Research On Surface Defects Detection For Smartphone Glass Cover Based On Convolutional Neural Network

Smartphone glass cover is located in the outermost layer of the mobile phone screen,which is the solid shell of the screen and touch medium,in the production process will inevitably produce all kinds of defects,such as: chipping,scratches,dirt,pits,etc..At present,the quality inspection of smartphone glass covers is mainly carried out by manual visual inspection methods in industrial sites,which have problems such as high inspection costs,high labour intensity,low efficiency and low yields.The traditional machine vision based smartphone glass cover defect detection method has low detection versatility,poor robustness,high image quality requirements,low detection accuracy and other problems.To address the above problems,a convolutional neural network based smartphone glass cover surface defect detection method is proposed,and the effectiveness of the method is verified from both theoretical analysis and experimental research,the main research content is as follows.First,the principles related to the target detection algorithm based on convolutional neural network are analysed and the detection performance of the detection algorithm in the dataset is compared.The principles related to detection algorithms such as Faster R-CNN,a two-stage target detection algorithm,and YOLOv4 and SSD,single-stage target detection algorithms in convolutional neural network algorithms,are analysed.Comparing the detection m AP and detection speed of each algorithm in data sets such as VOC and MS COCO,the results show that YOLOv4 has a greater advantage in detection accuracy and detection speed,so YOLOv4 is chosen as the base algorithm for smartphone glass cover defect detection.Secondly,the YOLOv4 algorithm is improved for the specific defect characteristics of smartphone glass covers.The improved Canopy+K-means algorithm is used for clustering the dataset to solve the problem that the pre-defined anchor frame is not applicable to the small target smartphone glass cover dataset.The Mobile Net V3 network is invoked instead of the CSPDarknet53 network as the feature extraction network,and the MYOLOv4 algorithm is proposed to solve the problems of many small-sized defects and relatively simple backgrounds of defective images of smartphone glass covers.The MYOLOv4 algorithm is improved by using the anchorless frame detection network,and the MYOLOv4-X algorithm is proposed to solve the multi-scale problem of smartphone glass cover defects.Finally,a smartphone glass cover defect dataset is constructed and the performance of the convolutional neural network-based algorithm for smartphone glass cover defect detection is verified.An experimental platform for acquiring defect images is built in the production and inspection workshop of smartphone glass cover,and the obtained images are made into a glass cover defect dataset.Comparing the defect detection performance of the improved YOLOv4 algorithm and the typical convolutional neural network,the experiments show that the proposed MYOLOv4-X algorithm detects 87.53% of m AP and 52.95 fps,which is 8%,0.35% and 1.5% higher than the original YOLOv4,MYOLOv4 and YOLOX algorithms respectively.Compared to other convolutional neural network target detection algorithms,the detection speed and detection accuracy were improved.The proposed algorithm is applied to the test dataset to achieve high accuracy and speed for defect detection of smartphone glass cover defects.