Encoder decoder based CNN for single image dehazing with a semi-supervised approach
Muhammad Ismail, Zakir Ali, Salem Moqa, Jianfeng Lu · 2021 International Conference on Innovative Computing (ICIC) · 2021
In the present era of computer vision, the dehazing technique has a substantial role in numerous real-world applications, including outdoor surveillance, video-assisted conveyance, and support system for the driver. Previous approaches have utilized prior-based learning algorithms or methods depending on the light scattering model. These conventional techniques are not good in practice since they rely on statistical resources where the presumed priors do not hold for a particular appearance. We suggested a well-organized trainable end-to-end Convolutional Neural Network (CNN) with a semi-supervised framework independent of any statistical model to restore visibility in tempestuous weather conditions. The proposed deep learning model consists of decoupling CNNs and uses both synthetics and real-world hazy images as the training set. The supervised part is trained with multiple losses such as multi-structure similarity index measure-based loss and mean absolute error (MS-SSIM + 11) combined with perceptual loss. In contrast, the unsupervised part is trained with total variation loss. Specifically, we designed CNN with encoder-decoder architecture. In the reconstruction part of hazy images, we used attention mechanism to remove the distortion and artifacts in the focused part of the hazy images. We observed that the attention mechanism helps to achieve better results. Our wide range simulation shows that our designed framework performed compared to other state-of-art single image dehazing models on real-world data and synthetics hazy data.