摘要
在金属制造、桥梁隧道工程、建筑行业等过程中,佩戴安全帽可以极大地保护生命安全。目标检测方法可用于检测头盔是否佩戴。但目前的安全帽佩戴检测方法多集中于监督学习,依赖于大量精确标记的数据。但在现实中,标记数据的成本非常高,训练数据的获取不足可能成为性能提升的瓶颈。与有标签的数据相比,无标签的数据更丰富、更便宜、更容易获得。基于这一问题,将伪标签技术引入到传统安全帽检测方法中,提出了一种半监督安全帽检测方法。它在训练模型时同时使用有标签的数据和无标签的数据,只需要少量的有标签的数据,而使用大量的无标签数据来辅助模型的训练。在自制头盔数据集上的实验结果表明,该方法能在有限的标记数据下取得良好的性能,准确率达到92.7%,平均准确率提高3.7%。在标记数据不足的情况下,满足头盔检测的要求。
In the process of metal manufacturing, bridge and tunnel engineering, and construction industry, wearing a safety helmet can greatly protect the safety of life. The target detection method can be used to detect whether a helmet is worn or not. The current safety helmet wearing detection methods mostly focus on supervised learning, which relies on a large number of accurately labeled data. However, in reality, the marked data is very expensive, and the insufficient acquisition of training data may become a bottleneck for performance improvement. Compared with labeled data, unlabeled data are more abundant, cheaper and easier to obtain. Based on this problem, this paper introduces the pseudo-label technology into the traditional safety helmet detection method, and pro-poses a semi-supervised safety helmet detection method. It utilizes both labeled and unlabeled da-ta when training the model, and it requires only a small amount of labeled data, while assisting the training of the model with a large amount of unlabeled data. The experimental results on the self-made helmet data set show that this method can achieve good performance under limited labeled data, with an accuracy rate of 92.7% and an average accuracy increase of 3.7%. It meets the requirements for helmet detection in case of insufficient marking data.
出处
《数据挖掘》
2023年第1期67-74,共8页
Hans Journal of Data Mining
