玉米品种识别多算法模型比较研究

A Comparative Study of Multi-algorithm for Maize Varieties Identification

为了比较玉米品种图像识别中各种神经网络识别模型的性能,搭建了一套基于统计特征提取和模式识别分类算法的玉米品种识别系统。采用扫描仪获得了11个玉米品种每个品种50粒子粒图像,基于图像的统计特征,分别研究了7种人工神经网络(ANN)模型(BP、rbf、grnn、pnn、compet、sofm、ELM)的识别性能,进一步考察了极限学习机(ELM)、支持向量机(SVM)模式分类过程性能。结果表明,在同样的情况下SVM模型较ANN模型的特征识别率高,另外神经网络模型grnn和ELM识别效果较好,其他识别模型性能较差。对11个玉米品种种子的最高检出率为91.73%,另外,所采用的特征降维方法、特征维数、初始权值的随机性选择等因素都会影响模型的识别效果。这对玉米种子纯度和品种真实性检验中人工神经网络模型的构建具有指导意义。

In order to compare the performance of various neural network identification model for maize varieties image recognition, a set of maize variety identification system classification algorithm based on statistical feature extraction and pattern recognition was constructed. Each variety 50 particle images of 11 varieties were obtained by scanner, and based on the statistical features of these images, the recognition performances of seven kinds of artificial neural network（ANN） models（BP,rbf, grnn, pnn, compet, sofm, ELM） were analyzed, then the performance of extreme learning machine（ELM） and support vector machine（SVM） pattern classification process were further analyzed. The test results showed that, in the same circumstances, the feature recognition rate of SVM model was higher than that of ANN model. In addition, the recognition effect of grnn and ELM were better than others neural network model. For the 11 maize varieties, the highest detection rate was91.73%. In addition, feature reduction method, feature dimension, initial weights of random selection and other factors will affect the model identification effect. The method and conclusion in this article has the guiding sense to the construction of artificial neural network model for purity and authenticity for testing varieties of maize seeds.