficulties such as hardw...

AI Technical Summary

Problems solved by technology

The above classical image super-resolution reconstruction methods all need to learn image priors from large training data sets, but the learning ability of a single sparse dictionary model, convolutional neural network model or generative adversarial network model is not enough to accommodate the image priors in the training data set. test

Therefore, using a single machine learning model cannot achieve the optimal image super-resolution reconstruction effect

Images