Single image haze removal using CNN based encoder-decoder architecture

Haze is a natural phenomenon, which severely obscure the visibility of the distant objects makes it difficult for both the autonomous and human driving vehicles to take appropriate decision and may cause an accident. An efficient and robust solution in removing haze will help in reducing accidents. Recently, many prior based dehazing algorithms were proposed and doing fairly good but computationally intensive. In this contribution, we have proposed a computationally efficient encoder-decoder model based on up sampling and down sampling was used in convolutional neural network (CNN). The model was trained to update weight matrix using different datasets such as RESIDE and FRIDA for getting model to experience wide variety of data. Maxpooling and dropout layers were used to get advantage in both computation and for better generalization on new data. Objective analysis has shown that the proposed architecture doing relatively better in terms of SSIM & PSNR compared with the recent methods.