基于椭圆型度量学习的小麦叶部病害识别

特征提取和相似性度量是基于图像处理的农作物病虫害识别方法中的两大关键问题。以感染小麦白粉病的叶片为研究对象，提出了一种基于椭圆型度量学习的小麦叶部病害严重度识别算法。该算法首先给出了一种滑窗最大值（Moving window maximum, MWM）特征提取方法，对分割后的病斑图像采用滑窗法提取HSV颜色特征和LBP纹理特征，在同一水平条滑窗上取每一维特征的最大值作为这一水平条的特征，这种MWM特征表示方法能有效减弱小麦叶片弯曲、倾斜、拍摄角度不同等对识别率的影响；然后，引入对样本数据具有更好区分性的椭圆型度量，根据样本的类内与类间高斯分布的对数似然比定义椭圆型度量矩阵，为了保持最大化的分类信息，将特征子空间学习和椭圆型度量学习同时进行；最后，利用得到的椭圆型度量计算特征向量之间的距离实现不同严重度病害的识别。对比实验结果表明,本文算法使得小麦白粉病严重度的识别正确率达到了100%，优于SVM方法的88.33%、BP神经网络方法的90%。

Recognition of Wheat Leaf Diseases Based on Elliptic Metric Learning

Feature extraction and similarity measurement are two key problems of crop pest recognition based on image processing. The leaves of wheat powdery mildew were treated as the research objects, and an algorithm of wheat leaf disease severity recognition based on elliptical metric learning was proposed. Firstly, a method of moving window maximum (MWM) feature extraction was presented in the algorithm. The HSV color features and LBP texture features were extracted by using the sliding window method from the segmented lesion images. The maximum value of each dimension feature on the same horizontal sliding window was taken as the feature of this horizontal bar. The MWM feature representation method can effectively reduce the influence of curvature, tilt and different shooting angles of wheat leaves on the recognition rate. Then, an elliptical metric with better distinguishability for sample data was introduced, and the elliptic metric matrix was defined based on the log-likelihood ratio of Gaussian distributions on the intrapersonal sample and the extrapersonal sample. In order to maintain the maximal classification information, the feature subspace learning and elliptic metric learning were performed simultaneously. Finally, to recognize the severity of diseases, the elliptic metric was used to calculate the distance between the eigenvectors. The results of comparison experiments showed that the recognition rate of wheat powdery mildew severity was 100%, which was better than 88.33% for SVM method and 90% for BP neural network method. The research result can provide valuable help for the intelligent recognition of crop disease severity.