Detection-friendly Hybrid Attention Network for Single Image Dehazing
Qinfei Bao, Hongmin Ren · 2024
Object detection algorithms have made significant improvement in recent years. However, these detection methods still struggle in adverse weather like haze. Existing image dehazing methods often focus on producing haze-free images but may introduce noise that is harmful for object detection networks. Therefore, this paper proposes a detection-friendly Hybrid Attention Network (HAN). It integrates a dehazing network and a detection network. The dehazing network utilizes an encoder-decoder architecture comprising two downsampling modules and a feature transform module in the encoder, and two upsampling modules in the decoder. Each module integrates hybrid attention, combining spatial attention to highlight important spatial infortion within feature maps, and self-attention to capture long-range dependencies across all features. This hybrid attention mechanism significantly enhances the network's ability to learn features comprehensively. Furthermore, we adopt YOLOv5 as our detection network which adopts features extracted by the encoder and computes the object detection loss to optimize the dehazing network. Extensive experimental results demonstrate that HAN is effective in enhancing the clarity of hazy images and improving object detection performance under hazy conditions.