MSF$^2$DN:Multi Scale Feature Fusion Dehazing Network with Dense connection

Guangfa Wang, Xiaokang Yu; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 2950-2966

Single image dehazing is a challenging problem in computer vision. Previous work has mostly focused on designing new encoder and decoder in common network architectures, while neglecting the connection between the two. In this paper, we propose a multi-scale feature fusion dehazing network based on dense connection, MSF^2DN. The design principle of this network is to make full use of dense connection to achieve efficient reuse of features. On the one hand, we use a dense connection inside the base module of the encoder-decoder to fuse the features of different convolutional layers several times, and on the other hand, we design a simple multi-stream feature fusion module which fuses the features of different stages after uniform scaling and feeds them into the base module of the decoder for enhancement. Numerous experiments have demonstrated that our network outperforms the existing state-of-the-art networks in real-world datasets.