关键词: 红枣缺陷检测, 域自适应, 元学习, 双层优化, L2正则化

Abstract: To address the challenge posed by variations in red date defects across different varieties and lighting conditions in automated sorting tasks, this study proposed a novel meta-learning-based algorithm for domain-adaptive defect detection. First, a cross-domain dataset was constructed by collecting images of red date defects from multiple varieties and environmental conditions.To mitigate sample imbalance,additional defect samples were generated using the StyleGAN3 network, and data augmentation techniques were applied to enhance the diversity of test dataset. Next, a bi-level optimization meta-learning framework was introduced for domain-adaptive red date defect detection. A convolutional neural network was employed as the base learner, while a dual-layer optimization strategy was used to construct the meta-learner. An L2 regularization term was incorporated into the loss function to reduce overfitting. Average accuracy was used as the evaluation metric. Ablation experiments were conducted on both the base learner and the meta-learner, and the proposed method was compared against various deep learning and meta-learning algorithms to validate its performance. Experimental results demonstrated that the proposed method achieves average accuracies of 78.6% on the original target domain dataset and 86.5% on the augmented datasets, outperforming the MAML algorithm by 6.4% and 7.6%, respectively. These findings confirm the methods effectiveness in adapting to cross-domain red date defect detection under diverse conditions.

Key words: jujubes defect detection, domain adaptation, meta learning, bi-level optimization, L2 regularization