Different Learning Model Fusion Method under Small Labeled Samples for Potatoes External Defect Detection

Chen, Fengnong, Xiang, Yuheng, Li, You, Chen, Zhaoqing, Wei, Qiquan, Sun, Hongwei
Potato Research · Springer Nature · 2026-10-13 · DOI: 10.1007/s11540-026-10147-3

Аннотация

Potato defect detection is a critical step in post-harvest sorting. Traditional supervised methods rely on large manually labeled datasets, which is time-consuming and labour-intensive. This study proposes a novel semi-supervised learning (SSL) approach for potato defect detection. We first compared multiple neural networks and selected ConvNeXt v2 as the backbone. We then evaluated various SSL architectures and selected FixMatch. The proposed ConvNeXt v2 + FixMatch framework achieved superior performance across all experiments. With only 50, 100, 200, and 400 labeled images, the test accuracies reached 0.880, 0.899, 0.937, and 0.943, respectively, outperforming fully supervised learning (FSL) baselines by 0.117, 0.118, 0.120, and 0.050. Furthermore, the model trained with limited labeled samples can serve as a seed model to automatically annotate unlabeled samples, enabling large-scale dataset labeling. These results demonstrate that SSL methods effectively leverage unlabeled images to enhance neural network performance for potato defect detection.

Читать у издателя →

Похожие статьи