基于稀疏表示的枸杞分类研究

Wolfberry classification based on sparse representation

为实现利用计算机视觉技术提升枸杞品种分类效果的目的,研究基于稀疏表示(SR)的枸杞品种分类方法。首先获取枸杞的图像,并提取枸杞图像的颜色和形态特征参数,得到枸杞训练样本的数据词典矩阵。在此基础上,利用稀疏表示方法对枸杞测试样本进行分类。基于稀疏表示分类方法的第一步是利用数据词典矩阵对测试样本进行稀疏性表示,得到测试样本的稀疏表示系数;第二步,利用测试样本的稀疏表示系数,对测试样本进行重构;第三步,计算重构样本与测试样本之间的残差,通过比较残差的大小来确定测试样本的类别。本文将稀疏表示分类方法与深度神经网络(DNN)、最小二乘支持向量机(LSSVM)、BP网络和支持向量机(SVM)等方法的识别结果做了对比和分析。试验结果表明,稀疏表示分类方法对于3个枸杞品种的综合分类准确率为98.33%,获得了最好的分类效果。

In order to achieve better effect of wolfberry classification based on computer vision technology, a classification method based on sparse representation is proposed for discriminating the varieties of wolfberry precisely. Firstly, we obtain the wolfberry images, extract the color and shape feature parameters of the wolfberry images, and then, get the matrix of data dictionary. Secondly, sparse representation method is applied to accomplish the wolfherry classification. To be more specific, the first step of this classification process is to represent the test samples by the matrix of data dictionary and to obtain the sparse representation coefficients; By making use of the sparse representation coefficients, the second step is to reconstruct the test samples; In the third step, residuals between the reconstructed samples and the test samples are calculated and varieties that test samples should belong to can be confirmed. At last of this paper, the sparse representation method is compared with some other methods, such as the deep neural network （DNN）, least squares support vector machine （LSSVM）, BP network and support vector machine （SVM）. Experimental results demonstrate that the overall classification accuracy of the sparse representation method is 98.33%, which has the best classification effect among five methods.