DOI: 10.35377/saucis...1945069 ISSN: 2636-8129

Fuzzy Logic-Based Hybrid Deep Learning System for Metal Surface Defect Detection

Ethem Belka Şahin, Seçkin Arı
Show PDF Cite
Metal surface defect detection is a fundamental quality control challenge in the manufacturing industry. Surface defects, such as holes and crescent-shaped edge defects, directly affect the structural integrity of metal components, making their accurate and timely identification critical in the production pipeline. Single-model deep learning approaches tend to exhibit inconsistent performance across varying imaging conditions, such as blur, noise, and illumination changes, motivating the investigation of multi-model fusion strategies. This paper proposes a hybrid object detection system that integrates YOLOv8s and Faster R-CNN through a Sugeno-type fuzzy inference mechanism for metal surface defect detection. The system operates in three sequential stages: independent inference by both models, fuzzy-weighted fusion using a rule-based confidence evaluation mechanism, and hybrid non-maximum suppression. A greedy intersection-over-union matching algorithm pairs the predictions from both models, and the resulting confidence scores are fuzzified through a Sugeno-type fuzzy inference system to produce adaptive fusion weights. A parametric ablation study was conducted across nine system configurations by varying the membership function boundaries, unmatched detection filters, class-specific rule tables, and hybrid NMS thresholds. The system was evaluated on a custom metal surface dataset containing two defect classes under seven types of synthetic image distortion across 26 sub-levels comprising 2,080 test images. The best-performing configuration achieved a macro-average F1 score of 0.9100 and precision of 0.9437, outperforming four comparison methods, including the standard NMS ensemble, equal-weight fusion, Maximum Confidence Selection, and Weighted Box Fusion. The hybrid system demonstrated superior robustness to Gaussian blur and motion blur conditions compared to either standalone model, with an inference latency of 87.9 ms at 11.4 FPS, confirming its suitability for production-line deployment.

More from our Archive