Qiyu Zhang, Xinping Mo, Huan Zhang, Wenjing Li, Guangfei Liu, Ranbing Yang, Zhiguo Pan
The improved Inception ResNet combined with Savitzky-Golay preprocessing achieved the best performance, with five-fold cross-validation accuracy, precision, recall, and F1 score of 96.13 ± 0.82%, 96.21 ± 0.79%, 96.13 ± 0.82%, and 96.15 ± 0.80%, respectively. On the independent external validation set, the corresponding values were 94.27 ± 0.89%, 94.39 ± 0.85%, 94.27 ± 0.89%, and 94.22 ± 0.87%. Spectral analysis showed distinct deterioration-related responses in the visible and near-infrared regions, associated with surface colour variation, moisture redistribution, and internal quality degradation.
Potato is an important staple crop worldwide, and accurate tuber detection is essential for intelligent harvesting, yield estimation, and quality management. During mechanized harvesting, potatoes are conveyed with large amounts of soil clods on conveyor belts. The similar appearance between soil-covered tubers and clods makes accurate identification and separation difficult under practical operating conditions. Existing methods are often limited by scale variation, object occlusion, irregular boundaries, and complex backgrounds, restricting their application in real-time harvesting systems.This study proposes DFINE-PFM, an end-to-end instance segmentation framework for automatic potato-clod discrimination on harvesting conveyor belts. The framework incorporates four targeted improvements: a multi-scale feature diffusion module to enhance feature representation, a parallel large-kernel decomposition module to improve boundary extraction, a heterogeneous multi-branch convolution module to capture multi-granularity features with reduced redundancy, and an asymmetric receptive field convolution module to improve localization of small and irregular objects. A self-collected conveyor belt dataset under practical harvesting conditions was used for evaluation. DFINE-PFM achieved Precision, Recall, F1-score, and mAP values of 98.80%, 98.70%, 98.75%, and 98.03%, respectively, with an inference time of 43.41 ms. Compared with the baseline model, it improved Precision and mAP by 6.20 and 5.60 percentage points, respectively. Deployment experiments demonstrated stable real-time performance at 20 FPS under conveyor speeds of 0.1-0.6 m/s. The proposed method provides an effective solution for automated potato-clod separation and intelligent harvesting quality monitoring.