基于多视图特征融合的植物病害识别算法

Recognition algorithm of plant disease based on multi-view feature fusion

近年来,基于传统机器学习和深度学习的叶片病害识别算法取得显著进展。然而现有病害识别模型大多采用单一类型特征,即对叶片病害图像提取颜色纹理等传统特征或采用深度学习自动学习特征。一方面采用深度学习自动学习特征需要大量样本,计算开销较大;另一方面传统特征往往应用于小规模病害数据集。因此提出基于多视图特征融合的病害识别算法,首先提取病害叶片图像的Gist特征以及基于深度学习预训练模型VGG16的深度特征,通过深度典型相关分析(DCCA)发掘传统特征与深度特征的相关性,获得更加鲁棒的特征子空间,从而提高识别效果。在Plant Village上的试验结果表明,采用DCCA融合传统特征和深度特征的识别方法比单类型特征识别法的识别精度要高,其平均识别精度可达88.45%。

Plant leaf diseases are a serious problem in agricultural production.In recent years,significant progress has been made in plant disease identification based on traditional machine learning and deep learning algorithms.However,most of the existing disease recognition algorithms use either the traditional features such as color texture extraction of leaf disease images or deep features based on the pretrained deep convolutional neural networks.On the one hand,the use of deep features requires a large number of samples with expensive computational cost.On the other hand,traditional features are often applied to small-scale disease data sets.In this paper,it was proposed a multi-view feature fusion approach for plant disease recognition.First,the Gist features of the plant disease image and the deep features based on VGG16 model were extracted.Through deep canonical correlation analysis(DCCA),the correlation between Gist feature vectors and deep feature vectors was explored,and a more robust feature subspace was obtained,thereby improving the recognition effect.The experimental results on plant village database showed that the accuracy of our proposed approach could reach 88.45%,which was higher than that of single-type feature recognition algorithms.