Image Dehazing based on Multi-scale Feature Fusion under Attention Mechanism

To solve the problems of insufficient feature extraction and the loss of too much image information in existing methods, a dehazing network based on multi-scale feature fusion under attention mechanism is proposed. Firstly, the base convolutional layer in U-Net is built using improved fully connected residual blocks to reduce the amount of computation. Secondly, the self-convolution block based on the self-attention mechanism is added to extract more delicate feature information of the image. Finally, to increase feature reuse and reduce feature information loss, the feature maps of different levels are fused using various scale gated units. In order to improve the capacity of the restored image to be recognized subjectively, the mixed loss function of multi-scale structural similarity and minimal absolute error is introduced. Experiments are carried out with synthetic haze data sets. Compared with other neural networks, the multi-scale structural similarity and peak signal-to-noise of the dehazed image of the proposed network are increased by 4.31% and 18.33% on average, respectively. The experiment results demonstrate that the network can efficiently avoid color distortion, halo and strong edge effect around the object, and the image has high subjective recognition after haze removal.