Constraint-Aware PPO Network for Adaptive Food Surface Defect Detection in Industrial Sorting Scenarios

Abstract

For industrial sorting lines, food-surface fault detection should be highly accurate visually and satisfy production needs. It is challenging to differentiate small bruises, fractures, stains, mold spots, and texture anomalies from the inherent characteristics of food, and real-time sorting must make solid choices in the face of conveyor movement and changing light. A constraint-aware PPO model for adaptive food defect detection is presented in this research. Create visual states using the model's multi-scale surface characteristics, then define inspection as a restricted policy optimization problem. Under latency, missed-defect, and false-rejection limitations, the PPO agent gains knowledge of inspection focus modification, confidence control, and sorting decision refinement. Based on 42,600 food surface photos gathered from the three sorting lines, the suggested method's inference latency is 18.7 ms, its detection accuracy is 96.4%, and its defect recall rate is 94.8%. The strategy improved subtle-defect recall by 7.9% and decreased false sorting by 4.6% compared to YOLO, attention-based, and rule-based baselines.