DNASM -TID: Atmospheric scattering model construction based on light source distance and image dehazing for nighttime traffic scenes

Nighttime traffic scene image dehazing has an important application value for nighttime traffic driving and security monitoring activities. Aiming at the existing nighttime image dehazing optical model for the non-uniform distribution of artificial light sources is not enough to consider the problems of halo and uneven illumination in the dehazed image. A nighttime image imaging atmospheric scattering model based on the distance of the light source is constructed. An image dehazing algorithm for nighttime traffic scenes is proposed based on the constructed optical model. Firstly, the near light source region in the foggy image at nighttime is segmented based on Gaussian pyramid, and in order to deal with the scattering effect of artificial light source more accurately, a discriminative index for the proximity of scene pixels to the center of the light source is proposed. Then, the atmospheric light matrix conforming to the regional distribution of light sources is estimated using Gaussian low-pass filtering, and to obtain a more accurate atmospheric light matrix, a channel map is constructed as a guided map by utilizing the luminance and saturation components of the image, and fast guided filtering is carried out on the preliminarily obtained atmospheric light matrix. Subsequently, the transmittance matrices of the near-source and far-source regions are solved using the bright and dark channel a priori, respectively. The optimal fusion model with L₂ norm regularization is established to solve the fused transmittance matrix. To avoid the block partition effect in the process of transmittance fusion, anisotropic Gaussian filtering is carried out on the fused transmittance matrix. Finally, a dark-adaptive adjustment function is designed to adaptively adjust the image's brightness level after dehazing. The experimental results show that the proposed nighttime traffic scene image dehazing algorithm dehazes the image after the subjective visual effect of dehazing more thoroughly, at the same time, effectively mitigates the problem of uneven halo and luminance distribution. Compared with other nighttime image dehazing algorithms, it shows substantial superiority in subjective and objective evaluation experiments.