A Hierarchical hybrid architecture with CNN autoencoder for foreign object detection on jig surfaces in smartphone screen manufacturing

Van Chung Luong; Xuan Tuan Dong; Trung Kien Ha; Quy Thinh Dao

Authors

Van Chung Luong School of Electrical and Electronic Engineering, Hanoi University of Science and Technology, Vietnam
Xuan Tuan Dong School of Electrical and Electronic Engineering, Hanoi University of Science and Technology, Vietnam
Trung Kien Ha The Department of Electrical Engineering, School of Electrical and Electronic Engineering (SEEE), Hanoi University of Industry (HaUI)
Quy Thinh Dao School of Electrical and Electronic Engineering, Hanoi University of Science and Technology, Vietnam

Keywords:

Anomaly Detection, Automated Optical Inspection, CNN Autoencoder, Industrial Image Processing, Jig Surface Inspection, Robust Statistics

Abstract

Foreign object and surface defect detection on inspection jigs is critical in smartphone screen manufacturing, demanding both real-time speed and high accuracy. This paper proposes a Hierarchical Hybrid Pipeline comprising two phases: Phase 1 employs traditional image processing with rigid Euclidean alignment, absolute background subtraction, and robust-statistics-based dynamic thresholding (Median + k × MAD) for high-speed full-frame screening; Phase 2 employs an unsupervised CNN Autoencoder with Ensemble anomaly scoring (Local MSE + Feature-based MSE) for in-depth verification of suspect regions. Evaluated on a dataset collected from an actual Lamination production line, Phase 1 achieves 99.25% Accuracy on 1,200 images (120–200 ms processing time). Phase 2 attains an AUC-ROC of 0.971 and F1-score of 0.961 with 97.1% Recall. The complete pipeline executes within 300–350 ms on industrial CPU hardware, satisfying real-time production requirements.