A Residual Network with Efficient Transformer for Lightweight Image Super-Resolution

Abstract

In recent years, deep learning approaches have achieved remarkable results in the field of Single-Image Super-Resolution (SISR). To attain improved performance, most existing methods focus on constructing more-complex networks that demand extensive computational resources, thereby significantly impeding the advancement and real-world application of super-resolution techniques. Furthermore, many lightweight super-resolution networks employ knowledge distillation strategies to reduce network parameters, which can considerably slow down inference speeds. In response to these challenges, we propose a Residual Network with an Efficient Transformer (RNET). RNET incorporates three effective design elements. First, we utilize Blueprint-Separable Convolution (BSConv) instead of traditional convolution, effectively reducing the computational workload. Second, we propose a residual connection structure for local feature extraction, streamlining feature aggregation and accelerating inference. Third, we introduce an efficient transformer module to enhance the network’s ability to aggregate contextual features, resulting in recovered images with richer texture details. Additionally, spatial attention and channel attention mechanisms are integrated into our model, further augmenting its capabilities. We evaluate the proposed method on five general benchmark test sets. With these innovations, our network outperforms existing efficient SR methods on all test sets, achieving the best performance with the fewest parameters, particularly in the area of texture detail enhancement in images.