基于边界样本分位数的葡萄霜霉病自适应识别方法

Adaptive identification method of grape downy mildew based on quantile of boundary sample

针对葡萄霜霉病病斑组织图像阈值难以确定的问题,提出一种基于边界样本分位数的自适应阈值确定方法,通过高斯滤波识别病斑边界,并采用边界样本的50%分位数确定为病斑阈值。之后采用蒙特卡洛方法,通过随机采样方法估算病斑比例。结果表明,与其他阈值确定方法对比,所提方法能够自适应获取病斑灰度阈值,识别精度达到92.2%,明显高于其他阈值确定方法;与传统的机器学习方法对比,在识别精度上高于BP神经网络、卷积神经网络、支持向量机,略低于VGG16模型的94.3%与ResNet50模型的96.26%,但计算时间为1.410 s,远快于VGG16模型与ResNet50模型的5.588 s与20.317 s,说明方法能够在较短的运行时间内实现较高的精度。

Aiming at the problem of difficulty to determine the threshold of grape downy mildew lesion tissue image,an adaptive threshold determination method based on quantiles of boundary samples was proposed.The boundary of diseased spot was identified by Gaussian filtering,and the threshold of diseased spot was determined by 50%quantiles of boundary sample.Then Monte Carlo method was used to estimate the proportion of diseased spots by random sampling method.The results show that compared with other threshold determination methods,the proposed method can adaptively obtain the grayscale threshold of the lesion,with a recognition accuracy of 92.2%,which is significantly higher than other threshold determination methods.Compared with the traditional machine learning methods,the recognition accuracy of this method is higher than that of BP neural network,convolutional neural network,and support vector machine,slightly lower than 94.3%of the VGG16 model and 96.26%of the ResNet50 model.However,the calculation time is 1.410 s,which is much faster than 5.588 s and 20.317 s of the VGG16 model and ResNet50 model,indicating that this method can achieve high accuracy in a shorter running time.