Dynamic Multi-Attention Dehazing Network with Adaptive Feature Fusion

This paper proposes a Dynamic Multi-Attention Dehazing Network (DMADN) for single image dehazing. The proposed network consists of two key components, the Dynamic Feature Attention (DFA) module, and the Adaptive Feature Fusion (AFF) module. The DFA module provides pixel-wise weights and channel-wise weights for input features, considering that the haze distribution is always uneven in a degenerated image and the value in each channel is different. We propose an AFF module based on the adaptive mixup operation to restore the missing spatial information from high-resolution layers. Most previous works have concentrated on increasing the scale of the model to improve dehazing performance, which makes it difficult to apply in edge devices. We introduce contrastive learning in our training processing, which leverages both positive and negative samples to optimize our network. The contrastive learning strategy could effectively improve the quality of output while not increasing the model’s complexity and inference time in the testing phase. Extensive experimental results on the synthetic and real-world hazy images demonstrate that DMADN achieves state-of-the-art dehazing performance with a competitive number of parameters.