| dc.description.abstract | Atmospheric phenomena like haze, fog, and smoke degraded image visibility. Asa result, there is less contrast, colour distortion, etc. in the obtained image. Inremote sensing, computer vision, and photography, haze reduction is greatly desired.in photography, dehazing can increase the visibility and quality of outdoorimages and landscapes, making them more vibrant and appealing. In computervision, dehazing can improve the quality of object detection task, recognition,and tracking algorithms, especially in outdoor and low-light environments. Inremote sensing, dehazing can improve the quality of satellite and aerial images,making them more useful for environmental monitoring, disaster management,and urban planning. Dehazing removes haze, improves scene vision, and adjuststhe airlight�s colour change. There are two types of haze, homogeneous andnon-homogeneous. To increase the dehaze quality for both homogeneous andnonhomogeneous haze, several techniques were used. Methods �classical, Deeplearning-based, and GAN-based. For non-homogeneous haze, it is challenging toestimate the spread of haze. Due to the less availability of real world ground-truthimages, many recent methods focus on the unsupervised approach to solve thisissue. GAN, and cycle-GAN based unsupervised methods are highly used in thistechnique. But still, there is not any prominent unsupervised technique for nonhomogeneoushaze removal. This paper proposed the unsupervised cycle-GANbased approach, which has worked on both homogeneous and non-homogeneoushaze. Specifically, we use cycle-GAN with non-homogeneous and homogeneoushaze removal generator. Generator use modified Unet with pixel, channel attentionand pretrained resnet as haze removal. Overall proposed architecture givesbetter results for both non-homogenous and homogeneous images compared tothe existing unsupervised methods. | |
