基于改进YOLOv3的电容器外观缺陷检测

Capacitor appearance defect detection based on improved YOLOv3

针对部署于有限算力平台的YOLOv3(you only look once v3)算法对电容器外观缺陷存在检测速度较慢的问题,提出了基于YOLOv3算法改进的轻量化算法MQYOLOv3。首先采用轻量化网络MobileNet v2作为特征提取模块,通过利用深度可分离式卷积替换一般卷积操作,使得模型的参数量大幅度降低进而提高模型的检测速度,同时也带来了检测精度的降低;然后在网络结构中嵌入空间金字塔池化结构实现局部特征与全局特征的融合、引入距离交并比(distance intersection over union,DIoU)损失函数优化交并比(intersection over union,IoU)损失函数以及使用Mish激活函数优化Leaky ReLU激活函数来提高模型的检测精度。本文采用自制的电容器外观缺陷数据集进行实验,轻量化MQYOLOv3算法的平均精度均值(mean average precision,mAP)为87.96%,较优化前降低了1.16%,检测速度从1.5 FPS提升到7.7 FPS。实验表明,本文设计的轻量化MQYOLOv3算法在保证检测精度的同时,提高了检测速度。

Aiming at the problem that the you only look once v3(YOLOv3)algorithm deployed on the limited computing power platform has a slow detection speed for the appearance defects of capacitors,an improved lightweight algorithm MQYOLv3 based on YOLOv3 algorithm is proposed.First,the lightweight network MobileNet v2 is used as the feature extraction module,and by replacing the general convolution operation with the deep separable convolution,the amount of model parameters is greatly reduced and the detection speed of the model is improved,but also bring the reduction of detection accuracy.Then,the spatial pyramid pooling structure is embedded in the network structure to realize the fusion of local and global features,the distance intersection over union(DIoU)loss function is introduced to optimize the intersection over union(IoU)loss function,and the Mish activation function is used optimize the Leaky ReLU activation function to improve the detection accuracy of the model.This paper uses a self-made capacitor appearance defect data set for experiments.The mean average precision(mAP)of the lightweight MQYOLOv3 algorithm is 87.96%,which is 1.16%lower than before optimization,and the detection speed is increased from 1.5 FPS to 7.7 FPS.Experiments show that the lightweight MQYOLOv3 algorithm designed in this paper improves the detection speed while ensuring the detection accuracy.