Artificial intelligence (AI) and computer vision (CV) offer promising avenues for automated quality control in food manufacturing, yet many prior works in that sector focused on agricultural primary production tasks. This study evaluates object detection for in-line quality monitoring on a real production line for ready-to-eat chicken-type products. Overhead cameras captured images at four processing steps: forming, coating, frying, and cooking. For each step, we labeled 2000 images containing multiple products with multiple classes of quality deviations. Separate YOLOv12x models (default and hyperparameter-tuned) were trained per step and evaluated using mAP50–95, F1-curves, and confusion matrices. Step-specific models, i.e., models applicable solely for a specific processing step, achieved similar peak mAP50–95 (0.50–0.60), and hyperparameter tuning did not yield any major gains despite high computational cost. Performance was strongly tied to class frequency: common classes achieved high F1-Scores, whereas rare classes were often misclassified. To mitigate imbalance and improve robustness, we trained a single model on a combined dataset spanning all steps, which attained a higher peak mAP50–95 of 0.7331 ± 0.0040 and produced more balanced F1-curves, albeit with some loss of step-specific strengths, such as detection of certain deviations specific to that step. The results indicate that out-of-the-box detectors can add practical value to industrial CV-enhanced quality control in food processing, and that further improvements will primarily come from targeted data collection for minority classes, instance-centric datasets, higher-resolution or multi-scale training, and methods that address class imbalance.
Einsiedel et al. (Fri,) studied this question.