Multi-Stream Fusion Network With Generalized Smooth L1 Loss for Single Image Dehazing

Single image dehazing is an important but challenging computer vision problem. For the problem, an end-to-end convolutional neural network, named multi-stream fusion network (MSFNet), is proposed in this paper. MSFNet is built following the encoder-decoder network structure. The encoder is a three-stream network to produce features at three resolution levels. Residual dense blocks (RDBs) are used for feature extraction. The resizing blocks serve as bridges to connect different streams. The features from different streams are fused in a full connection manner by a feature fusion block, with stream-wise and channel-wise attention mechanisms. The decoder directly regresses the dehazed image from coarse to fine by the use of RDBs and the skip connections. To train the network, we design a generalized smooth L₁ loss function, which is a parametric loss family and permits to adjust the insensitivity to the outliers by varying the parameter settings. Moreover, to guide MSFNet to capture the valid features in each stream, we propose the multi-scale supervision learning strategy, where the loss at each resolution level is computed and summed as the final loss. Extensive experimental results demonstrate that the proposed MSFNet achieves superior performance on both synthetic and real-world images, as compared with the state-of-the-art single image dehazing methods.