Depth-guided Hybrid Attention Swin Transformer for Physics-guided Self-supervised Image Dehazing

Image dehazing is a critical preprocessing step in computer vision, enhancing visibility in degraded conditions. Conventional supervised methods often struggle with generalization and computational efficiency. This paper introduces a self-supervised image dehazing framework leveraging a depth-guided Swin Transformer with hybrid attention. The proposed hybrid attention explicitly integrates CNN-style channel and spatial attention with Swin Transformer window-based self-attention, enabling simultaneous local feature recalibration and global context aggregation. By integrating a pre-trained monocular depth estimation model and a Swin Transformer architecture with shifted window attention, our method efficiently models global context and preserves fine details. Here, depth is used as a relative structural prior rather than a metric quantity, enabling robust guidance without requiring haze-invariant depth estimation. Experimental results on synthetic and real-world benchmarks demonstrate superior performance, with a PSNR of 23.01 dB and SSIM of 0.879 on the RESIDE SOTS-indoor dataset, outperforming classical physics-based dehazing (DCP) and recent self-supervised approaches such as SLAD, achieving a PSNR gain of 2.52 dB over SLAD and 6.39 dB over DCP. Our approach also significantly improves object detection accuracy by 0.15 mAP@0.5 (+32.6%) under hazy conditions, and achieves near real-time inference (≈35 FPS at 256x256 resolution on a single GPU), confirming the practical utility of depth-guided features. Here, we show that our method achieves an SSIM of 0.879 on SOTS-Indoor, indicating strong structural and color fidelity for a self-supervised dehazing framework.