Image Super-Resolution via Lightweight Attention-Directed Feature Aggregation Network

Abstract

The advent of convolutional neural networks (CNNs) has brought substantial progress in image super-resolution (SR) reconstruction. However, most SR methods pursue deep architectures to boost performance, and the resulting large model sizes are impractical for real-world applications. Furthermore, they insufficiently explore the internal structural information of image features, disadvantaging the restoration of fine texture details. To solve these challenges, we propose a lightweight architecture based on a CNN named attention-directed feature aggregation network (AFAN), consisting of chained stacking multi-aware attention modules (MAAMs) and a simple channel attention module (SCAM), for image SR. Specifically, in each MAAM, we construct a space-aware attention block (SAAB) and a dimension-aware attention block (DAAB) that individually yield unique three-dimensional modulation coefficients to adaptively recalibrate structural information from an asymmetric convolution residual block (ACRB). The synergistic strategy captures multiple content features that are both space-aware and dimension-aware to preserve more fine-grained details. In addition, to further enhance the accuracy and robustness of the network, SCAM is embedded in the last MAAM to highlight channels with high activated values at low computational load. Comprehensive experiments verify that our proposed network attains high qualitative accuracy while employing fewer parameters and moderate computational requirements, exceeding most state-of-the-art lightweight approaches.