RSAMSR: A deep neural network based on residual self-encoding and attention mechanism for image super-resolution

Abstract

This paper proposes an improved residual self-encoding and attention mechanism super-resolution (RSAMSR) network. Firstly, we construct a new structure through the multi-path convolution and design an attention mechanism module. Then the input data are divided into high and low-frequency components sent to the residual network with different depths for processing based on the spatial scaling theory. Finally, we introduce a self-encoding network to remove image noise. The model uses the L1 loss function for data training on the DIV2K data set and is compared with some state-of-the-art SR networks in four different public datasets of Set5, Set14, B100, and Urban100 under the magnification factor ×2, ×3, and ×4. Detailed experimental results show that the proposed model has fewer model parameters, the best objective criteria PSNR and SSIM, and the best subjective visual effect.