Research on dehazing of surveillance images based on deep learning

Abstract

With the widespread application of surveillance systems in security, transportation, and other fields, image dehazing has become a key technology to improve the quality of surveillance images. This paper conducts research on surveillance image dehazing using the All-in-One Dehazing Network (AOD-Net) technology. The dataset consists of nearly 4,000 images, including the OTS subset from the RESIDE dataset and foggy images captured in real campus scenarios, which are divided into a training set and a test set at a ratio of 9:1. In terms of qualitative evaluation, the dehazed images conform to human visual perception, with details and colors close to the original images. For quantitative assessment, the mean Peak Signal-to-Noise Ratio (PSNR) reaches 15.25 dB, and the mean Structural Similarity Index (SSIM) is 0.839. Experimental results demonstrate that the AOD-Net technology achieves excellent dehazing performance for surveillance images.