计算机工程与应用 ›› 2017, Vol. 53 ›› Issue (22): 180-184.DOI: 10.3778/j.issn.1002-8331.1605-0073

• 图形图像处理 • 上一篇    下一篇

基于二维子空间的苹果病害识别方法

师  韵,黄文准,张善文   

  1. 西京学院 电子信息工程系,西安 710123
  • 出版日期:2017-11-15 发布日期:2017-11-29

Apple disease recognition based on two-dimensionality subspace learning

SHI Yun, HUANG Wenzhun, ZHANG Shanwen   

  1. Department of Engineering and Technology, Xijing University, Xi’an 710123, China
  • Online:2017-11-15 Published:2017-11-29

摘要: 如何准确、实时得到苹果病害信息是苹果病害管理的一个重要研究内容。根据苹果叶片症状准确、快速地诊断苹果病害是预防和控制苹果病害的基础。由于苹果同类病害叶片及其病斑图像的形状、颜色和纹理之间的差异很大,使得很多经典的模式识别方法不能有效地应用于苹果叶部病害识别。为此,提出了一种基于二维子空间学习维数约简(2DSLDR)的苹果病害识别方法。该方法将高维空间的观测样本点映射到低维子空间,使得类内样本点更加紧凑,而类间样本点更加分离,从而得到最佳的分类特征。该方法直接作用于叶片图像,不需要计算逆矩阵,从而克服了经典植物病害识别方法中特征提取与选择的难题,避免了经典子空间判别分析中的小样本问题,提高了识别效果。采用该方法对三种常见苹果叶部病害进行识别实验,并与其他苹果病害识别和监督流形学习方法进行比较。实验结果表明,2DSLDR对苹果叶部病害识别是有效可行的,识别精度高达90%以上。

关键词: 苹果叶部病害识别, 苹果病害叶片图像, 最近邻分类器, 二维子空间学习维数约简

Abstract: The accuracy and real-time recognition of apple diseases is an important research area. By the symptoms of the apple leaves, correct and fast identification of apple leaf diseases is the basis of preventing and controlling the apple leaf diseases. Because the apple leaves are different from each other in shape, color and texture, many classical pattern recognition methods are not effectively applied to recognizing the apple diseases. In this paper, a novel two-Dimensionality Subspace Learning Dimensionality Reduction method(2DSLDR) is proposed for the apple leaf disease recognition. By 2DSLDR, the multi-class sample points in high-dimensional space are projected to low-dimensional subspace and the optimal low-dimensional features from the point of view of classification are extracted. After being projected into a low-dimensional subspace, the sample points in the same class are close to each other, but the different-class sample points are far away from each other. 2DSLDR directly deals with the leaf images and does not calculate the inverse matrix, so the hard-problem in the classical plant recognition methods to extract and select the feature is overcome, and the small sample size problem occurring in traditional subspace discriminant analysis methods is naturedly avoided, and much computational time is saved by using 2DSLDR for dimensionality reduction. The recognition experiments are performed on a real apple disease leaf image database and compared with the traditional disease recognition methods and the supervised subspace learning algorithms in recognition performance. The experiment results show that the proposed method is effective and feasible for apple leaf disease recognition. The mean correct recognition ratio is above 90%.

Key words: apple leaf disease recognition, apple disease leaf image, nearest neighborhood classifier, Two-dimensionality Subspace Learning Dimensionality Reduction(2DSLDR)