Machine vision for visible mold detection in stored oil-bearing raw materials under simulated conveying conditions

Abstract

Post-storage oil-bearing raw materials may develop visible mold-related deterioration before processing or sorting, which can reduce raw material utilization, affect processing suitability, and increase the difficulty of front-end quality management. To support rapid non-destructive screening, this study developed a machine vision method for appearance-based visible mold detection in stored peanuts, corn kernels, and soybeans under simulated conveying-related conditions. Images of normal-appearing and visibly moldy samples formed under natural storage conditions were collected, and a six-class object detection dataset, OOSD-6, was constructed. To account for common imaging disturbances during conveying and screening, illumination variation, background replacement, scene recomposition, and motion blur augmentation were introduced. Based on YOLO11, an improved detection model, METW-YOLO, was developed for multi-class visible mold detection. On the validation set, METW-YOLO achieved Precision, Recall, [email protected], and [email protected]:0.95 values of 95.36%, 95.82%, 98.05%, and 95.90%, respectively, corresponding to improvements of 2.13%, 2.53%, 1.66%, and 1.82% over the baseline YOLO11. The results indicate that METW-YOLO improves detection performance under the constructed dataset and simulated conveying-related disturbance conditions. The proposed method provides a potential visual pre-screening approach for identifying visibly moldy samples before processing or sorting, and may support front-end quality management of post-storage oil-bearing raw materials.