The unexpected scenarios remain a major issue for the autonomous driving vision system, especially in low-visibility circumstances caused by weather deterioration. Image dehazing is crucial for road navigation and vehicle motion tasks to ensure a safe and efficient procedure under quickly changing adverse weather conditions. In order to preserve texture information and improve the caliber of dehazing results, this study proposes a novel Quadruple Module Network model. The Dehazing Module (DM), Formula: see text-estimation module, Edge Module (EM), and Feature Merging Dense Module (FMDM) are the four key modules that are included in this proposed approach. The DM is initially made up of three Deep Level Feature Boost (DLFB) blocks, which use Atrous convolutional layers and self-attention techniques to create feature mappings between the fuzzy pictures and their related ground truth. The EM is then utilized to build a haze-free edge picture by connecting hazy images to the dehazed images. When combined with DLFB, the information obtained from the EM provides more organized information that is useful for the dehazing process. The Formula: see text-estimation module, which includes cascaded capsule layers for estimating the depth and matching haze level, then receives the output from the DM branch. The final dehazed image is obtained by feeding the output of the Formula: see text-estimation module to FMDM, which displays dense blocks, pixel attention, and feature attention. As a result, the proposed study improves picture visibility through an efficient dehazing procedure. The proposed approach outperforms certain current state-of-the-art methods, according to experiments done on a benchmark foggy road images dataset. PSNR of 77.64Formula: see textdB, SSIM and FSIM of 0.99, latency of 23.20Formula: see texts, MAE of 0.024, MSE of 0.001, and RMSE of 0.033 are the performances attained by the proposed network.
Deshmukh et al. (Sat,) studied this question.