基于改进UNet++模型的葡萄黑腐病病斑分割和病害程度分级

Disease spot segmentation and disease degree classification of grape black rot based on improved UNet++model

为了解决葡萄病害图像边缘分割模糊和发病初期分割难的问题,基于PlantVillage数据集中的葡萄黑腐病图像,提出一种基于改进UNet++的葡萄黑腐病病斑分割模型。该模型在提取图像特征时:一方面,采用自适应软阈值化方法消除噪声影响,提高葡萄病斑边缘的分割精度;另一方面,采用长、短连接结合的方式构建UNet++中的跳跃式连接结构,降低模型的计算复杂度。同时,在模型的横向输出层中融合多尺度特征,增强病斑的语义信息,进一步提高目标分割精度。在该模型的损失函数中,将Dice损失函数和交叉熵损失函数进行线性加权组合,以解决病斑像素面积与叶片面积不平衡的问题。采用五折交叉验证进行模型训练与测试。结果显示,本文模型的像素准确率达到98.433%,平均交并比达到92.056%,病斑交并比为81.230%,Dice系数为0.941,均优于传统的UNet++模型。采用病斑占叶面积的比例对病害程度进行分级。结果表明,本文模型对病害等级的划分准确率达97.41%。该模型能精确实现对葡萄黑腐病病斑边缘和小病斑的分割,以及病害程度分级,具有良好的稳健性。

Based on the grape black rot images from PlantVillage dataset,an improved model for disease spot segmentation based on UNet++was proposed to solve the fuzzy edge segmentation and segmentation difficulties encountered at the early disese stage.For image feature extraction,on one hand,the adaptive soft thresholding method was introduced in the proposed model to improve the edge segmentation accuracy of grape disease image by filtering the influence of noise,on the other hand,the skip connection structure of UNet++was constructed by combining long and short connections to reduce the computational complexity of the model.Multi-scale features were fused in the lateral output layer of the model to enhance the semantic information of the disease spot and further improve the segmentation accuracy.In addition,the loss function of the model was weighted by adding Dice loss function to the cross-entropy loss function,to solve the imbalance between the pixel area of the disease spot and the leaf area.Five-fold cross validation was used for the model training and test.The results showed that the pixel accuracy of the proposed model was 98.433%,the mean intersection over union was 92.056%,the intersection over union for the disease spot was 81.230%,and the Dice coefficient was 0.941,which were all superior to the traditional UNet++model.Based on the area ratio of disease spot to leaf,the disease degree was classified,and the mean accuracy of disease degree classification was 97.41%.The proposed model could accurately segment the edge of diseased spots and minor diseased spots,realize the classification of disease degree with good robustness.