Mobile phone screen surface scratch detection based on optimized YOLOv5 model (OYm)

Abstract

Abstract To improve phone screen surface detection efficiency, an optimized YOLOv5s model (OYm) based on GhostNet(YOLOv5GHOSTs) and BottleneckCSP is proposed. For a given target sample, OYm could effectively reduce the computation of GFLOPS and detection time by optimizing the network structure. The detection results show that the mean average precision_0.5 (mAP_0.5) exceeds 95%, and the average detection rate is 16 ms. Compared with the traditional YOLOv5s model, the loss of average accuracy is ensured to be controlled within 3%, the detection frame rate of OYm is risen by 56.25%, and GFLOPS is decreased by 64.2%. The principle of OYm is explained in detail, and the proposed model is then experimentally validated.

Keywords

• BottleneckCSP
• deep learning
• defect detection
• GhostNet
• YOLOv5