NDPC-Net: A dehazing network in nighttime hazy traffic environments

Nighttime haze seriously affects the objects detection and recognition for autonomous vehicles. Although great progress has been made in daytime dehazing research, it is not suitable for night traffic scenes. Lack of large-scale training datasets of nighttime haze images is one of the main reasons that hinders the neural network model to achieve satisfied dehazing effects. To address it, we propose a novel method to synthesize nighttime traffic dehazing dataset: NTHAZE, and also a dehazing network: NDPC-Net, which includes a Nested DDU-Net block and a Pyramidal convolution block. The former extracts hazy images' features of different complexity, the latter retains and restores spatial features of different scales. Experiment results show that NDPC-Net can effectively dehaze in low illumination and nighttime traffic environment scenes. The dehazing effect is natural and smooth in color and texture, in NH-HAZE, PSNR and SSIM up to 32.31dB and 0.8822 respectively, 67% higher than SOTA. The precision and recall of objects detection after our proposed dehazing are improved by 22.66% and 4.66% respectively compared with SOTA.