MPRNET-Based Enhancement Technique For Image Restoration

Image deblurring is a challenging process in image restoration as it requires a delicate balance between preserving spatial details and utilizing high-level contextual information. In this paper, we present an approach for gradual restoration function learning based on the MPRNet model and a multi-stage architecture. To acquire contextualized features, we used an encoder-decoder architectures, which are then fused with a high-resolution branch that preserves local information. We presented per-pixel adaptive design with directly supervised attention, with each stage enhancing the re-weighting of local features. Furthermore, our multi-stage design allows for information exchange both sequentially and via lateral connections between feature processing blocks, preventing data loss. The resulting MPRNet model achieved significant performance gains across multiple datasets, highlighting its efficiency in image deblurring. Our enhanced MPRNet model demonstrated superior performance compared to other existing models in the literature by achieving 34.8 for PSNR and 0.996 for SSIM, highlighting its potential to advance image restoration techniques.