基于半监督学习的三七叶片病害分割模型

Semi-supervised Model for Disease Segmentation in Panax notoginseng Leaves

摘要:

目前三七叶片病害分割存在标注数据稀缺、病害类别视觉相似性高和病斑边缘复杂问题，导致分割精度受限。为解决上述问题，本文提出多流扰动一致性约束的半监督分割模型（Multi-stream disturbance consistency constraint semi-supervised segmentation model，MDCC-SSeg）。首先，引入DINOv2作为模型主干网络，通过构建双流扰动机制生成具有差异性的增强视图，利用对比学习增强模型提高复杂病害形态的适应性;其次，提出注意力引导通道互补丢弃模块动态优化特征解耦，提升病理特征的表达能力;最后，引入动态阈值机制自适应数据分布调整伪标签筛选标准，确保未标注数据的有效利用。试验结果表明，在1/2标注数据量下，MDCC-SSeg模型对灰霉病、炭疽病等4类病害的整体均值交并比（mIoU）达到78.45%，优于U2PL、UniMateh和CorrMatch等其他分割模型，相比全监督SupBaseline模型提升10.83个百分点，其中灰霉病精度达86.10%，提升18.69个百分点，炭疽病精度达58.95%，提升12.08个百分点。消融试验结果和可视化也验证了本文方法具有强大的特征提取能力及低标注数据条件下的有效性，可为小样本三七叶片病害分割提供可靠解决方案。

Abstract:

Currently, the segmentation of Panax notoginseng leaf diseases faces several challenges, including scarce annotated data, high visual similarity among disease categories, and complex lesion boundaries, which collectively limit segmentation accuracy. To address these issues, a multi-stream disturbance consistency constraint semi-supervised segmentation model (MDCC-SSeg) was proposed. Firstly, DINOv2 was introduced as the backbone network. A dual-stream disturbance mechanism was constructed to generate diversified augmented views, and contrastive learning was leveraged to enhance the model's adaptability to complex disease patterns. Secondly, an attention-guided channel complementary dropout module was proposed to dynamically optimize feature decoupling, thereby improving the representation capability of pathological features. Finally, a dynamic thresholding mechanism was introduced to adaptively adjust pseudo-label selection criteria according to data distribution, ensuring the effective utilization of unlabeled data. Experimental results demonstrated that, under the setting of 1/2 labeled data, the proposed MDCC-SSeg achieved a mean intersection over union (mIoU) of 78.45% across four disease categories, including gray mold and anthracnose, outperforming other state-of-the-art segmentation methods such as U2PL, UniMatch, and CorrMatch. Compared with the fully supervised SupBaseline model, MDCC-SSeg achieved an overall improvement of 10.83 percentage points. Specifically, the segmentation accuracy for gray mold reached 86.10%, with an improvement of 18.69 percentage points, while anthracnose achieved 58.95%, improving by 12.08 percentage points. Furthermore, ablation studies and visualization results validated that the proposed method possessed strong feature extraction capability and demonstrated high effectiveness under low-annotation conditions, providing a reliable solution for small-sample segmentation of Panax notoginseng leaf diseases.