Contrast Aware Image Dehazing using Generative Adversarial Network

One of the most demanding aspects of computer vision is image dehazing. This paper presents a learning-based method to generate a dehazed image from a hazy input image using a densely connected end-to-end encoder-decoder-based GAN (Generative Adversarial Network) to facilitate feature extraction from images and their utilization. We added a Contrast-aware channel attention module to enhance and utilize the features extracted for generating more natural appealing dehazed images with better lightness computation, edges and structure detection. We used a traditional patch-based CNN type discriminator for pushing the generator to yield realistic results with regularized contrast and lower deformation of images. Moreover, we used cosine distance, a non-euclidean distance, and a weighted loss function to improve the perceptual quality of our output. Due to the unavailability of ground truth in real-world foggy images, we synthesized hazy images using a transmission map of the clear images for our dataset. On both synthetic and real-world foggy pictures, the results of our proposed model compete well with existing state-of-the-art techniques.