面向神经网络结构搜索的植物叶片病害增强识别方法

Enhanced Identification Method of Plant Leaf Disease Based on Neural Network Structure Search

针对植物病害识别模型结构复杂且依赖于人为设计网络结构等问题,通过神经网络结构搜索(NAS),提出一种基于队列分块的神经网络结构搜索方法(NNSS),可实现超轻量级高精度植物叶片图像识别模型的自动构建。首先将12种在经济和环境下有益的植物共计22类植物叶片图像作为训练样本,利用模糊c均值聚类(FCM)算法分割植物叶片的感染点,以获得叶片受关注的区域信息;通过图像像素的灰度空间相关性,采用快速灰度共生矩阵(FGLCM)算法提取6类受关注区域的纹理特征信息,获得的特征向量运用主成分变换选择重要特征;提出队列分块的局部搜索空间构造方法,将特征信息通过自动构建的模型进行分类。结果表明,NNSS方法取得了98.33%的准确率,特异性和灵敏性表现最优。相比于AlexNet、GoogLeNet、InceptionV3和VGGNet-16模型,改进VGG-INCEP16模型的性能得到进一步提升,但仍低于NNSS方法,这是由于该方法能结合数据集搜索合适的网络结构,对比次优VGG-INCEP16模型准确率至少提高了2.1%。研究结果显示,NNSS方法能够实现准确识别植物病害,对于神经网络模型结构自动搜索的未来具有较高的实际应用价值。

Aiming to solve the problems of the complex structure of plant disease recognition models and relying on human-designed network structure,a neural network structure search method based on queueing chunking an NNSS was proposed by neural network architecture search(NAS),which could realize the automatic construction of ultra-lightweight and high-precision plant leaf image recognition models.The method started with a total of 22 classes of plants leaf images of 12 economically and environmentally beneficial plants as training samples and used the fuzzy c-mean clustering(FCM)algorithm to segment the infected points of plant leaves to obtain information on the regions of interest of the leaves.Texture feature information of six classes of regions of interest by the gray-scale spatial correlation of image pixels was extracted by using the fast gray-scale co-generation matrix(FGLCM)algorithm and the important features of the obtained feature vectors were selected by using the principal component transform.A local search space construction method for queue chunking was proposed to classify feature information by an automatically constructed model.The experimental results showed that the NNSS method achieved 98.33%accuracy with the best performance in specificity and sensitivity.Compared with AlexNet,GoogLeNet,InceptionV3,and VGGNet-16 models,the performance of the improved VGG-INCEP16 model was further improved but still lower than the NNSS method due to its ability to search for a suitable network structure in combination with the dataset,which improved the accuracy over the suboptimal VGG-INCEP16 model by at least 2.1%.The NNSS method can accurately identify plant diseases and has a high practical application value for the future automatic search of neural network model structures.