基于改进级联神经网络的大豆叶部病害诊断模型

Diagnosis Model of Soybean Leaf Diseases Based on Improved Cascade Neural Network

针对大豆叶部病害性状特征与病种之间的模糊性和不确定性,将数字图像处理技术与神经网络智能推理技术相结合,充分挖掘大豆受病害胁迫后表现性状与病种之间的潜在规律,提出了基于改进级联神经网络的大豆病害诊断模型。首先利用自制载物模板无损采集大田大豆叶部病害数字图像,计算病斑区域的形状特征、颜色特征及纹理特征14维度特征参数;为突显各方面特征对于不同病害种类决定作用的差异性,构建各子神经网络并联的第1级网络,第2级网络的输入为第1级网络的输出,利用多维特征各自优势来自动取得病种模式推理规则,建立了用于大豆叶部病害自动诊断的两级级联神经网络模型,仿真实验准确率为97.67%;同时应用量子遗传计算优化级联神经网络参数,平均迭代次数为743,平均网络误差为0.000 995 445,提高了学习效率,实现了大豆叶部病害的高效自动诊断和精确测报,为大田农作物全面系统地开展作物病害监测、智能施药及自动防治提供了理论依据。

Crop disease is an important factor to restrict high-yielding,high-quality and high efficiency of products. Soybean is a critical crop,but incidence of soybean diseases increases year by year during their growth,so diagnosis of soybean diseases timely and accurately can provide reliable basis for prevention and control of soybean. Therefore,aiming at the fuzzy and uncertainty between disease traits and diseases of soybean leaf diseases,combining digital image possessing and neural network technology,the diagnosis model of soybean diseases was proposed based on improved cascade neural network after the potential rules of disease traits and diseases was fully mined. Firstly,the diseases images were acquitted by homemade slide template,the 14 dimensional characteristic parameters were calculated based on the geometry characteristic,color characteristic and texture characteristic of disease areas. Secondly,in order to highlight all aspects of characteristics for different kinds of diseases,the first level of each parallel neural network was constructed,the output of the first level was the input of the second level. Thirdly,the two slopes cascade neural network model was established for diagnosis soybean leaf diseases automatically,which based on inference rules of diseases using respective advantages of multidimensional characteristics,the simulation accuracy was 97. 67%. Meanwhile,the cascade neural network parameters were optimized by quantum genetic algorithm. The average number of iterations was 743,and the average network error was 0. 000 995 445. The proposed method realized the automatic diagnosis and precise forwards,which also provided important theory basis for disease monitoring and smart pesticide spraying.