CasDyF-Net: Transforming Single-Image Dehazing through Federated and Adaptive CNNs

The Cascaded Dynamic Filter Network (CasDyF-Net) is introduced as a privacy-preserving framework for single-image dehazing, integrating Federated Learning (FL) with a lightweight, task-specific Convolutional Neural Network (CNN). Unlike centralized methods that require raw data transmission, CasDyF-Net ensures confidentiality by transmitting only model updates from edge devices. Its cascaded design incorporates multi-scale feature extraction, dynamic filters for adaptive haze removal, and progressive attention mechanisms to effectively address non-uniform haze patterns. A composite loss function, combining L1, Learned Perceptual Image Patch Similarity (LPIPS), and Structural Similarity Index Measure (SSIM), enhances perceptual fidelity beyond single-loss approaches. Evaluation on the RESIDE-6K dataset demonstrates performance with PSNR of 22.0, SSIM of 0.85, and LPIPS of 0.08, outperforming AOD-Net and SADnet while remaining competitive with FFA-Net and DR3DF-Net. By balancing privacy preservation, computational efficiency, and robustness to heterogeneous data, CasDyF-Net establishes a practical solution for real-world dehazing applications in domains such as intelligent transportation and remote sensing.