Object detection algorithm based on attention mechanism in foggy weather

Abstract

In this paper, we propose a foggy weather object detection model based on an attention mechanism, to address the problem of low detection accuracy, missed detection and false detection when general object detection models are applied directly to foggy scenes. Firstly, to enhance the detection network's multi-scale expression ability and sensitivity to the target, a residual module that integrates the attention mechanism replaces the BottleNeck module of the backbone network. This design improves the network's ability to extract features and locate targets at a fine-grained level. Secondly, the CIOU loss function replaces the original loss function, improving the stability of the bounding box regression process. Thirdly, the K-means++ clustering algorithm is used to generate anchors suitable for the dataset in this paper. Furthermore, the object detection dataset in foggy scenes is further enriched based on the atmospheric scattering model. Experimental results indicate that the proposed method's mAP in light fog, medium fog and dense fog scenes is increased by 7.4%, 6.05% and 6.36%, respectively, compared to the original YOLOv5s. This improvement in accuracy significantly reduces the missed detection rate and false detection rate, effectively enhancing object detection performance in foggy weather.