DeepVideoDehazeNet: A Comprehensive Deep Learning Approach for Video Dehazing Using Diverse Datasets

Abstract

Video dehazing is a technique commonly used to enhance the quality of videos that appear hazy or degraded due to factors like air scattering and light absorption. Unlike working with individual frames, video-based approaches leverage information from neighboring frames to achieve better dehazing results. This study proposes a straightforward yet powerful real-time video dehazing method utilizing a Convolutional Neural Network (CNN). The process involves dividing the video into frames, dehazing each frame, and merging them to produce a clear video output. To train the network, a dataset comprising synthetic hazy videos and haze-free reference videos is created using various datasets such as NYU depth, NYU, D-HAZY, NH-HAZE, and RESIDE. The forward half of RES2NET is used as an encoder, while an Image generator, CNN, is employed to generate dehazed images. The study's findings show how well the suggested strategy clarifies haze from outdoor sceneries in synthetic and real-world videos. In terms of dehazing performance, it performs better than current cutting-edge techniques. The proposed CNN-based video dehazing model demonstrates strong performance, achieving an average SSIM of 0.987 and PSNR of 38.86 across multiple datasets. Video dehazing has many uses including medical imaging, surveillance imaging, underwater imaging, and outdoor imaging.