Aiming at the detection of non-compliant glove usage by operators,the high-precision YOLOv5 is adopted as the target detection framework and its backbone network is improved to enhance its ability of small-object recognition.Based on its strong small-object recognition capabilities,the attention mechanism(Visual Transformer)modules is incorporated to improve overall recognition accuracy.Additionally,the original loss function is replaced to further enhance recognition speed and accuracy.Finally,a data set collected from operators is trained and validated.Experimental results show that compared to the original network,the proposed optimized YOLOv5 structure has a significant improvement in accuracy,whose average recognition accuracy reaches 95%on the validation dataset.
关键词
目标检测/YOLOv5/卷积网络/注意力机制/CIoU损失函数
Key words
target detection/YOLOv5/convolutional network/attention mechanism/complete intersection over union(CloU)loss
维普
万方数据