Method based on the cross-layer attention mechanism and multiscale perception for safety helmet-wearing detection. (October 2021)

• Record Type: Journal Article
• Authors: Han, Guang; Zhu, Mengcheng; Zhao, Xuechen; Gao, Hua
• Is Part Of: Computers & electrical engineering. Volume 95(2021)
• Publication Date: 2021-10
• Subjects: Object detection -- Safety helmet detection -- Attention mechanism -- Feature pyramid -- Multiscale perception; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2021.107458
• Languages: English
• ISSNs: 0045-7906

Abstract

Highlights: A cross-layer parallel attention network is proposed to further refine the semantic information of the high-level fusion features and the fine-grained information of the low-level fusion features. A multiscale perception module is proposed to improve the robustness of the network to object scale change. An effective anchor box allocation strategy is proposed to detect small objects at the low-level feature maps. Our model can achieve mAP of 88.1% on the GDUT-HWD dataset and mAP of 80.5% on the PASCAL VOC 2007 dataset respectively. Abstract: To solve the problem of low accuracy in existing safety helmet detection methods, a novel object detection algorithm based on Single Shot Multibox Detector (SSD) is proposed in this paper. The algorithm uses the spatial attention mechanism for low-level features and the channel attention mechanism for high-level features, this cross-layer attention mechanism can further refine the feature information of the object region. The proposed detection algorithm introduces a feature pyramid and multiscale perception module to improve its robustness to object scale change. In addition, an effective anchor box adaptive adjustment method is designed to adaptively adjust the scale distribution of each layer of the anchor boxes based on the feature map size. Experiment results demonstrate that our detection model has mean Average Precision (mAP) of 88.1% and 80.5% on helmet dataset and VOC 2007 dataset respectively, which is better than baseline by 15.65% and 3.4%. Graphical abstract: The network framework of CASP detection algorithm for safety helmet-wearing detection. Image, graphical abstract