基于改进YOLOv4网络的手机曲面玻璃缺陷检测

Mobile phone curved glass defect detection based on improved YOLOv4 network

针对手机曲面玻璃缺陷难以识别、检测效率低、识别精度低的问题,提出改进的YOLOv4网络模型。模型的主干网络由CSPDarknet53网络替换为MobileNetv3网络,该网络可以减少参量的计算,减轻设备的负担,提高算法的效率,加强对缺陷细节的识别。采用K-means++聚类自动生成锚框,提高算法的识别精度。该算法改进空间金字塔池化层,使用SPPF模块,该模块既能实现SPP模块相同的功能,又提高了算法效率。损失函数使用CIoU代替IoU,丰富网络感受视野的同时增加了远距离目标间的交互,提升了对微小目标的检测精度,更好地识别手机曲面玻璃中的细微缺陷。对比实验结果表明,改进的YOLOv4检测4种缺陷的AP值均得到提升,mAP值达到了97.57%,相较于传统的YOLOv4算法精度提升了1.23%,检测时间提高了10 ms,能有效地识别手机曲面玻璃的各种缺陷。

In view of the difficult identification,low detection efficiency and low identification accuracy of curved glass defects of mobile phones,an improved YOLOv4 network model is proposed.As for the backbone network of the model,CSPDarknet53 is replaced by MobileNetv3,which can reduce the calculation of parameters,reduce the burden of equipment,improve the efficiency of the algorithm and strengthen the identification of defect details.The anchor frame is automatically generated by K⁃means++clustering to improve the recognition accuracy of the algorithm.In the algorithm,the space pyramid pooling layer is improved,and the SPPF module is used.This module can realize the same function as SPP module and improve the efficiency of the algorithm.CIoU is used as the loss function instead of IoU,which enriches the network perception field and increases the interaction between distant objects,improves the detection accuracy of tiny objects,and better identifies the subtle defects of the curved glass of the mobile phones.The results of comparative experiments show that the AP values of the four defects of the improved YOLOv4 are all improved.The value of mAP reaches 97.57%.In comparison with the traditional YOLOv4 algorithm,its accuracy increases by 1.23%,and its detection time increases by 10 ms,so the improved model can effectively identify various curved glass defects of mobile phones.