A super-resolution image reconstruction method based on WPSN and CNN

Abstract

The present invention provides a super-resolution image reconstruction method based on WPSN and CNN. The method comprises: constructing a training data set using LR image data and HR image data; constructing a new feature extractor wave packet scattering network WPSN, setting experimental parameters; constructing a new super-resolution image reconstruction model WPSN-CNN based on WPSN, setting model hyperparameters, training the network model WPSN-CNN through the training data set, and obtaining a trained WPSN-CNN network model; and performing super-resolution image reconstruction using the trained WPSN-CNN network model. The present invention introduces a wave packet system into WSN, removes the modulus and convolution operations with a low-pass filter in the WSN, and establishes a new feature extractor, so that the newly extracted features have translational exchangeability, and can make full use of the multi-directional, multi-scale, and multi-band features of the image. Using WPSN-CNN for super-resolution image reconstruction tasks can extract richer image features, thereby more accurately reconstructing the high-frequency structure and complex texture of the image, with faster convergence speed on small-scale data sets, significantly improved PSNR and SSIM indicators, and more realistic visual effects.

Images