基于YOLOv13-EEAttn的实验室安全帽识别研究OA

To address the issue existing in practical teaching of engineering,mechanical,and electrical majors,such as difficulties in supervising the compliant wearing of safety helmets by personnel,and traditional visual recognition algorithms being susceptible to occlusion and significant interference from instrument backgrounds,an improved structure named YOLOv13-EEAttn is proposed based on the YOLOv13 algorithm,which integrates efficient channel attention and efficient multi-scale attention.The improvements are as follows:(1)Introduce the efficient multi-scale attention(EMA)after the deep feature extraction module A2C2F in the Backbone to suppress redundant features from the laboratory background.(2)Add the Para-ECA structure composed of multiple parallel efficient channel attention(ECA)modules before the output in the Head,so as to enhance the channel weights of edge and corner regions of small-scale safety helmets in parallel.(3)Incorporate images of students taken from a top-down view and images of ordinary hats into the datasets to improve the recognition accuracy of safety helmets under surveillance perspectives,and reduce false defections and missed defections caused by interference from ordinary hats.Experimental results show that compared with the original YOLOv13 model,the improved model achieves a 1.44%increase in Recall and a 0.76%improvement in mAP@50,and outperforms models such as YOLOv12 and SSD.Ablation experiments are conducted on the YOLOv13-EEAttn structure to quantitatively analyze the contributions of the EMA and Para-ECA modules to safety helmet recognition performance.Finally,verification with the confusion matrix shows that the detection accuracy rates for safety helmets and ordinary hats are increased by 1.1%and 3.1%respectively.This study provides a technical reference for laboratory safety management in universities.