AI-Based Non-Destructive Online Quality Detection of Potatoes Using Cross-Modal Imaging and Spectroscopy

At present, india is one of the leading countries in the world in terms of potato cultivation area and overall potato production. Rapid and non-destructive evaluation of potato quality before processing is very important for improving post-harvest management, supporting rural development, and increasing farmers’ income. However, most existing potato quality sorting techniques are mainly limited to laboratory-based or theoretical studies, while the market still lacks a fully integrated intelligent detection system suitable for real production conditions. Therefore, there is a strong need to develop an online potato quality detection method that can work effectively under actual post-harvest processing environments. This study developed a potato quality sorting approach based on cross-modal technology. For external quality assessment, an industrial camera was used to capture potato images, and the YOLOv5s algorithm was applied to detect surface defects such as green skin, sprouting, rotting, and mechanical damage. For internal quality evaluation, visible/near-infrared spectroscopy was used to collect spectral data, and a convolutional neural network model was employed to identify internal blackheart disease. The external defect detection model achieved a mean average precision of 0.892 at an IoU threshold of 0.5, while the internal defect detection model obtained an accuracy of 98.2%. The complete online detection system achieved a real-time dynamic defect detection rate of 91.3%, with an average detection time of 350 ms per potato Compared with studies conducted under ideal laboratory conditions, the results of this research are more practical because the system was tested in a real-time online production environment. This work provides a useful reference for developing intelligent online quality inspection systems for potatoes and other similar agricultural products.