Multi-Scale Attentive Feature Fusion Network for Single Image Dehazing

Single image dehazing is a challenging ill-posed problem due to the severe degeneration of image information, which hinders the practical application of images. Due to its ill-posed quality, it faces great challenges especially in dealing with dense haze. In this paper, we propose a Multi-Scale Attentive Feature Fusion Network (MAFFNet) based on the U-Net structure to process dense haze regions from the hazy images while accurately preserving the image details. Further, we introduce the Boost Information Connection (BIC) module, Dense Residual (DR) module, and Feature Fusion (FF) module in MAFFNet to guide the feature extraction and feature fusion and eventually construct the dehazing image. Since the connection of shallow and deep information is helpful for network feature extraction, we use BIC to perform phase and phase subtraction operations on the feature maps of the front and back blocks. DR is used to identify dense haze regions and color channels. In addition, FF aims to accelerate convergence by aggregating features at several layers. The approach is thoroughly evaluated on the SOTS and the NH-HAZE datasets. The result indicates that our MAFFNet surpasses the current state-of-the-art single image dehazing approaches in improving the dehazing performance and maintaining image details.