Deep learning-based double-sided fudge detection system with integrated physical components

This study proposes an automated defect detection system for large-scale fudge production, addressing the limitations of manual inspection, which is both labor-intensive and error-prone. Deep learning-based image recognition was employed to classify normal samples and four defect types using multiple object detection models, including SSD and YOLOv4/YOLOv5/YOLOv7/YOLOv8/YOLOv11. Instead of relying on a single model, the system integrates a multi-model strategy combining confidence-weighted voting, rule-based selection for specific defects, and Non-Maximum Suppression (NMS), enabling complementary strengths across models and improving robustness. The selected models were deployed in a real-time inspection system equipped with a flipping mechanism that allows each piece of fudge to be inspected on both sides, thereby expanding the coverage of defects. In evaluations using 1,000 real production samples, first-round accuracies were 47.5% (hole), 56.7% (leak), and 60.9% (white). After applying the flipping mechanism for a second inspection, accuracies increased to 75.8%, 83.6%, and 89.3%, respectively, with hole defects showing the largest improvement (28.3%). Our validation-learned fusion achieves 0.995 mAP@0.5 (on par with the best single model) and 0.944 mAP@0.5:0.95, outperforming YOLOv11 and YOLOv5 by + 0.2 and + 2.4 percentage points, respectively; gains are most evident for White defect and Hole defect. These results indicate that combining multi-model detection with dual-side inspection significantly enhances accuracy and enables the reliable screening of defects in real-time production environments.