MSDB-based CNN architecture for image dehazing in driverless cars

With the booming development of image processing technology and computer vision technology, scene detection and image processing in special weather has become an important research direction in this field. Among them, images taken in foggy days are easily affected by fog or haze, resulting in blurred details and low contrast to the loss of important image information, and to solve such problems image defogging algorithms are born. To address these challenges, a lightweight convolutional neural network based on multi-scale dense connectivity, called MSDL, is proposed in this paper for reconstructing blurred images. DehazeNet is the End-to-End defogging system that takes the fogged image as input, its transmission map as output, and then uses an atmospheric scattering model for image reduction. The proposed MSDL uses the transformed atmospheric scattering model to jointly estimate the transmission map and atmospheric light. In addition, a novel feature extraction module MSDB is proposed. Finally, extensive experiments are carried out using synthetic and natural hazy images. The experimental results show superiority over both non-deep learning and deep learning methods in both qualitative and quantitative evaluation. The PSNR, SSIM and MSE metrics were measured on different datasets, and the advantages were obtained on NYU2 dataset.