An Investigation Into a Lightweight Safety Helmet Detection Approach Utilizing the MSS-YOLO Framework

An Investigation Into a Lightweight Safety Helmet Detection Approach Utilizing the MSS-YOLO Framework

Motivated by the challenges of low detection accuracy, false positives, false negatives, and the excessive computational demand in safety helmet detection within complex power scenarios, this paper introduces a lightweight safety helmet object detection algorithm based on MSS-YOLO. Initially, the pr...

Full description

Saved in:
Bibliographic Details
Main Authors: Bing Zeng, Zhanpeng Liao, Yu Zhou, Dilin He, Zhihao Zhou, Kexin Yi, Yuwen Li, Xiaopin Yang, Yunmin Xie
Format: Article
Language:English
Published: IEEE 2024-01-01
Series:IEEE Access
Subjects:
MSS-YOLO
safety helmet
lightweight
MobileNetV3-SC
simSPPF-N
SC attention mechanism
Online Access:https://ieeexplore.ieee.org/document/10767142/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Motivated by the challenges of low detection accuracy, false positives, false negatives, and the excessive computational demand in safety helmet detection within complex power scenarios, this paper introduces a lightweight safety helmet object detection algorithm based on MSS-YOLO. Initially, the proposed lightweight backbone network, MobileNetV3-SC, replaces the original DarkNet53 to reduce complexity. Subsequently, an improved lightweight spatial pyramid pooling module, simSPPF-N(simplified SPPF-NAM), is appended to the backbone network, facilitating a more effective amalgamation of local and global features. Additionally, the SC(Spatial and Channel Attention Mechanism) lightweight attention mechanism module is incorporated into the upsampling layers to enhance the algorithm’s focus on targets. The CIoU loss function supersedes the conventional IoU, thereby boosting detection accuracy. The experimental outcomes reveal that compared to its predecessor, the proposed algorithm achieves a significant reduction in parameters from 61.949M to 13.547M and in computational load from 66.171G to 7.895G, while simultaneously elevating mAP@0.5 from 81.99% to 91.20%. This demonstrates the algorithm’s capability to maintain high detection accuracy while drastically minimizing its parameter size and computational overhead.
ISSN:2169-3536

Similar Items