A Deep-learning-Based Dehazing Framework for Non-Homogenous Scenes

The advancement of single-image dehazing techniques has been rapid in recent years. Several existing algorithms that are based on deep learning have shown remarkable efficiency for dealing with homogeneous hazing-free issues, but convolutional neural networks (CNNS) frequently fail on non-homogeneous dehazing datasets. Meanwhile, dehaze results from dense haze regions are often blurry because the information of these regions is typically unknown and difficult to estimate. To address these issues, an efficient image enhancement dehazing algorithm that utilizes deep learning techniques, and a non-uniform atmospheric scattering model had been proposed. Unlike the majority of existing dehazing methods, the medium transmission function was automatically computed using the transmission map estimation module from hazy images only as a regularizer without depth map data and dehaze images, allowing training the network on natural images. To guide network training, a mean square error 1oss function was utilized for calculating the difference between the enhanced image and the target image. Using real-world datasets, the proposed model was compared with several state-of-the-art approaches to evaluate its effectiveness. The suggested method enhanced color and details and achieved higher scores in the two full-reference metrics while reducing network size, as demonstrated in the experimental results.