Local Multiscale Aggregation with a Global Attention Super-Resolution Network

In response to the issues of insufficient feature capture rooted in uniformly sized convolutional filters in super-resolution reconstruction algorithms, as well as the increased complexity of networks caused by the need to capture both global and local features, this research details a new model incorporating local multiscale aggregation and global attention for enhanced super-resolution. The algorithm adopts a deep residual structure, utilizing a receptive field block for shallow feature extraction. The deep feature extraction module reduces the parameter count through aggregation via the addition and dimension reduction of adjacent blocks while introducing channel attention at global positions to dynamically fine-tune the weighting of hierarchical features, thereby efficiently utilizing features from different depth blocks. Additionally, this paper designs a multiscale perception aggregation global convolution block as the basic block for deep feature extraction, extracting both local multiscale and global features. Finally, nonlocal sparse attention modules are introduced before and after the deep feature extraction module to efficiently aggregate local and nonlocal representations between pixels, achieving multilevel feature reuse. Compared with the EDSR and OISR networks, the proposed algorithm has only approximately one-fourth of the parameter count but achieves superior reconstruction performance. Compared with all contrastive methods in this paper, the devised method achieves average enhancements of 1.07 dB (2×) and 0.53 dB (4×) in the peak signal-to-noise ratio (PSNR) and average enhancements of 0.0098 (2×) and 0.0353 (4×) in structural similarity (SSIM). Extensive research evidence suggests that the presented method offers enhanced performance and reconstruction quality compared with similar methods.