Combining Variable Convolution and Transformer Network for Image Super-Resolution

In recent years, network models combining convolutional and Transformer architectures have shown strong performance in image super-resolution. However, these fusion models face two main challenges: they do not adequately consider the relationship between network depth and receptive field, and the complex structure of the Transformer module reduces computational efficiency. To address these issues, we propose a novel image super-resolution network, CVCT, which integrates variable convolution and Transformer. The Variable Convolution Group (VCG) expands the receptive field by using convolutional kernels of varying sizes, improving efficiency with fewer parameters. Additionally, we introduce the Simplified Transformer Group (STG), which reduces the complexity of normalization operations and simplifies the self-attention mechanism with a more efficient Simplified Self-Attention (SSA) approach. Experiments on five benchmark datasets show that CVCT outperforms or achieves competitive performance compared to state-of-the-art methods. Notably, at a 4× sampling factor, CVCT has 209k fewer parameters than SwinIR, yet improves PSNR by 0.18 dB on the Manga109 dataset and about 0.1 dB on Set5 and Set14. CVCT surpasses other methods in both quantitative and qualitative analyses, demonstrating its effectiveness and efficiency with fewer parameters.