基于光谱技术和多分类器融合的异物蛋检测

为了提高鸡蛋中的血斑和肉斑的检测准确率,给消费者提供高品质的鸡蛋,该文利用微型光纤光谱仪采集鸡蛋的透射光谱,在单分类器的基础上,通过多分类器的融合对异物蛋进行检测。首先根据差异性度量选取朴素贝叶斯,Ada Boost和SVM分类器作为单分类器,然后通过特征级融合选取了5个基分类器。最后,5个基分类器以加权投票机制进行决策级融合。多分类器融合对正常蛋和异物蛋的检测准确率分别为92.86%和91.07%。试验结果表明,利用多分类器融合所建立的模型优于单一分类器的模型,提高了对异物蛋的检测准确率。

Abnormal eggs detection based on spectroscopy technology and multiple classifier fusion

Abstract: The aim of the research was to improve detection accuracy of the blood spots and meat spots in eggs, which can provide consumers with high-quality eggs. Spectroscopy technology and multiple classifier fusion for abnormal egg detection were investigated. Micro fiber spectrometer (Ocean Optics company, USB2000+) was used to collect the transmittance spectroscopy of both normal and abnormal eggs, which were from Hubei Shendan Healthy Food Co., Ltd. After outliers detection and elimination, there were 336 eggs in all, which were randomly assigned to training set and test set; among the 336 eggs, 224 (about two-thirds of the total) were assigned to the training set, and the remaining 112 (about one-third of the total) were assigned to the test set. Before multiple classifier fusion, all data collected from micro fiber spectrometer was preprocessed including the methods of SNV (standard normal variate), smoothing and MSC (multiplicative scatter correction). usion of multiple classifiers was used to detect the foreign bodies of eggs on the basis of single-classifier. Firstly, five single-classifiers which was inclusive of Naive Bayes classifier, Mahalanobis distance classifier, PLS-DA (partial least squares-discriminate analysis) classifier, AdaBoost (adaptive boosting) classifier and SVM (support vector machine) classifier were all trained, and five groups of classification results were attained. In order to choose suitable one from these five single-classifiers, according to diversity measure, output disagreement measure and error agreement measure were introduced and used, and Naive Bayes, AdaBoost and SVM classifiers were selected as the single-classifiers; then, on the basis of these three selected single-classifiers, through feature level fusion, 21 base-classifiers were obtained. In order to get the final base-classifiers which were used to accomplish the multiple classifier fusion, in a similar way, through output disagreement measure and error agreement measure again, 5 base-classifiers were chosen from these 21 base-classifiers. Finally, 5 basic-classifiers were fused by weight vote strategy on the decision level. The weight vote strategy was that each base-classifier was allocated a weight value according to its accuracy rate, and the higher accuracy rate a base-classifier had, the larger weight value it would be allocated, because it was more trusted. Detection accuracy rate of ensemble classifier, which was formed after multiple classifier fusion, was 92.86% and 91.07%, respectively for normal eggs and abnormal eggs. As a contrast, among all the single-classifiers and base-classifiers, the highest detection accuracy of normal eggs in the test set was 91.07%, which came from AdaBoost (500-600 nm), and the highest detection accuracy of abnormal eggs in the test set was 89.29%, which came from SVM (550-600 nm). The experiment results showed that the model established by multiple classifier fusion could take full advantage of the information which came from each single-classifier or base-classifier, and in the aspect of the detection accuracy of either normal eggs or abnormal eggs, the model established by multiple classifier fusion was indeed superior to the model established by each single-classifier or base-classifier. Even though the detection accuracy was enhanced by a small margin, considering that a large number of either normal eggs or abnormal eggs were being produced and being detected in lots of companies that were involved in eggs, in this meaning, the slight promotion of detection accuracy was of great significance.