Application Of Machine Learning In Detection Of Meat Freshness Based On Hyperspectral Imaging

Meat is a major source of nutrition and trace elements for human, and its quality directly affects the quality of life, nutrient levels and food security. With the development of economic and the improvement of people’s living standards, there are higher requirements for the quantity and quality of meat products. In recent years, the hyperspectral imaging which integrats the advantages of machine vision and near infrared spectral analysis technology, which is the reason to introduce the hyperspectral imaging into the nonedestructive detection of meat freshness. However, how to build a safety, high efficient and high precision model for the detection of meat freshness based on hyperspectral imaging is still a problem needed to be solved. In this paper, the main work had been done in research of building model for the detection of meat freshness based on hyperspectral imaging, and the related research achievements in the field of machine learning were introduced into the nondestructive testing of meat freshness based on hyperspectral imaging to realize the high precision, fast and nondestructive testing of meat freshnessfor the freshness of meat. The main research works of this paper are as follows:1. The feature fusion was applied to build the detection model of meat freshness based on hyperspectral image technology. This method wew used to extract entropy features of image and the mean feature of spectrumto make up for the deficiency of the single feature.However, the redundancy problem could be caused by much variable information. The SPA combined with PCA obtained the main component of information and reduced the redundancy of information. The results showed that the fusion feature of entropy feature of image and mean feature of spectrum could obtain better model than single feature; in addition, compared with the single dimension reduction of SPA and PCA, the dimension reduction method of SPA combined with PCA could obviously improve the accuracy and stability of model.2. The selection of training samples is an important factor to the accuracy of the detection for meat freshness based on hyperspectral image. In this paper, the sampling strategy based on Active Learning algorithm was used to select training samples. This method used the weighted risk of prediction function as random variables.The training samples were selected by Active Learning algorithm to minimize the risk of expected error of prediction function and labeled by artificial markers. The research results showed that training samples chosen by Active Learning sampling strategy could improve the accuracy of model, save the time and the cost of modeling under the condition of small training samples.3. The model has a decisive influence on prediction results. All the current methods of traditional modeling were still to be used a single model. However, the single model is possible to make parameter be caught in local optimum, which could limit the performance of the model. In addition, if the sample set includes the abnormal samples or different kinds of samples, the single model obviously unable to meet the requirements of stability and accuracy. Therefore, this paper presented a method of Multi-Model Coordinated Recursive. Firstly, the different LSSVM models were built by training samples and the T test was used to judge performance of model. Then the standard values of samples were obtained by weighted fusion for results of multiple model and the samples were joined in the training sample set to train new model, which was used to predict the standard values for next sample. In this circle until all the test samples had been predicted. Research results showed that, compared with traditional modeling methods, this method could availably improve the accuracy and stability of model and the effects are obviously in the case of a small number of training samples.