基于语义分割的水稻叶瘟病分割与分级方法

Semantic segmentation-based method for rice leaf blast segmentation and grading

针对传统叶瘟病分割和分级方法在效率、准确率等方面存在的问题,提出一种基于语义分割的水稻叶瘟病分割与分级方法。首先,对CO39品种水稻叶片进行图像采集,使用Labelme标注软件对图像叶片和病斑进行标注,建立叶片数据集;然后采用不同的卷积神经网络作为U-Net、DeepLabV3+的主干特征提取网络,构建3种水稻叶片分割模型,分别为VGG16-UNet、ResNet50-UNet、MobileNetV2-DeepLabV3+,对水稻叶片、病斑进行分割,根据叶瘟病分级标准与等级计算公式,确定水稻叶片的叶瘟病等级,在此过程中,对比3种模型的分割性能。结果表明,VGG16-UNet模型为最优模型,在平均像素精度、平均交并比和F1分数上分别达到了86.87%、80.68%和88.48%,能够有效满足水稻叶瘟病分割和分级的实际需求。该方法为开发叶瘟病智能分级系统提供了理论依据,可为其他作物病害的分级研究提供参考。

A method based on semantic segmentation for rice leaf blast segmentation and grading was proposed to address the issues of efficiency and accuracy in traditional methods.Firstly,images of rice leaves from the CO39 variety were collected,and the leaf and lesions were annotated using the Labelme annotation software to create a leaf dataset.Then,three rice leaf segmentation models,namely VGG16-UNet,ResNet50-UNet,and MobileNetV2-DeepLabV3+,were constructed by using different convolutional neural networks as the backbone feature extraction networks.These models were used to segment rice leaves and lesions.Based on the grading criteria and grading formula of rice blast,the disease grade of rice leaves was determined.During this process,the segmentation performance of the three models was compared.The results showed that the VGG16-UNet model performs the best,achieving an average pixel accuracy,average intersection over union,and F1 score of 86.87%,80.68%and 88.48%,respectively.It effectively met the practical requirements of rice leaf blast segmentation and grading.The proposed method provided a theoretical basis for developing an intelligent grading system for rice blast and serves as a referenced for grading research on other crop diseases.