基于深度对比学习的番茄病害可信诊断

Trustworthy diagnosis of tomato disease based on deep contrastive learning

中文关键词: 番茄病害诊断  对比学习  不确定性估计  可信融合

英文关键词: tomato disease diagnosis  contrastive learning  uncertainty estimation  trustworthy fusion

中文摘要:

      现有基于对比学习的番茄病害诊断模型在聚合原始数据和增强数据时通常假设视角融合权重不变，导致番茄病害诊断准确度受限．为此，提出一种基于深度对比学习的番茄病害可信诊断方法．首先，构建番茄叶片图像数据增强模块，利用仿射变换等增强函数缓解对比学习框架中正负样本对不平衡问题．然后，定义深度对比模块，耦合神经网络和对比学习，捕获具有语义不变特征的鲁棒低维表示．同时，设计可信聚合模块，利用狄利克雷函数建模类别划分的不确定性估计，实现原始数据和增强数据决策信息的样本特异性可靠融合，极大促进番茄病害诊断可信性．2个公开数据集上的实验结果表明，所提方法在准确率、敏感性、F1值和精确率等评价指标上均优于9个基线方法，证明了所提方法的优越性和有效性．

英文摘要:

      The existing contrastive learning-based tomato disease diagnosis models usually assume that the view fusion weights remain constant when aggregating raw data and augmented data, which limits the accuracy of tomato disease diagnosis. To address this, a trustworthy diagnosis method for tomato disease based on deep contrastive learning is proposed. First, a tomato leaf image data augmentation module is constructed, using affine transformations and other augmentation functions to alleviate the imbalance problem between positive and negative sample pairs in the contrastive learning framework. Then, a deep contrastive module is defined, coupling neural networks and contrastive learning to capture robust low-dimensional representations with semantic-invariant features. Meanwhile, a reliable aggregation module is designed, using Dirichlet functions to model the uncertainty estimation in category division, enabling sample-specific reliable fusion of decision information from both raw data and augmented data. This greatly enhances the reliability of tomato disease diagnosis. Experimental results on two public datasets show that the proposed method outperforms nine baseline methods in terms of accuracy, sensitivity, F1-score and precision, demonstrating the superiority and effectiveness of the proposed method.