基于卷积神经网络深度特征融合的番茄叶片病害检测

Tomato leaf disease detection based on deep feature fusion of convolutional neural network

为提高对叶片病害的识别性能和速度,针对传统手工设计特征识别能力有限的问题,提出了一种基于卷积神经网络(convolutional neural networks,CNN)深度特征融合的植物叶片病害检测方法。首先自动提取不同深度神经网络模型的深度特征,利用典型相关分析(canonical correlation analysis,CCA)进行特征融合,增加特征空间的丰富性和鲁棒性;然后与线性判别分析(linear discriminant analysis,LDA)降维算法相结合,利用LDA最大化类间距离,弥补CCA算法的弱点;最终得到精练良好的植物叶片特征表示。对最终得到的特征进行支持向量机(support vector machine,SVM)分类,在10类番茄叶片数据集上的分类准确率达98.7%,识别速度达30张/s。相比利用GoogleNet深度学习模型和深度特征级联融合分类,准确率分别提高了4%与1%;与使用CCA融合相比分类准确率下降了0.1%,但其识别速度远远高于CCA融合。对比实验表明,所提算法在去除特征冗余的同时较好地保留了相关特征信息,能快速而准确地对番茄叶片进行检测。

In order to improve the recognition performance and speed of leaf diseases,aiming at the problem of limited recognition ability of traditional manual design features,a plant leaf disease detection method based on convolutional neural networks(CNN)deep feature fusion was proposed.Firstly,deep features were automatically extracted by different neural networks,and canonical correlation analysis(CCA)algorithm was used to fuse these deep features,which increased the richness and robustness of the feature space.Secondly,linear discriminat analysis(LDA)algorithm was used to reduce the dimension of the fused features.The idea of LDA maximizing the distance between classes could make up for the weakness of the CCA algorithm,and we could finally get well-represented plant leaf feature.The final fused features were used for support vector machine(SVM)classification.The classification accuracy rate on the 10 types of tomato leaf dataset reached 98.7%,and the recognition speed reached 30 sheets/s.Compared with the classification results of GoogleNet model and deep feature cascade fusion,the accuracy rate was improved by 4%and 1%respectively.Compared with the results of CCA fusion,the classification accuracy is reduced by 0.1%,but the recognition speed is much faster than CCA fusion.In conclusion,the algorithm in this paper was used to remove feature redundancy and maximize the retention of relevant feature information,which can quickly and accurately detect tomato leaves.