| 摘要: |
| :目的 为了解决变电所场景下因模型参数量过大、推理时间较长导致错检和漏检问题较高等问题,提出一种
基于知识蒸馏和注意力机制的变电所安全帽佩戴实时检测算法 YOLO-FE(BCKD)。 方法 为了减少模型的大小和
推理速度,使用 FasterNet block,去替换原本骨干网络和 NECK 部分中的 C2f 模块;为了减少因为降低模型参数而
导致精度大幅下降的问题,使用高效多尺度的 EMA 注意力机制加入骨干网络之中,提高模型提取特征的能力;最
后为了让模型能具有更高的精度,选择 YOLOv8m 作为教师网络,YOLO-FE 作为学生网络,使用 BCKD 知识蒸馏方
法对 YOLO-FE 进行知识蒸馏。 结果 该研究使用在煤矿变电所收集到的安全帽数据集为基础,选用不同算法进行
对比实验。 最终表明:改进的 YOLOv8 模型未蒸馏前平均精度达到 74. 5%,相较于 YOLOv8n 提高了 2. 5%,参数量
下降 26. 7%,检测速度提高 26%,经过蒸馏之后模型的平均精度达到 75. 4%。 为了验证算法的泛用性,在公开数据
集 GDUT-HWD 上测试,其精度值为 84. 2%,比原始模型提高 3%。 结论 文中提出的安全帽佩戴检测算法在精度和
实时性上取得了良好的平衡,满足变电所环境下的实时检测需求。 |
| 关键词: 安全帽 YOLOv8 FasterNet 注意力机制 知识蒸馏 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Helmet-wearing Detection Combining Knowledge Distillation and Attention Mechanism |
|
LIU Yana, ZHOU Mengran b
|
|
a. School of Artificial Intelligence b. School of Electrical and Information Engineering Anhui University of Science and
Technology Huainan 232001 Anhui China
|
| Abstract: |
| Objective To address the issues of large model parameters long inference time and high rates of false
positives and negatives in safety helmet detection in substation scenarios a real-time helmet-wearing detection algorithm
YOLO-FE BCKD based on knowledge distillation and attention mechanism is proposed. Methods In this model to
reduce the model size and inference time FasterNet blocks were used to replace C2f modules in the original backbone
network and NECK section. Furthermore to mitigate the significant drop in accuracy due to reduced model parameters
an efficient multi-scale EMA attention mechanism was integrated into the backbone network to enhance the model?? s feature
extraction capability. Finally to achieve higher accuracy YOLOv8m was selected as the teacher network and YOLO-FE
as the student network and the BCKD knowledge distillation method was adopted to distill knowledge into YOLO-FE.
Results This study used a helmet dataset collected from coal mine substations and conducted comparative experiments
with different algorithms. The results showed that the improved YOLOv8 model achieved an average precision of 74. 5%
before distillation surpassing YOLOv8n by 2. 5% with a 26. 7% reduction in parameters and a 26% increase in detection speed. After distillation the model?? s average precision reached 75. 4%. To verify the algorithm?? s generalizability it was
tested on the public dataset GDUT-HWD achieving an accuracy of 84. 2% a 3% improvement over the original model.
Conclusion The proposed helmet-wearing detection algorithm achieves a good balance between accuracy and real-time
performance meeting the requirements for real-time detection in substation environments. |
| Key words: helmet YOLOv8 FasterNet attention mechanism knowledge distillation |
