Dual Attention Fusion Network for Single Image Dehazing

Abstract

Recently, there have been significant advances in image dehazing methods based on the convolutional neural network. However, the convolutional kernel is a fixed receptive field. It considers that image features are equally weighted on channels and pixels, which leads to the loss of some vital feature information. In this paper, we propose an end-to-end Dual Attention Fusion Network(DAF-Net). The network consists of three residual groups, and each group comprises three residual dual attention fusion modules. The module consists of a residual block and channel and pixel attention fusion to strengthen the global dependence and obtain pixel characteristics. Based on the different features of the fusion modules, adaptive learning is performed using channel and pixel attention to give more weight to essential features. This model structure also keeps the shallower information and transfers it to the deeper layers. The experimental results show that DAF-Net retains details and improves the visual effects by using a stack of fewer residual dual attention fusion modules on synthetic and real-world images.