Multi-Scenario Decision Framework for AI Model Selection in Food Visual Inspection

This study addresses the gap between algorithmic performance and manufacturing decision needs in bakery quality control by developing a decision-oriented framework for AI-based visual inspection model selection. Three YOLO object-detection architectures, namely YOLOv5, YOLOv8 and YOLOv11, were benchmarked on a bakery-defect dataset under nominal and controlled synthetic perturbation conditions, with Faster R-CNN included as a non-YOLO baseline. Standard detection metrics, including precision, recall, mAP@50 and mAP@50–95, together with inference latency, were translated into three managerial indicators: Risk-Weighted Total Cost of Quality (TCQ), Throughput Efficiency (TE) and Waste Reduction Index (WRI). These indicators enabled model comparison across high-volume commodity, brand-protection and resource-constrained SME production scenarios. Results show that model suitability is scenario-dependent rather than determined by aggregate detection accuracy alone. YOLOv11 achieved the most favourable profile in high-speed commodity production because its lower latency avoided downtime costs. YOLOv8 provided the lowest TCQ in brandprotection and SME scenarios, where weighted missed-defect costs were more influential than throughput losses. YOLOv5 did not yield the lowest TCQ under the adopted assumptions, but remained relevant as a lowercomplexity alternative when false-positive control, waste avoidance or legacy deployment are prioritised. Faster R-CNN was consistently dominated in the investigated case. The proposed cost-quality-throughput decision matrix supports small and medium-sized enterprises (SMEs) in selecting “good-enough” architectures based on constraints and risk appetite rather than technical accuracy alone.