MSFRNet: Multiscale Feature Recomposition Network for SingleImage Dehazing

The article addresses the critical issue of image dehazing in computer vision, which is essential for improving visibility and enhancing the performance of high-level tasks such as object detection and tracking. It begins by discussing the limitations of traditional physical prior-based dehazing techniques, which often suffer from distortion issues and struggle with complex real-world haze distributions. The article then introduces a novel deep learning model, MSFRNet, designed to efficiently extract, refine, and recombine features across multiple scales, ensuring improved haze removal while preserving fine structural details. The proposed model incorporates two key components: the Twin Prior Fusion Module (TPFM) and the Pyramid Dilated Split Attention Unit (PDSAU). The TPFM integrates Gradient-Based Prior (GBP) and Channel Residual Prior (CPR) to enhance feature separation and preserve edge and structural information. The PDSAU performs multi-scale feature extraction using dilated convolutions with different receptive fields, followed by a dual attention mechanism that selectively enhances important features while suppressing redundant information. The article presents extensive experiments on benchmark datasets, demonstrating that MSFRNet achieves higher PSNR and SSIM scores compared to prior methods. It also highlights the model's ability to generalize effectively across synthetic and real-world hazy conditions, making it a robust solution for real-time and resource-constrained applications. The qualitative and quantitative results showcase the superior performance of MSFRNet, making it a significant contribution to the field of image dehazing.