On Line Detection Of Appearance Defects Of Communication Circulator Based On Deep Learning

As one of the important parts of 5g base station,the stable quality of communication circulator directly affects the progress and precision of the project.In the production process of communication ring,due to the current processing technology,there will be solder joint,missing block,filaments and other defects.In view of these surface defects detection,the current communication circulator detection relies on artificial experience for identification and classification,which seriously affects the production efficiency,and visual fatigue,body state and working environment will affect the accuracy of manual detection,coupled with the increasing cost of employment,so the production line of communication circulator urgently needs online automatic detection system.With the rapid development of digital image processing technology,machine vision detection has played an pivotal role in industrial production.A set of stable automatic detection system will bring great commercial value.Its advantages are:(1)improving production efficiency;(2)optimizing production cost;(3)Can defect big data statistics and analysis,engineers can be based on large data analysis process,improve product percent of pass and also improve the company in the market competition ability.From the perspective of engineering,research and problem-solving,based on the framework of Faster-RCNN and YOLO,actively absorb the latest research results,optimize the appearance defect detection of 5g communication circulator,improve the detection accuracy and development efficiency.The main contents of this paper are as follows:(1)In the stage of sample collection,through the enhancement of limited samples,in addition to the commonly used data enhancement methods,the morphological transformation of finite defects is proposed to realize the diversity of sample number and morphology;(2)By reusing a sample,this paper proposes a target detection method of cascading Faster-RCNN and YOLO,and trains the Faster-RCNN and YOLO models respectively;in the model training stage,through the transfer learning and optimization of super parameters,the target detection algorithms based on Faster-RCNN and YOLO are respectively used to obtain the training model with high precision;(3)In the reasoning stage,by cascading Faster-RCNN and YOLO,two excellent target detection algorithms,this paper proposes a new decision conflict resolution algorithm,and finally obtains a higher comprehensive detection accuracy;(4)In the deployment phase,in order to solve the problem of performance and computing power overhead caused by algorithm cascading,this paper proposes a solution to share the GPU in pieces;(5)Finally,the research summary and prospect are made;Experimental results show that:based on the deep learning method of communication circulator surface defect detection system,effectively locate and identify all kinds of defects,and finally the recognition accuracy of about 98.5%,recognition rate can reach about 0.4 S,compared with the traditional detection method,greatly improving the recognition speed and accuracy.