Defect Detection Of Mobile Phone Midframe Based On Deep Learning

With the development of communication technology and mobile Internet,smart phones have become indispensable items in people's lives and work,and consumers' quality requirements for the appearance of mobile phones have also increased.Defect detection is one of the important links of mobile phone appearance quality monitoring.Traditional surface defect detection algorithms are subject to the influence of detection environment factors,and it is often difficult to extract effective features.The backward manual inspection methods can not meet the production needs under the background of intelligent manufacturing.The development of deep learning technology has brought an opportunity for the defect detection of the mobile phone frame.Due to its data-driven characteristics,it has no prior knowledge requirements for developers,and has good robustness and versatility.In this way,this paper focuses on the small defect objects in the mobile phone frame detection,low detection recall rate,and poor real-time performance.Through the elaboration and experimental analysis of the past defect detection network,a deep learning network for mobile phone frame defect detection is proposed.Among them,the main research content and innovation are as follows:(1)In view of the fact that there is still no open source mobile phone midframe defect data set,the collection of this subject data set has been completed.Aiming at the problem of small sample size and unbalanced data in the subject data set,this paper proposes a defect sample generation method based on a generative adversarial network.The generated results show that the method can effectively expand the defect samples and increase the diversity of the data set.(2)In-depth research on the three mainstream detections network of Faster R-CNN,SSD and YOLO,and analysis of their respective detection principles from the network structure.And conduct a horizontal comparison experiment on the open source data set and the data set of this subject to analyze the performance difference between the networks.It laid a theoretical foundation for the subsequent improvement and optimization of the mobile phone midframe detection network.(3)Aiming at the problem of small target detection under high-resolution background,this paper uses k-means clustering to optimize the measurement distance to recalculate the Anchor box,so that the network can better fit curve targets of different sizes.In terms of network feature extraction,CSPDarknet53 is proposed to replace Darknet53 as the feature extraction backbone of the network.At the same time,the feature pyramid pooling layer SPP module is introduced to effectively integrate feature maps of different scales to further expand the receptive field and improve the detection ability of small targets.Finally,the Anchor box matching strategy is adjusted,and the recall rate of small target defects is further improved by reducing the IOU matching threshold.(4)In order to further improve the convergence speed and detection accuracy of the detection network,a stochastic gradient descent algorithm SGDR with hot restart is proposed as the training optimizer of this network.The feature of hot restart can more effectively adjust the change of the learning rate,while avoiding falling into the local optimal value.Compared with SGD and Adam,the experimental results show that SGDR has better convergence speed and detection accuracy in the defect detection of mobile phone midframe.(5)Based on the theoretical foundation and optimization improvement of the detection network,a defect detection system for the midframe of the mobile phone was designed,and the related hardware and software modules were built.The experimental results show that the modified detection system has better detection performance and can fully meet the accuracy and real-time requirements of the industrial detection environment.