Nondestructive perception of potato quality in actual online production based on cross-modal technology

Nowadays, China stands as the global leader in terms of potato planting area and total potato production. The rapid and nondestructive detection of the potato quality before processing is of great significance in promoting rural revitalization and augmenting farmers’ income. However, existing potato quality sorting methods are primarily confined to theoretical research, and the market lacks an integrated intelligent detection system. Therefore, there is an urgent need for a post-harvest potato detection method adapted to the actual production needs. The study proposes a potato quality sorting method based on cross-modal technology. First, an industrial camera obtains image information for external quality detection. A model using the YOLOv5s algorithm to detect external green-skinned, germinated, rot and mechanical damage defects. VIS/NIR spectroscopy is used to obtain spectral information for internal quality detection. A convolutional neural network (CNN) algorithm is used to detect internal blackheart disease defects. The mean average precision (mAP) of the external detection model is 0.892 when intersection of union (IoU) = 0.5. The accuracy of the internal detection model is 98.2%. The real-time dynamic defect detection rate for the final online detection system is 91.3%, and the average detection time is 350 ms per potato. In contrast to samples collected in an ideal laboratory setting for analysis, the dynamic detection results of this study are more applicable based on a real-time online working environment. It also provides a valuable reference for the subsequent online quality testing of similar agricultural products.