Abstract: There are problems of distorted colors and blurred edges in the results of the state-of-the-art image dehazing algorithms. For solving the problems，an image dehazing algorithm based on deep learning is proposed. The proposed algorithm consists of two modules: attentional feature fusion module and haze model parameter estimation module. Attentional feature fusion module is designed to extract the color and edge features of hazy images sufficiently. Haze model parameter estimation module based on densely connected dilated convolution auto encoder is used to estimate the parameter of haze model and deal with the network degeneration in image dehazing. Experiments on images with thin haze and thick haze show that the proposed algorithm performs well on image dehazing， and the proposed dehazing algorithm has higher structural similarity (SSIM)， lower mean-square error (MSE)， lower edge error e_○edge than the state-of-the-art image dehazing algorithms.

Key words: attentional mechanism； feature fusion； densely connected； dilated convolution； auto encoder