Single Image Dehazing Using Deep Convolution Neural Networks

Haze removal is urgently desired in multi-media system. A deep learning-based method, called dehazingCNN, is proposed to estimate an approximate clear image. The proposed learning model is different from traditional learning based method. We adopts Deep Convolution Neural Networks (CNN) to take a hazy image as the input and outputs the corresponding clear image directly. The output of the network is high quality except some block artifacts and color distortions. We can remove the color distortion in the approximate clear image via atmospheric scattering model and guided filter effectively. Experimental results on different type of images, such as synthetic and benchmark of hazy images, demonstrate that the proposed method is comparative to and even better than many complex state-of-the-art methods.