基于改进DSSD网络的机械材料缺陷识别方法

The mechanical material defect identification method based on improved DSSD network

针对机械材料缺陷识别准确率低的问题,提出一种基于改进反卷积单激发多框探测器(DSSD)网络的机械材料缺陷识别方法。首先以Residual101为基础网络,通过减少反卷积层数量以提高检测速度,然后采用K-means聚类算法优化目标检测框的长宽比,提高对机械材料缺陷识别的准确率和检测速度。结果表明,通过改进DSSD网络可实现机械材料缺陷检测,且相较于DSSD网络和常用目标检测网络YOLO、SSD、Faster-RCNN,改进DSSD网络的缺陷识别平均准确率更高,平均检测时间更短,分别达到85.9%和0.141 s,具有一定的有效性和优越性。

Aiming at the low accuracy of mechanical material defect identification,a mechanical material defect identification method based on improved deconvolution single shot detector(DSSD)network is proposed.Firstly,Residual101 is used as the basic network,and improving detection speed by reducing the number of deconvolution layers.Then,K-means clustering algorithm is adopted to optimize the length-width ratio of the target detection frame.Finally,the accuracy and detection speed of mechanical material defect identification is improved.The simulation results show that the proposed improves DSSD network can achieve mechanical material defect detection.In addition,compared with DSSD network and common target detection networks YOLO,SSD and Faster-RCNN,the improved DSSD network has higher average defect identification accuracy and shorter average detection time,reaching 85.9%and 0.141 s respectively,which has certain effectiveness and superiority.