Widely-activated network merging perceptual loss via discrete wavelet transform for image super-resolution

There are some problems with image super-resolution techniques, such as insufficient utilization of image features in different layers and lack high frequency information resulting in blurred texture edges in reconstructed images. A new method for the image super-resolution reconstruction of Widely Activated Network Fused Perceptual Loss (P-WAN) is proposed, which can improve the usage of image features from different layers. The method uses a pre-training model to extract features before activation to obtain the perceptual loss, meanwhile, draws on the adversarial loss in the adversarial generation network, and combines the pixel loss of the image to form a new loss function. Finally, we optimize the loss function by adjusting the weights of the three loss terms. Based on this research, a Widely Activated Network Based on Discrete Wavelet Transform and Fused Perceptual Loss (DP-WAN) is further proposed, which can reconstruct better high frequency information and higher quality texture edges. The method mainly adds the discrete wavelet transform on the basis of P-WAN, trains the different components obtained by the transform separately, and finally reconstructs the super-resolution image through the inverse discrete wavelet transform. To validate the feasibility and effectiveness, 4 representative methods are selected to test on 5 datasets. Experimental results have shown that the proposed method achieves the best performance in objective evaluation, and can obtain a good visual experience in subjective visual evaluation.