FUSION BASED GENERATIVE ADVERSARIAL NETWORK FOR SINGLE IMAGE DEHAZING

Abstract

Hazy environments significantly reduce the quality of digital images since haze varies with the scene depth. Consequently, the removal of haze from original images assumes great importance. Deep learning approaches have greatly improved the efficiency of the dehazing process for images. This work aims to provide a detailed analysis of various deep learning-based approaches employed for image dehazing. In this work, we proposed a model known as FAG-Net for image dehazing. We also proposed a new generator architecture which consists of a dense block, a transition block and a new feature attention (FA) block at each layer which enhance the realistic nature of the haze-free image. FA block consists of one channel attention (CA) and one pixel atten tion (PA) block. Channel attention aims to enhance the relevant information in di↵erent color channels of the input image, while pixel attention aims to selectively emphasize or suppress certain pixels in the image, which helps the models to concentrate on the most a↵ected areas of an image by haze. Perceptual loss and reconstruction loss are used along with adversarial loss to give more attention to the pixel which contains more haze and to maintain the realistic nature of generated haze-free image. Our FAG-Net trained on RESIDE(ITS), O-HAZE and I-HAZE datasets to conduct the experimental analy sis. Extensive experimental study demonstrates that our FAG-Net better than previous state-of-the-art methods.