Progressive Multi-Scale Residual Network for Single Image Super-Resolution

Multi-scale convolutional neural networks (CNNs) achieve significant success in single image super-resolution (SISR), which considers the comprehensive information from different receptive fields. However, recent multi-scale networks usually aim to build the hierarchical exploration with different sizes of filters, which lead to high computation complexity costs, and seldom focus on the inherent correlations among different scales. This paper converts the multi-scale exploration into a sequential manner, and proposes a progressive multi-scale residual network (PMRN) for SISR problem. Specifically, we devise a progressive multi-scale residual block (PMRB) to substitute the larger filters with small filter combinations, and gradually explore the hierarchical information. Furthermore, channel- and pixel-wise attention mechanism (CPA) is designed for finding the inherent correlations among image features with weighting and bias factors, which concentrates more on high-frequency information. Experimental results show that the proposed PMRN recovers structural textures more effectively with superior PSNR/SSIM results than other small networks. The extension model PMRN$^+$ with self-ensemble achieves competitive or better results than large networks with much fewer parameters and lower computation complexity.