An image super-resolution reconstruction method based on wavelet coefficients learning

Abstract

The invention discloses an image super-resolution reconstruction method based on wavelet coefficient learning, belonging to the field of image super-resolution. the method includes selecting that image train set and the test set; obtaining The wavelet coefficients of LR image and HR image by two-dimensional stationary wavelet transform of low-resolution image and two-dimensional wavelet packet transform of high-resolution image respectively. 3, constructing a depth neural network; 4, extracting that depth network feature of the wavelet coefficients of the LR image by use the depth neural network, and obtaining the wavelet residual coefficients of the HR image; Step 5, restoring the obtained wavelet residual coefficients of the HR image to the wavelet coefficients of the HR, and carrying out the two-dimensional wavelet packet inverse transform to obtain the final high-resolution image. Compared with the prior art, the invention can perform super-resolution reconstruction on an image ina very short time by training a depth super-resolution model by using a large number of external training sets, and the reconstruction result is obviously superior to most algorithms.

Description

technical field [0001] The invention belongs to the field of image super-resolution, in particular to an image super-resolution method based on wavelet coefficient learning. Background technique [0002] Single image super-resolution reconstruction technology (Single Image Super-Resolution, SISR) refers to a given single low-resolution image, through a specific algorithm to restore the details of the image lost due to downsampling, so as to obtain more detailed information , The process of high pixel density images with more delicate picture quality. Due to the ability to recover finer details with limited information, this technology has been widely used in various application scenarios, such as high-definition television broadcasting, video surveillance, and satellite imaging. Since the downsampling process of a high-resolution image under different conditions will obtain a low-resolution image with different details, the reconstruction process from a low-resolution image...

AI Technical Summary

Problems solved by technology

2) For gradient disappearance / explosion, it is difficult to converge

[0005] However, the above CNN-based methods also have some shortcomings: 1) It is easy to produce over-smoothed (Over-Smoothed) results, lacking some detailed information; 2) Often use fixed upsampling operators, such as cubic spline interpolation operators, in Increasing the image resolution also introduces distortion information and increases the computational burden of the network; 3) It requires a lot of calculations and the training phase takes too long

Images