Defect Detection For Metal Base Of TO-Can Packaged Laser Diode Based On Deep Learning

With the transformation and upgrading of manufacturing industry in China,high-end manufacturing equipment has increasingly strict requirements on the production quality of products.Therefore,it is essential to detect the defects of parts in the complex industrial production process and control their ex-factory quality.TO-can packaged laser diode is the basic component in the field of optical communication with huge market demand.However,at present,the relevant manufacturing enterprises lack special automatic inspection equipment for parts and components.Most of them still use inefficient manual spot check to detect the appearance defect of parts,which is difficult to ensure product quality.Therefore,this paper takes the metal base of TO56 as the research object,and conducts a study on the detection technology of appearance defects of metal TO-base.The details are as follows:Based on analyzing the characteristics of the appearance defects on the metal base of TO56 cans,the overall design of the deep learning-based visual inspection system for the appearance defects on the metal base of TO-can packaged laser diode is determined.The hardware selection and system design are completed through comprehensive analysis of camera,optical lens and light source in order to build a microscopic visual image acquisition platform.The built microvision image acquisition platform is used to complete the acquisition of the defect image of the metal base of the coaxial packaging.All the defect images after image processing are labeled with the target detection frame through the Labelme image annotation tool.After data enhancement processing,the construction of the defect detection data set of the metal base of the coaxial packaging is completed.The YOLO-SO algorithm model is proposed for the appearance defect detection of the metal TO-base,enabling fast and accurate visual detection of defects in the appearance of the metal base of TO-can packaged laser diode.Based on the YOLO-V5 object detection model,a random-paste-mosaic(RPM)small object data augmentation method and K-Means++ anchor box clustering algorithm optimization are proposed,and the convolutional block attention mechanism(CBAM)module is incorporated in the backbone feature extraction network of the model for the problem of missed and wrong detection of small size target defects and defects with obscure features in the original network model.By reducing the influence of background features on defective features,the feature extraction capability of the network model is improved,enabling more accurate localisation and identification of defective targets and improving the generalisability of the network.The defect detection model is trained and evaluated based on the constructed defect detection dataset of metal TO-base.The experimental results on the test dataset demonstrate that the proposed YOLO-SO detection algorithm has an accuracy of 83% and a detection speed of 25 fps on the task of defect detection for metal TObase,which can meet the real-time detection of appearance defects of metal TO-base in actual industrial production.The trained YOLO-SO metal TO-base defect detection model is deployed on the embedded platform,and the software design of the defect detection system is completed based on PYQT5.The channel pruning algorithm is used to complete the lightweight of the network structure,which greatly reduces the network parameters and computation.By comparing the models with different sparse scale factors and pruning rates,the lightweight model YOLOSOLite,which can balance the detection accuracy and detection speed,is selected.The embedded deployment of the defect detection model of the metal TO-base is completed based on the Jetson Nano embedded development board.The built defect detection system for metal TO-base meets the project production requirements in terms of detection speed and defective product detection rate,which can achieve fast and accurate inspection.Guided by the actual needs and problems in the production process,this paper proposes a deep learning-based defect detection system for the metal base of TO56,which has certain theoretical significance and engineering application value for improving the automation level of TO-can packaged laser diode production.