基于改进YOLOv5的移动端螺栓缺失检测方法

An improved YOLOv5-based bolt missing detection method for mobile terminals

大型结构件中螺栓缺失往往会产生重大的安全隐患,而目前通过深度学习进行目标检测的设备通常部署繁琐、困难,为了保证检测的高效率并且足够轻量化的性能,提出一种通过改进YOLOv5进行模型训练,并将模型轻量化移植到移动端的方法,以实现螺栓的快速识别与个数缺失统计与报警。首先利用LabelImg软件对现有的螺栓数据进行标注,标注完成后对该数据集进行训练,再将得到的模型文件转化并降低位宽,转化为TensorFlow Lite(TFLite)模型文件,进而在安卓端进行部署。结果表明,将模型部署在小米MI 9 SE手机上,进行单目标检测时准确率可以达到98%,多目标检测时准确率可以达到97%,推理时间在40 ms左右,且报警功能可以正常使用,为螺栓缺失的检测提供了一种新的轻量化方法。

The lack of bolts in large structural components often leads to significant security risks.At present,the deployment of equipment for target detection through deep learning is cumbersome and difficult.In order to ensure the high efficiency and sufficient lightweight performance of detection, a method to train the model by improving YOLOv5,and transplant the model lightweight to the mobile was proposed, so as to realize the rapid identification of bolts and the statistics and alarm of the number of missing bolts. Firstly the LabelImg software to label the existing bolt data was used. After the labeling is completed,the data set was trained,and then the obtained model file was converted and the bit width was reduced to a TFLite model file,and then deployed on the Android side. The results show that when the model is deployed on Xiaomi MI 9 SE mobile phone,the accuracy of single target detection can reach 98 %,the accuracy of multi-target detection can reach 97 %,the reasoning time is about 40 ms,and the alarm function can be used normally,which provides a new lightweight method for bolt missing detection.