Utilizing image super-resolution to overcome information bottlenecks in vision transformers

Abstract

This research tends to solve the information bottleneck challenge in vision transformer-based solutions for image super-resolution, where the intensity of the feature map reduces in deeper network layers, thus affecting model performance. LITRL, the Layer-Interconnected Transformer with Residual Links, provides stability to the information flow by means of the dense residual connections between the layers, with the aim of preventing spatial information loss. The methodology involves the integration of the Swin transformer architecture and new schemes of interconnections to maintain vital spatial features in the whole network. Experimental results show that the LITRL-based method gives better results on traditional benchmark datasets (Set5, Set14, BSD100, Urban100, Manga109), in terms of quantitative (PSNR, SSIM) and qualitative evaluation. At 4×, LITRL obtains PSNR/SSIM of 40.37/0.9628 on Set 5 and 35.70/0.9408 on Urban100 with far higher performance than comparable methods. The proposed LITRL model dramatically reduces the information bottleneck of transformer-based super-resolution. It retains fundamental spatial information due to the dense-residual connections, giving rise to sharper images with more natural textures and fewer artefacts. Practical Implications: The excellent performance of LITRL in generating complex textures and structures that, in turn, enables accurate reconstruction, makes the method particularly useful for the tasks where the retention of a high level of fidelity of the image enhancement is imperative, i.e., for medical imaging, analysis of satellite images, and developing digital content while requiring a reasonable computational efficiency.