PMDNet: A multi-stage approach to single image dehazing with contextual and spatial feature preservation

Abstract

Hazy images suffer from degraded contrast and visibility due to atmospheric factors, affecting the accuracy of object detection in computer vision tasks. To address this, we propose a novel Progressive Multiscale Dehazing Network (PMDNet) for restoring the original quality of hazy images. Our network aims to balance high-level contextual information and spatial details effectively during the image recovery process. PMDNet employs a multi-stage architecture that gradually learns to remove haze by breaking down the dehazing process into manageable steps. Starting with a U-Net encoder-decoder to capture high-level context, PMDNet integrates a subnetwork to preserve local feature details. A SAN reweights features at each stage, ensuring smooth information transfer and preventing loss through cross-connections. Extensive experiments on datasets like RESIDE, I-HAZE, O-HAZE, D-HAZE, REAL-HAZE48, RTTS and Forest datasets, demonstrate the robustness of PMDNet, achieving strong qualitative and quantitative results.