Dehazify: An AI Powered Haze Removal Application
Tanvir Singh, Akshi Dagar, Ayaz, Nidhi Gupta · 2026
The existence of fog degrades the visual quality of single images which impacts the working of various applications in computer vision. To overcome this challenge, this paper proposes a previously developed fog removal strategy which is based on Fog-Free Attention (FFA), an attention based deep learning framework, however, it has some problems like: difficulty in convergence, unstable training process or the edge details problem, especially in complex fog images. To overcome these issues, we develop a dynamic optimizer switching strategy that uses Adam in the early stage for adaptive learning and Stochastic Gradient Descent (SGD) while fine-tuning for stability. Hybrid method for improving convergence speed and model generalizations, custom L1 spatial-edge loss is applied to maintain the global structure and edge details. Our model has a PSNR of 24.41 which is better than traditional approaches.