Improvement and Application of YOLOv3 for Smartphone Glass Cover Defect Detection

Abstract

Smartphone glass covers defects detected by human, which is inefficiency, high costs, low detection accuracy and labour intensive, while the automatic detection methods based on traditional machine vision is poor flexibility, low yield and poor generalisation capability. Therefore, this paper introduces YOLO (You Only Look Once) v3 to smartphone glass cover defects for the first time. The YOLOv3 algorithm was improved for the actual characteristics and specific requirements of defect detection. First of all, the channel attention mechanism SENet (Squeeze and Excitation Networks) was added to the feature extraction network to detect inconspicuous defect features. Moreover, a 104 × 104 scale detection layer was added to the YOLOv3 detection network to solve the problem of multi-scale defects. Finally, the scaling factor coefficient of the BN (Batch Normalization) layer in the convolutional network is used as the important factor for model pruning to improve the defect detection speed. The improved YOLOv3 algorithm is applied to smartphone glass cover defect detection, and a high accuracy and high detection speed method for smartphone glass cover defects is proposed. 15,914 production site images covering four types of defects, including chipped edges, pits point, soiling and scratches, were obtained from smartphone glass cover manufacturers, 14,321 were annotated as the training set and 1593 were used as the test set to compare and analyse the proposed method and the original YOLOv3 algorithm in this paper. These experiments showed that the mAP (mean average precision) of the detection was 81.0% and the detection speed was 43.1 sheets/s. Compared to the original YOLOv3 algorithm, the mAP of the detection increased by 3% and the detection speed increased by 6.7 frames/s, which meets the need for high precision and efficient detection of defects in the industrial production of smartphone glass covers.