基于一致性半监督学习的苹果叶片病斑分割模型研究

Apple Leaf Spot Segmentation Model Based on Consistency Semi-supervised Learning

摘要:

快速准确的病斑分割对于病害严重程度评估及科学施药具有重要意义。基于深度学习的语义分割为构建高精度病斑分割模型提供了技术支撑。然而，苹果病斑标注费时费力。为了解决这一问题，以陇东苹果为研究对象，提出了一种基于轻量级一致性半监督学习框架的苹果叶片病斑分割模型。首先，遵循Mean Teacher半监督学习框架，使用2个轻量化的DeepLabV3+模型，构建病斑语义分割模型，以提高模型从有限标注数据中提取特征描述符的能力。其次，对比19种一致性正则化方法，发现MSE+Huber 组合对图像的细微差异更敏感、抗噪性更高，可提高模型对病斑过小、分布不均、边缘模糊的适应性。接着，使用贝叶斯优化算法对模型涉及的6个超参数进行寻优，以加快模型收敛速度和稳定性。结果表明，优化后模型仅使用30%的标注数据，病斑分割精确率达到95.60%，平均交并比为94.85%，平均像素准确率为96.50%。效果均优于全监督和自训练半监督学习框架。

关键词:苹果病害;病斑分割;半监督学习;一致性正则化;贝叶斯优化;DeepLabV3+

Abstract:

Rapid and accurate lesion segmentation was essential for assessing disease severity and ensuring precise pesticide application. Deep learning-based semantic segmentation offered the technical foundation necessary for developing high-precision disease detection models. However, the annotation of apple leaf spots was both time-consuming and labor-intensive. To address this issue, a model for apple leaf spot segmentation was proposed based on a lightweight consistency semi-supervised learning framework, using Longdong apples as the research subject. Firstly, following the Mean Teacher semi-supervised learning framework, two lightweight DeepLabV3+ models were utilized to build the lesion semantic segmentation model, which improved its ability to extract feature descriptors from limited annotated data. Secondly, a systematic comparison of 19 consistency regularization methods revealed that the combination of MSE and Huber was more sensitive to subtle image differences and exhibited higher noise resistance, thereby improving the model’s adaptability to small, unevenly distributed, and blurred-edge lesions. Next, a Bayesian algorithm was utilized to optimize six hyperparameters involved in the model, which accelerated convergence speed and enhanced stability. The results demonstrated that the optimized model, using only 30% of the annotated data, achieved a precision of 95.60%, a mean intersection over union (mIoU) of 94.85%, and a mean pixel accuracy (mPA) of 96.50%. These outcomes surpassed those of fully supervised and self-training semi-supervised learning frameworks. The findings offered agricultural practitioners an efficient and reliable tool for disease detection.

Key words:apple diseases;disease spot segmentation;semi-supervised learning;consistency regularization;Bayesian optimization;DeepLabV3+