Generative adversarial network for intelligent haze removal from high quality images

Abstract

Suspended atmospheric particulates like haze, mist, and fog greatly degrade captured images, creating considerable challenges for computer vision applications operating in safety-sensitive areas such as autonomous driving, surveillance, and remote sensing. In this paper, we treat the important challenge of single-image haze removal by proposing a novel and robust conditional generative adversarial network (cGAN)-based framework. The proposal utilizes a U-Net-based generator with self-attention and skip connections to preserve spatial fidelity, and a PatchGAN discriminator to enforce local realism. At the heart of our contribution is a carefully weighted multi-component loss function that applies reconstruction, perceptual, edge, structural similarity (SSIM), and adversarial losses to optimize pixel-level accuracy and perceptual quality. We trained and evaluated our proposal on the large-scale real-world LMHaze dataset. Experimental results demonstrate state-of-the-art performance with a peak signal-to-noise ratio (PSNR) of 33.42 dB and SSIM of 0.9590. Our qualitative and comparative analyses further support our claims by assessing our proposed model's capacity to recover clear and artifact-free images from hazy images - outperforming the existing methods on this challenging real-world benchmark.