An Annotation-Efficient Parallel Processing Framework for Robust Defect Detection in TFT-LCD Manufacturing

Results in Engineering, pp. 113476

Abstract

High-performance automated defect detection is essential for achieving cost-effective thin-film transistor liquid crystal display (TFT-LCD) manufacturing. This paper proposes a weakly-supervised computer vision framework that integrates hybrid edge detection, deep learning, and parallel ensemble methods for robust industrial quality control. The primary innovation is a weakly-supervised region-of-interest (ROI) detection approach that utilizes only image-level class annotations (ILCA), reducing the manual data-labeling labor effort by approximately 71% while maintaining high center-pinpointing capability. The framework utilizes an optimized majority voting ensemble (OMVE) comprising DenseNet-201-GGC, PYLON-RN50E, and EfficientNetV2-s-CBAM, paired with a parallelized sliding window mechanism (SWM) to isolate variable-scale anomalies at native pixel resolutions. To demonstrate cross-domain generalizability, the framework was extensively benchmarked against modern attention-driven Vision Transformers (RT-DETR-L) and classic two-stage anchor-based (Faster R-CNN) fully-supervised architectures using the public NEU steel strip database and the DAGM 2007 texture database. Experimental results on the evaluated industrial validation datasets demonstrate a peak overall classification accuracy of 96.33%, an overall macro precision of 95.46%, a macro recall of 95.98%, and a balanced F1-score of 95.72%. Wilcoxon signed-rank tests ( p < 0.05) confirm that the proposed OMVE statistically outperforms individual model topologies, establishing a highly accurate and annotation-efficient pipeline for advanced manufacturing quality inspection.

Authors 8

1. Muhammad Irshad

   Minhaj University Lahore · Centre for Artificial Intelligence and Robotics

   Affiliation as printed

   Hong Kong Industrial Artificial Intelligence and Robotics Centre, 19W, Hong Kong

   School of Computer Science, Faculty of Computer Science & Information Technology, Minhaj University Lahore, Pakistan

2. Poorya Khanali Aachen Information Management

   RWTH Aachen University

   Affiliation as printed

   Information Management, FIR an der RWTH Aachen e.V., Campus Boulevard 55, Aachen, 52074, Germany

3. Sameed Sajid

   Centre for Artificial Intelligence and Robotics

   Affiliation as printed

   Hong Kong Industrial Artificial Intelligence and Robotics Centre, 19W, Hong Kong

4. Martin Wanly Aachen Information Management

   RWTH Aachen University

   Affiliation as printed

   Information Management, FIR an der RWTH Aachen e.V., Campus Boulevard 55, Aachen, 52074, Germany

5. Susanna Weber Aachen Information Management

   RWTH Aachen University

   Affiliation as printed

   Information Management, FIR an der RWTH Aachen e.V., Campus Boulevard 55, Aachen, 52074, Germany

6. Benny Drescher

   Centre for Artificial Intelligence and Robotics

   Affiliation as printed

   Hong Kong Industrial Artificial Intelligence and Robotics Centre, 19W, Hong Kong

7. Chang Liu

   Centre for Artificial Intelligence and Robotics

   Affiliation as printed

   Hong Kong Industrial Artificial Intelligence and Robotics Centre, 19W, Hong Kong

8. ASM Ashraf Mahmud corresponding

   University of the West of Scotland

   Affiliation as printed

   School of Computing, Engineering and Physical Sciences, University of the West of Scotland, United Kingdom