Ultrasonic scanning image defect detection of plastic packaging components based on FCOS

Abstract

Defect detection of ultrasonic scanning images of plastic packaging components is mainly rely on manpower and not suitable for traditional feature extraction methods, to solve this problem, this paper put forward an optimized FCOS deep learning network to identify its delaminated defects. We redesign the backbone IResNeSt that consists of new bottleneck and data transmission path as the feature extraction module to enhance the information expression ability, furthermore, we introduce a feature pyramid network TF-FPN to improve the feature utilization. Finally, the complete proposed structure FCOS-ITN realizes the identification of various defects and retains more feature details. The experimental results show that compared with the typical object detection method, our FCOS-ITN applied on ultrasonic scan data set locates the delaminated region more accurately. As a matter of fact, the average accuracy (mAP) achieved 90.27% on all defect types, which is 6.58% higher than that of the original frame, indicating that our approach is feasible for non-destructive defect detection.