Single Image Dehazing using a Channel and Pixel Attention Network

In this work, a novel feature focus network is proposed to recover image information lost due to fog in an image. The network architecture can be divided into three main categories: 1) A feature attention block that combines the channel and pixel attention layers of the neural network to decompose images. Different channels of the image carry different weighted information about the intensity of fog at each pixel in the image. Considering this fact, this network provides additional robustness when trying to decompose images by treating different features and pixels differently. 2) The underlying structure of this network comprises feature attention and residual learning. Residual learning is used as a solution to the vanishing gradient problem faced by deep neural networks. It has now become one of the basic requirements of deep neural network. 3) Feature weights in this network are learned to prioritize salient features and adapt from the feature focus module. The proposed network outperforms previous dehazing methods in terms of structural similarity index and peak signal-to-noise ratio, as demonstrated by experimental results.