LiteDHAZE: An Adversarial Dehazing Network for Robust Robotic Perception in Challenging Visual Conditions

Haze and fog severely degrade image quality, hindering reliable perception in robotic systems performing navigation, mapping, and object detection. We present LiteDHAZE, a lightweight generative adversarial network (GAN) for real-time single-image dehazing, leveraging edge-aware frequency decomposition and attention-guided enhancement. The architecture employs directional wavelet transform to extract high-frequency sub-band features and utilizes Res2Net-based multi-scale encoding to preserve structural details. A streamlined frequency-guided attention module reinforces both spatial and spectral feature relevance with minimal overhead. Unlike multi-branch frameworks, LiteDHAZE adopts a compact single-path encoder–decoder design that ensures low latency and strong generalization. Trained on the RESIDE dataset and evaluated using PSNR and SSIM, LiteDHAZE delivers competitive dehazing performance with superior efficiency, making it well-suited for embedded and real-time robotic vision systems.