Single Frame-Based Video Dehazing with Adversarial Learning

Vijay M. Galshetwar, Prashant W. Patil, Sachin Chaudhary

Abstract: Hazy environmental conditions degrade the quality of captured videos which leads to poor visibility and color distortion in videos. Such deterioration of captured video quality is mainly because of the attenuation caused by the scattering of light due to the haze particles present in the environment. In this paper, we propose an adversarial learning based single frame video dehazing encoder-decoder network. The proposed method comprises of Dilated Residual Block (DRB) used as encoder and Skip connection. DRB module is used to gain more contextual information by achieving large receptive fields. Skip connections are established between each encoder and decoder, which helps to detect and give more attention to haze specific features by adjusting the weights of learned feature maps automatically. This helps to extract the haze-relevant feature maps and recovers the haze-free video frame by using Channel Attention Block (CAB) and Residual Block (ResB) respectively. An extensive quantitative and qualitative analysis of the proposed method is done on benchmark synthetic hazy video database namely DAVIS-16 and NYU depth. Experimental result shows that the proposed method outperforms the other existing state-of-the-art (SOTA) approaches for video dehazing.