Single Image Haze Removal Based on transmission map estimation using Encoder-Decoder based deep learning architecture

Abstract

Haze removal is an essential requirement in autonomous vehicle applications for identifying different objects on the road. Most of the available techniques are based on different constraints/priors. The important parameters required for recovering the ground truth from hazy image are transmission map and air light. In this paper, we proposed a learning-based Encoder-Decoder deep learning architecture for transmission map estimation. Based on the assumption that at least twenty percent of the outdoor image includes with sky region and hence airlight is calculated as average of the twenty percent brightest pixels of the image. These two parameters namely transmission map and airlight were applied in atmospheric scattering model for ground truth image recovery. In Encoder-Decoder architecture, Max pooling layer, dropout layer was used for feature learning and efficient generalization respectively. The proposed architecture was trained on different datasets like NYU Depth data set, FRIDA and RESIDE Dataset for better generalization on unseen data. Experimental results shows that the proposed method has shown better performance compared to the existing state of the art methods.