基于稀疏自编码和SPSO-SVM的稻瘟病早期病害识别

Early Disease Identification of Rice Blast Based on Sparse Automatic Encoder and SPSO-SVM

为及早准确识别水稻叶部慢性型、急性型、褐点型和白点型4种类型稻瘟病,将稀疏自动编码器和交换粒子群优化支持向量机(SPSO-SVM:Switching Particle Swarm Optimization Support Vector Machine)相结合,构建了一个深度神经网络。相较于其他算法,当神经网络输入的图像数量很多时,自动编码器可提取出原图片中最具代表性的信息,缩减输入中的信息量,再将缩减过后的信息放入神经网络中学习,降低了学习难度、减少了学习时间。首先依靠稀疏自动编码器编码、解码重构输入数据,对稻瘟病叶斑进行分层特征学习,并在自动编码器上加入稀疏性条件约束,对隐含层进行压缩,进而学习到更高层的隐含特征。其次应用交换粒子群优化的支持向量机对水稻稻瘟病类型识别。实验采用公开的Kaggle水稻病害图像数据库及实际采集的水稻稻瘟病图像作为数据集,每类选取350幅图像组成样本,并将每幅图像归一化为4 096维向量。从样本集中随机选取80%作为训练集,剩余20%作为测试集。通过10重交叉验证,测试集平均识别准确率达95.7%。实验结果表明,该方法能有效地从病斑特征中识别出水稻叶部稻瘟病早期病害,对水稻稻瘟病的早期预防有重要意义。

In order to identify chronic, acute, brown dot and white dot, four types of rice blast diseases early and accurately, a deep neural network is constructed by combining sparse automatic encoder and SPSO-SVM(Switching Particle Swarm Optimization Support Vector Machine). Compared with other algorithms, the neural network needs to input a large number of images, the autoencoder can extract the most representative information in the original image, reduce the amount of information in the input, and then put the reduced information into the neural network to learn, greatly reducing the difficulty and time of learning. Firstly, the input data is encoded, decoded and reconstructed by sparse automatic encoder to learn the hierarchical features of rice blast leaf spots, and the sparse condition constraint is added to the automatic encoder to compress the hidden layer, so as to learn the higher-level hidden features. Secondly, the support vector machine optimized by switching particle swarm optimization is used to identify the types of rice blast. The open Kaggle rice disease image database and the actually collected rice blast image are used as the data set. 350 images of each type were selected to form samples, and each image is normalized to 4 096 dimensional vector. 80% of the samples are randomly selected as the training set and the remaining 20% are used as the test set. Through 10 cross validation, the average recognition accuracy of the test set is 95.7%. The experimental results show that the proposed method can effectively identify the early disease of rice leaf blast from the features of disease spots, which is of great significance for the early prevention of rice blast.