Citation

An, Da (2026) Efficient and defect-aware deep learning-based object detection for industrial surface defect detection. PhD thesis, Multimedia University.

Full text not available from this repository.

Abstract

Surface defect detection is a key task in intelligent manufacturing because it has a direct impact on product quality, operational safety, and process stability. Although deep learning-based object detection methods have made substantial progress in general computer vision, their application to industrial surface inspection still faces significant challenges. In practical scenarios, defects are often characterized by low visual saliency, small size, and strong interference from complex surface textures, while industrial deployment also requires fast inference and efficient execution. Therefore, improving detection accuracy and enhancing realtime performance have become two important research directions in industrial surface defect detection. This dissertation focuses on industrial workpiece surface defect detection and conducts the study from two complementary aspects: accuracy improvement and real-time optimization. On the one hand, targeted methods are investigated to enhance defect recognition performance under challenging conditions such as complex backgrounds, weak defect features, and tiny defect patterns. On the other hand, lightweight and efficient design strategies are explored to satisfy the practical requirements of high inference speed, low computational cost, and deployable implementation in industrial environments. Through these two research perspectives, this work aims to improve both the effectiveness and the practical applicability of surface defect detection in intelligent manufacturing. Specifically, background-aware feature modulation is explored to selectively strengthen defect-relevant responses while suppressing background-induced activations, thereby improving feature discrimination on complex textured surfaces. In addition, a defect-oriented texture augmentation strategy is developed to increase local texture diversity within defect regions and enhance the learnability of small and subtle defects. Furthermore, lightweight and small-object-oriented structural enhancement strategies are investigated to strengthen high-resolution feature utilization and restore fine-grained defect details while maintaining low computational overhead. Extensive experiments and ablation studies are conducted on public and realworld industrial defect datasets to evaluate the proposed methodology. The results demonstrate that the proposed approach consistently outperforms representative state-of-the-art detectors in terms of detection accuracy, recall for tiny defects, and robustness under complex backgrounds, while achieving real-time inference with a lightweight model footprint. The contributions of this dissertation provide a practical and reproducible solution for balancing accuracy and efficiency in industrial defect detection, and offer methodological insights for the design of deployable deep learning–based inspection systems in intelligent manufacturing.

Item Type:	Thesis (PhD)
Additional Information:	Call No.: Q325.73 .A53 2026
Uncontrolled Keywords:	Deep learning (Machine learning)—Industrial applications
Subjects:	Q Science > Q Science (General) > Q300-390 Cybernetics
Divisions:	Faculty of Computing and Informatics (FCI)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	05 Oct 2026 05:55
Last Modified:	05 Oct 2026 05:55
URII:	http://shdl.mmu.edu.my/id/eprint/16877

Downloads

Edit (login required)