Dual-Channel Residual Network for Image Super-Resolution Reconstruction

Aiming at the shortcomings of the existing deep learning based image super-resolution algorithms in image high-frequency detail reconstruction, a dual-channel residual network emphasizing image high-frequency detail reconstruction is proposed. Residual structure with channel attention mechanism is leveraged as the main channel of the network. Targeting on retaining more delicate geometric structure and edge information of the original image during the reconstruction, the auxiliary channel of the network is designed by an adaptive structured convolution hence the evolved dual-channel residual network has a stronger ability to capture high-frequency information during the learning process. To make the reconstructed image well coincide with the subjective visual experience of human eye, L1 loss function is combined with the multi-scale structural similarity loss function to train the network, so that the network completely retains the visual effect of the image during the training process. Experiments on the benchmark database show that combining the auxiliary channel based on adaptive structured convolution outside the main channel can heighten the peak signal-to-noise ratio of the reconstructed image by 2 dB. The simultaneous operation of L1 loss function and the multi-scale structural similarity loss function can heighten the peak signal-to-noise ratio of the reconstructed image by 3 dB and the structural similarity by 0.05. The objective and quantitative comparison with the competing networks exhibits the proposed network's outstanding effectiveness on two public data sets. © 2022, Editorial Office of Journal of Xi'an Jiaotong University. All right reserved.