Feature-adaptive anomaly detection model for onion inspection system

• A novel onion inspection system integrating object detection and feature-adaptive anomaly detection for real-time recognition and filtering of defect onion. • Proposed a feature-adaptive anomaly detection method that optimizes feature layer weights, which enhances sensitivity to deviations while effectively addressing the inherent variability of normal data. Experiments demonstrate strong performance of our method with an accuracy of 97.02 %. • The proposed system was validated in a real-world onion processing facility, achieving a defect detection accuracy of 86.37 % and a filtering accuracy of 78.5 % during industrial tests. • The system reduces manual labor dependency by enabling two workers with AI assistance to match the performance of four workers. Traditional industrial onion inspection relies heavily on manual labor, making it both time-consuming and prone to human error. We propose a novel automated onion inspection system comprising a Vision-Based Inspection Module (VBIM) for detecting defective onions and a Defective Onion Filtering Module (DOFM) for their physical removal. The VBIM introduces a new feature-adaptive anomaly detection method, which optimizes feature layer weights in a student-teacher anomaly detection model. This approach enhances the model’s ability to focus on common features of normal onions while disregarding irrelevant variations, improving the discrimination of defective onions. The proposed system achieved 97.02 % accuracy on a test dataset and 86.37 % accuracy in real-world inspections. With an average processing speed of 4 FPS , it demonstrates strong potential for real-time, high-precision defect detection in industrial applications.