| With the rapid development of social production and manufacturing technology,the use of transparent materials is becoming more and more high.And has also been widely used in the packaging industry.The surface defect of transparent packaging is a very important factor to determine the quality of its products,so the detection of its surface defect is an indispensable link in the production process of transparent packaging.In recent years,deep learning technology develops rapidly and occupies a place in the field of defect detection with high accuracy and stable detection rate.Therefore,this paper chooses the YOLOV3 network model of convolutional neural network to study the surface defect detection of transparent packaging.The main research contents of this paper are as follows:First of all,this paper describes the research purpose and significance of the surface defect detection of transparent packaging box,chooses to use YOLOV3 network model to detect the three surface defects of the transparent packaging box,such as scratch,crush and crack,and analyzes the basic theory of convolutional neural network.For example,the basic principle of convolution operation,nonlinear mapping,pooling operation and other operational steps,as well as the more representative algorithms of the two mainstream target detection algorithms.Secondly,the surface defect detection method of transparent packaging box based on YOLOV3 was studied,and an optical acquisition platform was built to collect images of the transparent packaging box.The collected images were standardized and VOC data set was made.Since under the original anchors,the YOLOV3 network model is not agile enough to detect the new target size,this has resulted in inefficient target detection.Therefore,in this article,WE chose to optimize the K-means clustering algorithm,improve the distance formula in the clustering process,analyze the relationship between K value and IOU in the clustering algorithm,and re-determine the K value,thus obtaining nine new anchors.The results show that with the new anchors,the model has improved the accuracy of defect detection;Finally,in order to solve the problem of slow detection rate of YOLOV3 model,the computation in the model detection process is optimized in this study,and the traditional convolution structure is replaced by the depth-separable convolution structure.The final experimental results show that,after the traditional convolution is replaced by the depth-separable convolution,YOLOV3 network model can guarantee certain accuracy of defect detection,and its detection speed is also improved,which can meet our real-time detection standards... |
