CN106446942A - Crop disease identification method based on incremental learning - Google Patents

Abstract

Translated fromChinese

本发明提供一种基于增量学习的农作物病害识别方法，在新数据到达时在原有学习结果的基础上继续学习，具有渐进学习的能力，即在有效保留已有知识的情况下，能从分批获得的新样本中获取新知识，逐步提高性能；首先，收集病害样本库，利用负相关集成神经网络为主要技术手段对样本库病害图像模拟增量学习，从而确定负相关学习系统的初始参数，并用此初始参数初始化一个基于负相关学习的集成神经网络分类器，利用初始阶段的样本训练此分类器；增量学习阶段，当专家将新的样本数据加入样本库时，基于负相关学习的集成神经网络分类器只对新加入的样本数据进行训练并更新分类器，从而达到增量学习的目的；最后，将病害图片的诊断结果及防治措施反馈给用户,从而准确的识别诊断病虫害，达到农作物综合防治的目的。

The invention provides a method for identifying crop diseases based on incremental learning. When new data arrives, it continues to learn on the basis of the original learning results, and has the ability of gradual learning, that is, under the condition of effectively retaining existing knowledge, it can learn from points to points. Acquire new knowledge from the new samples obtained in batches, and gradually improve performance; first, collect the disease sample database, and use the negative correlation integrated neural network as the main technical means to simulate incremental learning on the disease image of the sample database, so as to determine the initial parameters of the negative correlation learning system , and use this initial parameter to initialize an integrated neural network classifier based on negative correlation learning, and use the samples in the initial stage to train the classifier; in the incremental learning stage, when experts add new sample data to the sample library, the negative correlation learning based The integrated neural network classifier only trains the newly added sample data and updates the classifier, so as to achieve the purpose of incremental learning; finally, the diagnosis results and control measures of the disease pictures are fed back to the user, so as to accurately identify and diagnose the pests and diseases, and achieve The purpose of integrated crop management.

Description

Translated fromChinese

基于增量学习的农作物病害识别方法Crop Disease Identification Method Based on Incremental Learning

技术领域technical field

本发明涉及模式识别和机器学习领域，尤其涉及一种基于增量学习的农作物病害识别方法。The invention relates to the field of pattern recognition and machine learning, in particular to a method for identifying crop diseases based on incremental learning.

背景技术Background technique

传统的农作物病害识别方法，通常是从病害样本数据库中提取的特征向量使用神经网络进行训练得到一个分类器，当用户上传病害图像时，根据训练得到分类器对病害种类进行识别。为了得到较高的分类精度，要求训练样本集越完备越好，但在实际应用中，完整的样本集很难获得，由于对问题理解的局限性和实际应用中的复杂性，很难精确、完整地定义训练样本集，使得一次性获得和保存所有数据的代价会随着时间增长变得越来越高；许多实际问题也不允许等到获得所有数据后再进行学习，当新样本加入时，为了得到更精确的学习结果，需要将以前训练集中的数据和新的训练集中的数据合并以后进行训练。这样操作复杂,每次都需要重新训练神经网络也需要消耗大量的时间和内存容量。特别地，如果训练数据特别大，内存容量可能无法满足训练神经网络的需求。The traditional crop disease identification method usually uses the feature vector extracted from the disease sample database to train a classifier using a neural network. When the user uploads a disease image, the classifier is trained to identify the disease type. In order to obtain a higher classification accuracy, the training sample set is required to be as complete as possible. However, in practical applications, it is difficult to obtain a complete sample set. Due to the limitations of understanding the problem and the complexity of practical applications, it is difficult to accurately and Completely define the training sample set, so that the cost of obtaining and saving all the data at one time will become higher and higher as time increases; many practical problems do not allow learning after obtaining all the data. When new samples are added, In order to obtain more accurate learning results, it is necessary to combine the data in the previous training set with the data in the new training set for training. This operation is complicated, and the neural network needs to be retrained every time, which also consumes a lot of time and memory capacity. In particular, if the training data is extremely large, the memory capacity may not be sufficient for training the neural network.

发明内容Contents of the invention

本发明的目的在于提供一种识别效率高、识别种类多的基于增量学习的病害识别方法，该方法具有病害同类之间变化鲁棒性强、不同类之间相似性敏感的特点。The purpose of the present invention is to provide a disease recognition method based on incremental learning with high recognition efficiency and multiple recognition types. The method has the characteristics of strong robustness to changes between diseases of the same type and sensitivity to similarities between different types.

为实现上述目的发明采用如下方案：In order to realize the above object, the invention adopts the following scheme:

基于增量学习的农作物病害识别方法，包括以下步骤：A crop disease identification method based on incremental learning, including the following steps:

步骤1收集原始病害图像进行图像预处理；Step 1 collects the original disease image for image preprocessing;

步骤1-1收集病害原始图像，标记其种类；Step 1-1 collects the original image of the disease and marks its type;

步骤1-2对病害图像进行灰度变换，把彩色病害图像转换成灰度图像,提取每一个像素的R、G、B分量，转换公式如下：Step 1-2: Carry out grayscale transformation on the diseased image, convert the color diseased image into a grayscale image, and extract the R, G, and B components of each pixel. The conversion formula is as follows:

Gray＝0.299*R+0.587*G+0.114*BGray＝0.299*R+0.587*G+0.114*B

步骤1-3对病害图像进行图像增强，先使用中值滤波进行去噪，接着对去噪后的图像进行直方图均衡化；Steps 1-3 perform image enhancement on the diseased image, first use median filtering to denoise, and then perform histogram equalization on the denoised image;

步骤1-4对病害图像进行分割，采用阈值分割，将病害图像中病斑与叶片背景进行分离，获得只含有病斑的图像；Steps 1-4 segment the disease image, and use threshold segmentation to separate the lesion in the disease image from the leaf background to obtain an image containing only the lesion;

步骤1-5对病害图像进行轮廓提取；Steps 1-5 carry out contour extraction to the disease image;

步骤1-6对病害图像进行病害提取，将轮廓图像与原图叠加进行与运算，得到去除了叶片背景的病斑图像；Steps 1-6 carry out disease extraction on the disease image, superimpose the outline image and the original image and perform AND operation, and obtain the lesion image with the leaf background removed;

步骤2对经过预处理的病害图像进行特征提取，包括颜色特征、纹理特征以及形态特征，把提取的三方面的特征作为识别分类的特征向量，构建样本数据库；Step 2 is to extract features from the preprocessed disease images, including color features, texture features and morphological features, and use the extracted three features as feature vectors for identification and classification to build a sample database;

步骤3利用负相关集成神经网络对样本库病害图像进行演化计算模拟增量学习过程，确定负相关学习系统的初始参数，并用此初始参数初始化一个基于负相关学习的集成神经网络分类器；Step 3: Use the negative correlation integrated neural network to perform evolution calculation on the sample library disease image to simulate the incremental learning process, determine the initial parameters of the negative correlation learning system, and use this initial parameter to initialize an integrated neural network classifier based on negative correlation learning;

步骤3-1对于病害样本训练集合D，Step 3-1 For the disease sample training set D,

D＝{(x(1),d(1)),…,(x(N),d(N))}D={(x(1),d(1)),...,(x(N),d(N))}

x,d分别表示样本输入和输出，N为训练样本数；x, d represent sample input and output respectively, and N is the number of training samples;

步骤3-2该系统由M个子分类器的个体神经网络集成而成，采用平均输出作为集成系统的输出：Step 3-2 The system is integrated by the individual neural networks of M sub-classifiers, and the average output is used as the output of the integrated system:

其中F_i(n)是第n个训练样本作为输入时个体神经网络i的输出，F(n)是第n个训练样本作为输入时集成系统的输出；where F_i (n) is the output of the individual neural network i when the nth training sample is taken as input, and F(n) is the output of the integrated system when the nth training sample is taken as input;

负相关学习在每个个体网络的误差函数中引入了一个相关惩罚项，使得个体网络的误差与其它网络的误差呈负相关性，进而使得所有网络能够在训练集D上同时并且交互地进行训练。该误差函数定义如下：Negative correlation learning introduces a correlation penalty term in the error function of each individual network, so that the error of the individual network is negatively correlated with the errors of other networks, so that all networks can be trained simultaneously and interactively on the training set D . The error function is defined as follows:

E_i(n)是第n个训练样本作为输入时个体网络i的误差函数；公式(2)中第一项是个体网络i的经验风险函数，第二项中p_i(n)为相关性惩罚函数；通过最小化p_i，使得每个个体网络的误差和其余网络的误差呈负相关性；通过调节λ，0≤λ≤1来控制惩罚力度；E_i (n) is the error function of individual network i when the nth training sample is used as input; the first item in formula (2) is the empirical risk function of individual network i, and p_i (n) in the second item is the correlation Penalty function; by minimizing p_i , the error of each individual network is negatively correlated with the error of the rest of the network; by adjusting λ, 0≤λ≤1 to control the degree of punishment;

相关性惩罚函数p_i(n)定义如下:The correlation penalty function p_i (n) is defined as follows:

第n个训练样本作为输入时，E_i(n)关于网络i的输出的偏导为:When the nth training sample is used as input, the partial derivative of E_i (n) with respect to the output of network i is:

这里假定F(n)相对于F_i(n)是个常量,采用BP算法以顺序模式更新权值，即对于每个输入训练样本，所有个体网络的权值更新是通过公式(4)同时进行的，所有样本训练一遍，称为一个epoch；It is assumed here that F(n) is a constant relative to F_i (n), and the weights are updated in a sequential mode using the BP algorithm, that is, for each input training sample, the weights of all individual networks are updated simultaneously through formula (4) , all samples are trained once, called an epoch;

训练过程中，所有的个体网络通过误差公式中的惩罚项相互作用；个体网络i的权值不仅要使F_i(n)与d(n)的差异最小，还要使F(n)与d(n)的差异最小；During the training process, all individual networks interact through the penalty term in the error formula; the weight of individual network i should not only minimize the difference between F_i (n) and d(n), but also make F(n) and d (n) has the smallest difference;

步骤3-2通过演化学习模拟增量学习过程，获得一组最优的负相关集成神经网络的负相关惩罚因子强度、子网络数、子网络隐节点、子网络学习系数、学习误差及初始权重等参数；Step 3-2 Simulate the incremental learning process through evolutionary learning, and obtain a set of optimal negative correlation penalty factor strength, number of sub-networks, hidden nodes of sub-networks, learning coefficients of sub-networks, learning errors and initial weights of the integrated neural network of negative correlations and other parameters;

该方法包括以下步骤：The method includes the following steps:

(1)将初始训练样本集S随机分成m个子集{S₀,S₁,…S_m}，初始迭代次数k＝0；(1) Randomly divide the initial training sample set S into m subsets {S₀ , S₁ ,...S_m }, and the initial number of iterations k=0;

(2)初始化N个负相关神经网络集成作为初始群体，每一个个体对应一个神经网络集成，每个网络集成的初始参数在取值范围内随机产生，设训练批次I＝0；(2) Initialize N negatively correlated neural network integrations as the initial group, each individual corresponds to a neural network integration, and the initial parameters of each network integration are randomly generated within the value range, assuming training batch I=0;

(3)新加入一批样本S_I，用S_I，0<I<m通过负相关学习方法训练每一个网络集成，直到在S_I上错误率小于学习误差；(3) Add a new batch of samples S_I , use S_I , 0<I<m to train each network integration through the negative correlation learning method, until the error rate on S_I is less than the learning error;

(4)如果I大于分批样本数m，则转向下一步，否则I＝I+1返回步骤(3)；(4) if I is greater than the batch sample number m, then turn to the next step, otherwise I=I+1 returns to step (3);

(5)在确认集上测试每个负相关网络集成的分类正确率，作为每个个体的适应度；(5) Test the classification accuracy rate of each negative correlation network integration on the confirmation set, as the fitness of each individual;

(6)删除N/2个适应度较差的个体，对于剩余的适应度较高的个体，每个个体随机选择另一个个体，通过交叉、变异产生一个后代；(6) Delete N/2 individuals with poor fitness, and for the remaining individuals with high fitness, each individual randomly selects another individual, and generates an offspring through crossover and mutation;

(7)得到新一代群体，并用对应的参数重新初始化群体；(7) Obtain a new generation of groups, and reinitialize the groups with corresponding parameters;

(8)如果k大于最大迭代次数，则转向下一步，否则k＝k+l，I＝0转向步骤(3)；(8) if k is greater than the maximum number of iterations, then turn to the next step, otherwise k=k+1, and I=0 turns to step (3);

(9)选择适应度最高的个体的参数作为负相关神经网络集成增量学习算法的参数；(9) Select the individual parameters with the highest fitness as the parameters of the negative correlation neural network integration incremental learning algorithm;

所采用的交叉策略是，后代个体的参数从父代个体的参数所确定区间内随机选择，公式如下所示：The crossover strategy adopted is that the parameters of offspring individuals are randomly selected from the interval determined by the parameters of parent individuals, and the formula is as follows:

x′_ij＝x_i+(x_j-x_i)·rand(0,1)+k·N(0,1)x′_ij ＝x_i +(x_j -x_i )·rand(0,1)+k·N(0,1)

公式中x′_ij中表示的是父代i，j的子代个体参数且j≠i，rand(0，1)表示(0，1)区间的均匀分布随机数，N(0,1)表示(0,1)正态分布随机数,k为变异强度；In the formula, x′_ij represents the individual parameters of the offspring of the parent i, j and j≠i, rand(0,1) represents a uniformly distributed random number in the (0,1) interval, and N(0,1) represents (0,1) Normal distribution random number, k is the variation intensity;

每一个网络集成都经过一定的增量学习过程，然后进行演化选择，最终获得的个体针对当前问题将具有良好的增量学习能力；Each network integration undergoes a certain incremental learning process, and then undergoes evolutionary selection, and the finally obtained individuals will have good incremental learning capabilities for the current problem;

步骤3-3根据步骤3-2中获得的初始参数初始化一个基于负相关学习的神经网络集成分类器，利用初始阶段的样本训练此分类器；Step 3-3 initializes a neural network ensemble classifier based on negative correlation learning according to the initial parameters obtained in step 3-2, and uses the samples in the initial stage to train the classifier;

步骤4增量学习阶段：当将新的病害样本数据加入样本库时，基于负相关学习的集成神经网络分类器只对新加入的样本数据进行训练并更新分类器，从而达到增量学习的目的；Step 4 Incremental learning stage: When new disease sample data is added to the sample library, the integrated neural network classifier based on negative correlation learning only trains the newly added sample data and updates the classifier, so as to achieve the purpose of incremental learning ;

步骤5病害识别阶段：上传病害图片，通过图像预处理提取病害特征，然后将病害特征输入训练好的分类器进行识别。Step 5 Disease identification stage: Upload the disease picture, extract the disease features through image preprocessing, and then input the disease features into the trained classifier for identification.

进一步，所述步骤1-4在灰度直方图上选取阈值，进行分割，采用基于OTSU算法和基本粒子群优化算法的双阈值进行图像分割：Further, the step 1-4 selects a threshold on the grayscale histogram to perform segmentation, and adopts a dual threshold based on the OTSU algorithm and the basic particle swarm optimization algorithm to perform image segmentation:

OTSU自适应阈值求法与粒子群算法结合，将OTSU算法作为粒子群算法的适应值函数，计算每个粒子的适应度与最优阈值相比较，经过多次迭代最后取得联合算法优化的双阈值，利用所取得的阈值将病害图像中病斑与叶片背景进行分离。The OTSU adaptive threshold calculation method is combined with the particle swarm algorithm, and the OTSU algorithm is used as the fitness value function of the particle swarm algorithm to calculate the fitness of each particle and compare it with the optimal threshold. After multiple iterations, the double threshold optimized by the joint algorithm is finally obtained. The obtained threshold is used to separate the lesion from the leaf background in the disease image.

进一步，所述步骤1-5采用Canny算法对病斑轮廓进行检测，用高斯滤波器平滑病斑图像，用一阶偏导有限差分计算病斑图像梯度幅值和方向，对梯度幅值进行非极大值抑制，最后用双阈值算法检测和连接边缘。Further, the steps 1-5 use the Canny algorithm to detect the lesion outline, use a Gaussian filter to smooth the lesion image, and use the first-order partial derivative finite difference to calculate the gradient magnitude and direction of the lesion image, and perform non-destructive analysis on the gradient magnitude. Maximum suppression, and finally a dual-threshold algorithm to detect and connect edges.

进一步，所述步骤2颜色特征提取时，综合运用RGB颜色空间、HIS颜色空间和YCbCr颜色空间分析病斑颜色，提取RGB颜色空间原始颜色R、G、B分量，归一化的颜色分量r、g、b，HIS颜色空间色调H、亮度I、饱和度S，YCbCr颜色空间色彩Cb、Cr值，以及一阶矩、二阶矩和三阶矩，共14个颜色特征向量；Further, during the color feature extraction in step 2, comprehensively use RGB color space, HIS color space and YCbCr color space to analyze lesion color, extract RGB color space original color R, G, B components, normalized color components r, g, b, HIS color space hue H, brightness I, saturation S, YCbCr color space color Cb, Cr value, and first-order moment, second-order moment and third-order moment, a total of 14 color feature vectors;

进一步，所述步骤2纹理特征提取时，利用灰度共生矩阵算法和计盒维数法分析病害图像纹理分布，提取熵、能量、惯性、对比度、相关性和相关信息测度以及分形维数等7个纹理特征；Further, during the texture feature extraction in the step 2, the texture distribution of the diseased image is analyzed by using the gray level co-occurrence matrix algorithm and the box counting dimension method, and the entropy, energy, inertia, contrast, correlation and related information measures and fractal dimension etc. are extracted7 a texture feature;

进一步，所述步骤2形态特征提取时，利用区域标记法和区域跟踪算法计算病斑面积、周长、圆度、球状性、形态因子、离散指数、等效面积半径以及内切圆半径等8个形态特征向量。Further, during the extraction of morphological features in the step 2, the lesion area, perimeter, roundness, sphericity, shape factor, dispersion index, equivalent area radius, and inscribed circle radius, etc., are calculated using the area labeling method and the area tracking algorithm. A morphological feature vector.

本发明基于增量学习的农作物病害识别方法，首先以农作物病害为对象，提取病害的颜色特征、纹理特征和形态特征构建特征向量；在初期阶段引入负相关学习策略，利用BP神经网络作为子网络构建集成神经网络，然后利用样本数据模拟增量学习的过程确定负相关学习系统的初始参数，并用此初始参数初始化一个基于负相关学习的集成神经网络分类器，利用初始阶段的样本训练此分类器，将新的样本数据加入样本库时，基于负相关学习的集成神经网络分类器只对新加入的样本数据进行训练并更新分类器，从而达到增量学习的目的；当有用户上传病害样本时，利用该分类器可以识别出该病害的种类，并给出相应的诊断方法，从而达到农作物综合防治的目的。The crop disease recognition method based on incremental learning in the present invention first takes the crop disease as the object, extracts the color feature, texture feature and morphological feature of the disease to construct a feature vector; introduces a negative correlation learning strategy in the initial stage, and uses the BP neural network as a sub-network Construct an integrated neural network, and then use the sample data to simulate the incremental learning process to determine the initial parameters of the negative correlation learning system, and use this initial parameter to initialize an integrated neural network classifier based on negative correlation learning, and use the samples in the initial stage to train the classifier , when new sample data is added to the sample library, the integrated neural network classifier based on negative correlation learning only trains the newly added sample data and updates the classifier, so as to achieve the purpose of incremental learning; when a user uploads a diseased sample , using the classifier can identify the type of the disease, and give the corresponding diagnosis method, so as to achieve the purpose of integrated crop control.

增量学习算法能够在新数据到达时在原有学习结果的基础上继续学习，具有渐进学习的能力，即在有效保留已有知识的情况下，能从分批获得的新样本中获取新知识，逐步提高性能。首先，收集病害样本库，利用负相关集成神经网络为主要技术手段对样本库病害图像模拟增量学习，从而确定负相关学习系统的初始参数，并用此初始参数初始化一个基于负相关学习的集成神经网络分类器，利用初始阶段的样本训练此分类器；增量学习阶段，当专家将新的样本数据加入样本库时，基于负相关学习的集成神经网络分类器只对新加入的样本数据进行训练并更新分类器，从而达到增量学习的目的；最后，将病害图片的诊断结果及防治措施反馈给用户,从而准确的识别诊断病虫害，达到农作物综合防治的目的。The incremental learning algorithm can continue learning on the basis of the original learning results when new data arrives, and has the ability of progressive learning, that is, it can acquire new knowledge from new samples obtained in batches while effectively retaining existing knowledge. Gradually improve performance. First, collect the disease sample database, and use the negative correlation integrated neural network as the main technical means to simulate incremental learning on the disease image of the sample database, so as to determine the initial parameters of the negative correlation learning system, and use this initial parameter to initialize an integrated neural network based on negative correlation learning. Network classifier, which uses the samples in the initial stage to train the classifier; in the incremental learning stage, when experts add new sample data to the sample library, the integrated neural network classifier based on negative correlation learning only trains the newly added sample data And update the classifier, so as to achieve the purpose of incremental learning; finally, the diagnosis results of the disease pictures and the control measures are fed back to the user, so as to accurately identify and diagnose the diseases and insect pests, and achieve the purpose of comprehensive crop control.

附图说明Description of drawings

图1为本发明的流程图Fig. 1 is a flowchart of the present invention

图2为发明基于增量学习的病害识别运行时的流程图Figure 2 is a flow chart of the runtime of the invention based on incremental learning for disease identification

具体实施方式detailed description

为使发明的目的、技术方案和优点更加清楚，下面发明中的技术方案进行清楚、完整地描述，显然，所描述的实施例是发明一部分实施例，而不是全部的实施例。基于发明中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于发明保护的范围。In order to make the purpose, technical solutions and advantages of the invention clearer, the technical solutions in the following invention are clearly and completely described. Apparently, the described embodiments are part of the embodiments of the invention, not all of them. Based on the embodiments of the invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts belong to the protection scope of the invention.

如图1所示，本发明基于增量学习的农作物病害识别方法，首先以农作物病害为对象，提取病害的颜色特征、纹理特征和形态特征构建特征向量；在初期阶段引入负相关学习策略，利用BP神经网络作为子网络构建集成神经网络，然后利用样本数据模拟增量学习的过程确定负相关学习系统的初始参数，并用此初始参数初始化一个基于负相关学习的集成神经网络分类器，利用初始阶段的样本训练此分类器，将新的样本数据加入样本库时，基于负相关学习的集成神经网络分类器只对新加入的样本数据进行训练并更新分类器，从而达到增量学习的目的；当有用户上传病害样本时，利用该分类器可以识别出该病害的种类，并给出相应的诊断方法，从而达到农作物综合防治的目的。As shown in Fig. 1, the present invention is based on incremental learning crop disease identification method, at first takes crop disease as object, extracts the color feature, texture feature and morphological feature of disease to construct feature vector; The BP neural network is used as a sub-network to build an integrated neural network, and then use the sample data to simulate the incremental learning process to determine the initial parameters of the negative correlation learning system, and use this initial parameter to initialize an integrated neural network classifier based on negative correlation learning. When the new sample data is added to the sample library, the integrated neural network classifier based on negative correlation learning only trains the newly added sample data and updates the classifier, so as to achieve the purpose of incremental learning; When a user uploads a disease sample, the classifier can be used to identify the type of the disease and provide a corresponding diagnosis method, so as to achieve the purpose of integrated crop control.

参照图2.本发明基于增量学习的农作物病害识别方法具体步骤如下：With reference to Fig. 2. the present invention is based on the crop disease identification method concrete steps of incremental learning as follows:

步骤1，收集病害原始图像，标记其种类。对病害图像进行预处理包括灰度变换、图像增强、图像分割、轮廓提取、病害提取等处理。Step 1. Collect the original images of the disease and mark its types. The preprocessing of disease images includes grayscale transformation, image enhancement, image segmentation, contour extraction, and disease extraction.

所述灰度变换，拍照采集的病害图像为彩色图像，需要将其转换成对应的灰度图像，为把彩色病害图像转换成灰度图像,需提取每一个像素的R、G、B分量，转换公式如下：The grayscale conversion, the disease image collected by taking pictures is a color image, which needs to be converted into a corresponding grayscale image. In order to convert the color disease image into a grayscale image, the R, G, and B components of each pixel need to be extracted. The conversion formula is as follows:

Gray＝0.299*R+0.587*G+0.114*BGray＝0.299*R+0.587*G+0.114*B

所述图像增强，先使用中值滤波进行去噪，过滤噪声，同时减少细节损失，接着对去噪后的图像进行直方图均衡化，不但可以去除噪声，还能得到对比度比较好的病害图像。In the image enhancement, median filtering is first used for denoising to filter noise while reducing detail loss, and then histogram equalization is performed on the denoised image, which can not only remove noise, but also obtain a diseased image with better contrast.

所述图像分割，采用的是阈值分割，分割图像目标是将病害图像中病斑与叶片背景进行分离，获得只含有病斑的图像，以消除噪声，得到更精确的病斑特征，以便后续对病斑进行特征提取，在灰度直方图上选取阈值，进行分割，然而阀值分割性能取决于阈值的选取。采用基于OTSU算法和基本粒子群优化算法的双阈值图像分割：The image segmentation adopts threshold segmentation, and the target of image segmentation is to separate the lesion in the disease image from the leaf background, obtain an image containing only the lesion, eliminate noise, and obtain more accurate features of the lesion, so that subsequent Feature extraction is performed on the lesion, and a threshold is selected on the gray histogram for segmentation. However, the threshold segmentation performance depends on the selection of the threshold. Using the dual-threshold image segmentation based on OTSU algorithm and basic particle swarm optimization algorithm:

OTSU自适应阈值求法与粒子群算法结合，将OTSU算法作为粒子群算法的适应值函数，来计算每个粒子的适应度与最优阈值相比较，经过多次迭代最后取得联合算法优化的双阈值，利用所取得的阈值就可以将图像背景和目标区分开来。The OTSU adaptive threshold calculation method is combined with the particle swarm algorithm, and the OTSU algorithm is used as the fitness value function of the particle swarm algorithm to calculate the fitness of each particle and compare it with the optimal threshold. After multiple iterations, the double threshold optimized by the joint algorithm is finally obtained. , the image background and the target can be distinguished by using the obtained threshold.

所述轮廓提取，病害叶片的病斑含有丰富的形态信息，而病斑的一些形状特征蕴含在病斑轮廓里，而形态特征的参数依此来计算，因此需要进一步提取病斑的轮廓，采用Canny算法对病斑轮廓进行检测，具体方法为用高斯滤波器平滑病斑图像，用一阶偏导有限差分计算病斑图像梯度幅值和方向，对梯度幅值进行非极大值抑制，最后用双阈值算法检测和连接边缘。In the outline extraction, the lesion of the diseased leaf contains rich morphological information, and some shape features of the lesion are contained in the outline of the lesion, and the parameters of the morphological feature are calculated according to this, so it is necessary to further extract the outline of the lesion, using The Canny algorithm is used to detect the lesion outline. The specific method is to smooth the lesion image with a Gaussian filter, calculate the gradient amplitude and direction of the lesion image with the first-order partial derivative finite difference, and suppress the gradient amplitude by a non-maximum value. Edges are detected and connected with a double threshold algorithm.

所述病害提取，将轮廓图像与原图叠加进行与运算，得到去除了叶片背景的病斑图像，病斑部位被清晰地分离出来。In the extraction of the disease, the contour image is superimposed on the original image and performed an AND operation to obtain a lesion image with the background of the leaf removed, and the lesion is clearly separated.

经过预处理的病害图像进行特征提取：颜色特征、纹理特征以及形态特征，把提取的三方面的特征作为识别分类的特征向量。Feature extraction is performed on the preprocessed disease image: color feature, texture feature and morphological feature, and the extracted three aspects of features are used as feature vectors for recognition and classification.

所述颜色特征，综合运用RGB颜色空间、HIS颜色空间和YCbCr颜色空间分析病斑颜色，提取RGB颜色空间原始颜色R、G、B分量，归一化的颜色分量r、g、b，HIS颜色空间色调H、亮度I、饱和度S，YCbCr颜色空间色彩Cb、Cr值，以及一阶矩、二阶矩和三阶矩，共14个颜色特征向量。Described color characteristic, comprehensively utilize RGB color space, HIS color space and YCbCr color space to analyze lesion color, extract RGB color space original color R, G, B component, normalized color component r, g, b, HIS color Space hue H, brightness I, saturation S, YCbCr color space color Cb, Cr value, and first-order moment, second-order moment and third-order moment, a total of 14 color feature vectors.

所述纹理特征，利用灰度共生矩阵算法和计盒维数法分析病害图像纹理分布，提取墒、能量、惯性、对比度、相关性和相关信息测度以及分形维数等7个纹理特征。For the texture features, use the gray level co-occurrence matrix algorithm and the box counting dimension method to analyze the texture distribution of the disease image, and extract 7 texture features such as entropy, energy, inertia, contrast, correlation and related information measurement, and fractal dimension.

所述形态特征，利用区域标记法和区域跟踪算法计算病斑面积、周长、圆度、球状性、形态因子、离散指数、等效面积半径以及内切圆半径等8个形态特征向量。For the morphological features, 8 morphological feature vectors, including lesion area, perimeter, roundness, sphericity, form factor, discrete index, equivalent area radius, and inscribed circle radius, are calculated using the area labeling method and the area tracking algorithm.

其中F_i(n)是第n个训练样本作为输入时个体神经网络i的输出，F(n)是第n个训练样本作为输入时集成系统的输出。where F_i (n) is the output of the individual neural network i when the nth training sample is taken as input, and F(n) is the output of the integrated system when the nth training sample is taken as input.

负相关学习在每个个体网络的误差函数中引入了一个相关惩罚项，使得个体网络的误差与其它网络的误差呈负相关性，进而使得所有网络能够在训练集D上同时并且交互地进行训练。该误差函数定义如下：Negative correlation learning introduces a correlation penalty term in the error function of each individual network, so that the error of the individual network is negatively correlated with the errors of other networks, so that all networks can be trained simultaneously and interactively on the training set D . The error function is defined as follows:

E_i(n)是第n个训练样本作为输入时个体网络i的误差函数。公式中第一项是个体网络i的经验风险函数，第二项中p_i(n)为相关性惩罚函数。通过最小化p_i，使得每个个体网络的误差和其余网络的误差呈负相关性。可以通过调节λ(0≤λ≤1)来控制惩罚力度。相关性惩罚函数p_i(n)定义如下E_i (n) is the error function of individual network i when the nth training sample is taken as input. The first item in the formula is the empirical risk function of individual network i, and the second item p_i (n) is the correlation penalty function. By minimizing p_i , the error of each individual network is negatively correlated with the errors of the rest of the networks. The penalty can be controlled by adjusting λ (0≤λ≤1). The correlation penalty function p_i (n) is defined as follows

第n个训练样本作为输入时，E_i(n)关于网络i的输出的偏导为When the nth training sample is used as input, the partial derivative of E_i (n) with respect to the output of network i is

假定F(n)相对于F_i(n)是个常量。采用BP算法以顺序模式更新权值。即对于每个输入训练样本，所有个体网络的权值更新是通过公式(4)同时进行的。所有样本训练一遍。Assume that F(n) is constant with respect to F_i (n). The weights are updated in a sequential mode using the BP algorithm. That is, for each input training sample, the weight updates of all individual networks are performed simultaneously by formula (4). All samples are trained once.

训练过程中，所有的个体网络通过误差公式中的惩罚项相互作用。个体网络i的权值不仅要使F_i(n)与d(n)的差异最小，还要使F(n)与d(n)的差异最小。也就是说，在训练一个网络的时候，负相关学习要兼顾所有其它网络的学习情况。During training, all individual networks interact through penalty terms in the error formula. The weight of individual network i should not only minimize the difference between F_i (n) and d(n), but also minimize the difference between F(n) and d(n). That is to say, when training a network, negative correlation learning should take into account the learning situation of all other networks.

该方法包括以下步骤：The method includes the following steps:

(1)将初始训练样本集S随机分成m个子集{S0，S1，.…Sm)，k＝0。(1) The initial training sample set S is randomly divided into m subsets {S0, S1, . . . Sm), k=0.

(2)初始化N个负相关神经网络集成作为初始群体，每一个个体对应一个神经网络集成，每个网络集成的初始参数在取值范围内随机产生，设训练批次I＝0；(2) Initialize N negatively correlated neural network integrations as the initial group, each individual corresponds to a neural network integration, and the initial parameters of each network integration are randomly generated within the value range, assuming training batch I=0;

(3)新加入一批样本S_I，用S_I，0<I<m通过负相关学习方法训练每一个网络集成，直到在S_I上错误率小于学习误差。(3) Add a new batch of samples S_I , use S_I , 0<I<m to train each network ensemble through the negative correlation learning method, until the error rate on S_I is less than the learning error.

(4)如果I大于分批样本数m，则转向下一步，否则I＝I+1返回步骤(3)；(4) if I is greater than the batch sample number m, then turn to the next step, otherwise I=I+1 returns to step (3);

(5)在确认集上测试每个负相关网络集成的分类正确率，作为每个个体的适应度。(5) Test the classification accuracy rate of each negative correlation network integration on the confirmation set, as the fitness of each individual.

(5)在确认集上测试每个负相关网络集成的分类正确率，作为每个个体的适应度；(5) Test the classification accuracy rate of each negative correlation network integration on the confirmation set, as the fitness of each individual;

(6)删除N/2个适应度较差的个体，对于剩余的适应度较高的个体，每个个体随机选择另一个个体，通过交叉、变异产生一个后代。(6) Delete N/2 individuals with poor fitness, and for the remaining individuals with high fitness, each individual randomly selects another individual to generate an offspring through crossover and mutation.

(7)得到新一代群体，并用对应的参数重新初始化群体。(7) Get a new generation of groups, and reinitialize the groups with corresponding parameters.

(8)如果k大于或等于最大迭代次数K，则转向下一步，否则k＝k+l，转向(3)。(8) If k is greater than or equal to the maximum number of iterations K, turn to the next step, otherwise k=k+l, turn to (3).

(9)选择适应度最高的个体的参数作为负相关神经网络集成增量学习算法的参数。(9) Select the parameters of the individual with the highest fitness as the parameters of the negative correlation neural network ensemble incremental learning algorithm.

本文所采用的交叉策略是，后代个体的参数从父代个体的参数所确定区间内随机选择，公式如下所示：The crossover strategy adopted in this paper is that the parameters of offspring individuals are randomly selected from the interval determined by the parameters of parent individuals, and the formula is as follows:

x′_ij＝x_i+(x_j-x_i)·rand(0,1)+k·N(0,1)x′_ij ＝x_i +(x_j -x_i )·rand(0,1)+k·N(0,1)

公式中x′_ij中表示的是父代i，j的子代个体参数且j≠i，rand(0，1)表示(0，1)区间的均匀分布随机数，N(0,1)表示(0,1)正态分布随机数,k为变异强度；In the formula, x′_ij represents the individual parameters of the offspring of the parent i, j and j≠i, rand(0,1) represents a uniformly distributed random number in the (0,1) interval, and N(0,1) represents (0,1) Normal distribution random number, k is the variation intensity;

步骤4根据步骤3-4中获得的初始参数初始化一个基于负相关学习的集成神经网络分类器，利用初始阶段的样本训练此分类器。Step 4 initializes an integrated neural network classifier based on negative correlation learning according to the initial parameters obtained in steps 3-4, and uses the samples in the initial stage to train the classifier.

步骤4，增量学习阶段，当专家将新的样本数据加入样本库时，基于负相关学习的集成神经网络分类器只对新加入的样本数据进行训练并更新分类器，从而达到增量学习的目的。Step 4, incremental learning stage, when experts add new sample data to the sample library, the integrated neural network classifier based on negative correlation learning only trains the newly added sample data and updates the classifier, so as to achieve the goal of incremental learning. Purpose.

步骤5，用户上传病害图片，分类器将病害图片的诊断结果及防治措施反馈给用户,从而准确的识别诊断病害类别，达到农作物综合防治的目的。Step 5: The user uploads the disease picture, and the classifier feeds back the diagnosis result and control measures of the disease picture to the user, so as to accurately identify and diagnose the disease category and achieve the purpose of comprehensive crop control.

最后应说明的是：以上实施例仅用以说明发明的技术方案，而非对其限制；尽管参照前述实施例对发明进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离发明各实施例技术方案的精神和范围。Finally, it should be noted that: the above embodiments are only used to illustrate the technical scheme of the invention, rather than limiting it; although the invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it still can The technical solutions described in the foregoing embodiments are modified, or some of the technical features are replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the invention.

Claims (6)

1. the corps diseases recognition methodss based on incremental learning are it is characterised in that comprise the following steps：

Step 1 is collected original disease geo-radar image and is carried out Image semantic classification；

Step 1-1 collects disease original image, its species of labelling；

Step 1-2 carries out greyscale transformation to disease geo-radar image, colored disease geo-radar image is converted into gray level image, extracts each pixelR, G, B component, conversion formula is as follows：

Gray=0.299*R+0.587*G+0.114*B

Step 1-3 carries out image enhaucament to disease geo-radar image, first carries out denoising using medium filtering, then the image after denoising is enteredColumn hisgram equalizes；

Step 1-4 is split to disease geo-radar image, using Threshold segmentation, scab in disease geo-radar image and blade background is carried out pointFrom acquisition comprises only the image of scab；

Step 1-5 carries out contours extract to disease geo-radar image；

Step 1-6 carries out disease extraction to disease geo-radar image, contour images is superimposed with artwork and carries out and computing, obtains eliminating leafThe scab image of piece background；

Step 2 carries out feature extraction to the disease geo-radar image through pretreatment, special including color characteristic, textural characteristics and formLevy, the characteristic vector that the feature of three aspects extracted is classified as identification, build sample database；

Step 3 carries out EVOLUTIONARY COMPUTATION simulation incremental learning process using negatively correlated integrated neural network to Sample Storehouse disease geo-radar image,Determine the initial parameter of negative correlation learning system, and initialize an integrated nerve based on negative correlation learning with this initial parameterNetwork classifier；

Step 3-1 for disease sample training set D,

D=(x (1), d (1)) ..., (x (N), d (N)) }

X, d represent sample input respectively and export, and N is number of training；

This system of step 3-2 is formed by M somatic nervess system integrating, using the average output exporting as integrated system：

$F\left(n\right)=\frac{1}{M}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{F}_i\left(n\right)---\left(1\right)$

Wherein F_iN () is the output as neutral net i individual during input of n-th training sample, F (n) is n-th training sampleOutput as integrated system during input；

Negative correlation learning introduces a related penalty term in the error function of each individual networks so that the mistake of individual networksDifference is in negative correlation with the error of other networks, and then all-network is entered in training set D simultaneously and alternativelyRow training, this error function is defined as follows：

$E_i=\frac{1}{N}\underset{n=1}{\overset{N}{\&Sigma;}}{E}_i\left(n\right)=\frac{1}{N}\underset{n=1}{\overset{N}{\&Sigma;}}\frac{1}{2}{\left(F_i\right(n)-d(n\left)\right)}^2+\frac{1}{N}\underset{n=1}{\overset{N}{\&Sigma;}}{\mathrm{\&lambda;p}}_i\left(n\right)---\left(2\right)$

E_iN () is n-th training sample error function as individual networks i during input；In formula (2), Section 1 is individual netThe empirical risk function of network i, p in Section 2_iN () is dependency penalty；By minimizing p_iSo that each individual networksError and remaining network error be in negative correlation；By adjusting λ, 0≤λ≤1 is controlling punishment dynamics；

Dependency penalty p_iN () is defined as follows:

$p_i\left(n\right)=\left(F_i\left(n\right)-F\left(n\right)\right)\underset{j\&NotEqual;i}{\mathrm{\&Sigma;}}\left(F_j\left(n\right)-F\left(n\right)\right)---\left(3\right)$

When n-th training sample is as inputting, E_iN the local derviation of () output with regard to network i is:

$\begin{array}{c}\frac{\&part;E_i\left(n\right)}{\&part;F_i\left(n\right)}=F_i\left(n\right)-d\left(n\right)+\mathrm{\&lambda;}\frac{\&part;p_i\left(n\right)}{\&part;F_i\left(n\right)}\\ =F_i\left(n\right)-d\left(n\right)+\mathrm{\&lambda;}\underset{j\&NotEqual;i}{\mathrm{\&Sigma;}}\left(F_j\left(n\right)-F\left(n\right)\right)\\ =F_i\left(n\right)-d\left(n\right)-\mathrm{\&lambda;}\left(F_i\left(n\right)-F\left(n\right)\right)\\ =\left(1-\mathrm{\&lambda;}\right)\left(F_i\left(n\right)-d\left(n\right)\right)+\mathrm{\&lambda;}\left(F\left(n\right)-d\left(n\right)\right)\end{array}---\left(4\right)$

It is assumed herein that F (n) is with respect to F_iN () is a constant, adopt BP algorithm to update weights with ordered mode, that is, defeated for eachEnter training sample, the right value update of all individual networks is carried out by formula (4) simultaneously, all sample training one time, claimFor an epoch；

In training process, the penalty term that all of individual networks are passed through in error formula interacts；The weights of individual networks i are notF only to be made_iN () is minimum with the difference of d (n), F (n) also to be made is minimum with the difference of d (n)；

Step 3-2 simulates incremental learning process by evolutional learning, obtains the negative of one group of optimum negatively correlated integrated neural networkRelated penalty factor intensity, sub-network number, sub-network hidden node, sub-network learning coefficient, learning error and initial weight etc. are joinedNumber；

The method comprises the following steps：

(1) initial training sample set S is randomly divided into m subset { S₀,S₁,…S_m, primary iteration number of times k=0；

(2) as initial population, each individuality corresponds to a neutral net collection to initialization N number of negative correlation Artificial neural network ensembleBecome, the initial parameter of each system integrating randomly generates in span, if training batch I=0；

(3) newly add a collection of sample S_I, use S_I, 0<I<M trains each system integrating, Zhi Dao by negative correlation learning methodS_IUpper error rate is less than learning error；

(4) if I is more than sample number m in batches, next step, otherwise I=I+1 return to step (3) are turned to；

(5) test the classification accuracy rate of each negatively correlated system integrating on confirming collection, as each individual fitness；

(6) delete the poor individuality of N/2 fitness, for the higher individuality of remaining fitness, each individuality randomly choosesAnother is individual, produces an offspring by intersecting, making a variation；

(7) obtain colony of new generation, and reinitialize colony with corresponding parameter；

(8) if k is more than maximum iteration time, turn to next step, otherwise k=k+l, I=0 turn to step (3)；

(9) select the individual parameter of fitness highest as the parameter of negatively correlated Artificial neural network ensemble Incremental Learning Algorithm；

The Crossover Strategy being adopted is that the parameter of offspring individuals randomly chooses in the determined interval of the parameter of parent individuality, publicFormula is as follows：

x′_ij=x_i+(x_j-x_i)·rand(0,1)+k·N(0,1)

X ' in formula_ijMiddle expression is parent i, the offspring individual parameter of j and j ≠ i, and rand (0,1) represents (0,1) interval equalEven distribution random numbers, N (0,1) represents (0,1) normal distribution random number, and k is variation intensity；

Through certain incremental learning process, then carrying out develops selects each system integrating, the final individual body acupuncture obtainingTo have good incremental learning ability to current problem；

Step 3-3 initializes an Artificial neural network ensemble based on negative correlation learning according to the initial parameter obtaining in step 3-2Grader, using this grader of sample training of starting stage；

The step 4 incremental learning stage：When new disease sample data is added Sample Storehouse, the integrated god based on negative correlation learningOnly the new sample data adding is trained and updates grader through network classifier, thus reaching the purpose of incremental learning；

The step 5 disease recognition stage：Upload disease picture, Disease Characters are extracted by Image semantic classification, then by Disease CharactersInput the grader training to be identified.

2. the corps diseases recognition methodss based on incremental learning according to claim 1 it is characterised in that：Described step1-4 selected threshold on grey level histogram, is split, using the dual threashold based on OTSU algorithm and fundamental particle colony optimization algorithmValue carries out image segmentation：

OTSU adaptive threshold asks method to be combined with particle cluster algorithm, using OTSU algorithm as particle cluster algorithm adaptation value function,The fitness calculating each particle, compared with optimal threshold, finally obtains the dual threashold of unified algorithm optimization through successive ignitionScab in disease geo-radar image is separated with blade background by value using acquired threshold value.

3. the corps diseases recognition methodss based on incremental learning according to claim 1 it is characterised in that：Described step1-5 is detected to scab profile using Canny algorithm, smooths scab image with Gaussian filter, with single order local derviation finite differenceDivide and calculate scab image gradient amplitude and direction, non-maxima suppression is carried out to gradient magnitude, finally use dual threashold value-based algorithm to detectAnd adjoining edge.

4. the corps diseases recognition methodss based on incremental learning according to claim 1 it is characterised in that：Described stepDuring 2 color feature extracted, integrated use RGB color, HIS color space and YCbCr color space analysis scab color, carryTake RGB color primitive color R, G, B component, normalized color component r, g, b, HIS color space hue H, brightness I,Saturation S, YCbCr color space color Cb, Cr value, and first moment, second moment and third moment, totally 14 color characteristics toAmount.

5. the corps diseases recognition methodss based on incremental learning according to claim 1 it is characterised in that：Described stepDuring 2 texture feature extraction, analyze disease geo-radar image grain distribution using gray level co-occurrence matrixes algorithm and Box-counting technique, extract entropy,Energy, inertia, contrast, dependency and relevant information are estimated and 7 textural characteristics such as fractal dimension.

6. the corps diseases recognition methodss based on incremental learning according to claim 1 it is characterised in that：Described stepDuring 2 morphological feature extraction, calculate lesion area, girth, circularity, spherical property, shape using zone marker method and area tracking algorithm8 morphological feature vectors such as the state factor, dispersion index, equivalent area radius and inscribed circle radius.