Lightweight hierarchical residual feature fusion network for single-image super-resolution

Abstract

In recent years, numerous lightweight convolution neural networks (CNNs) have made remarkable progress for single image super-resolution (SISR) and showed great power for image reconstruction under constrained resources. However, existing lightweight networks can not fully utilize the informative hierarchical features, which will lead to the degradation of network reconstruction. To alleviate this issue, we propose a hierarchical residual feature network named HRFFN. Specifically, we design an enhanced residual block (ERB) containing multiple mixed attention blocks (MABs) to boost the representative ability of the network. Compared with ordinary residual blocks, ERB can achieve better performance while reducing network parameters and computational complexity. To utilize more features from intermediate convolution layers, we introduce a hierarchical feature fusion strategy (HFFS) to efficiently fuse the detailed information from each ERB step by step. By fully utilizing the hierarchical details with this strategy, we can refine the hierarchical features more efficiently. Besides, we cooperate the global dense connection strategy (GDCS) and residual learning connection (RLC, at low, meditate, and high levels) to construct our HRFFN. By employing these strategies, we can maximize the utilization of hierarchical features with a slight increase in parameters. Comprehensive experiments show the superiority of our method on five benchmark datasets against other state-of-the-art methods, which achieves a comparable trade-off between visual quality and quantitative metrics.