论文部分内容阅读
目的计算机智能分析用户的饮食是一项有意义的研究课题。传统的分析方法侧重于分析食物类型,可是中国是个美食之国,食物类型在各个区域间表现出极大的多样性,造成很难实现通用的食物自动分类方法。为此尝试针对食物的原材料,即食材进行自动分析收集并建立了一个真实环境下的食材图像数据库(FOOD-SCUT),此数据库包括目前中国市面上常见的70种食材类别,共8 015幅图片。方法基于此数据库,本文尝试性地利用不同的传统特征和分类方法,对此食材图像数据库进行自动分类,以此来分析对比各种特征和分类方法的性能。对比性实验中所选用的特征包括:SIFT特征、颜色直方图特征、梯度直方图、SURF特征、LBP特征和Gabor特征等。除颜色直方图外其他特征都会利用词袋模型进行特征编码,而所选用分类方法包括:支持向量机(SVM)、朴素贝叶斯、随机森林、K-最近邻(KNN)算法。另外本文还尝试采用最近流行的深度神经网络方法对数据库进行特征学习和分类。结果通过实验验证基于各种传统特征分类方法的实验性能,其中各种特征包括单特征和多特征组合两种方式,通过不断调整不同特征组合和分类识别算法及其参数,得到基于传统特征分类方法的最好分类性能。同时通过实验验证深度卷积神经网络模型的实验性能,深度卷积神经网络模型使用直接训练和预训练两种不同训练模式,并调整不同的网络层数和权重初始化方法后获得最好的分类识别性能。本食材数据库基于传统特征分类方法的最好分类准确率为88.98%,而基于深度神经网络分类方法上可以获得最佳的实验性能,即95.7%,这个准确率比基于传统特征分类方法高出6.72%。结论数据库的统计结果表明此食材图像数据库类内数据具有极大的差异,可以作为分析食材的一个基础数据库。此数据库具有极高的应用价值,可以为后续各种基于食材分析应用提供相关分析数据,并且本文实验分析结果,对于后续用户开发相关的各种相关应用中,提供了模型和参数选择的建议,节省了用户选择模型和调参的实验过程。
The purpose of computer intelligent analysis of the user’s diet is a meaningful research topic. While traditional analytical methods focus on the analysis of food types, China is a food country and food types show great diversity across regions, making it difficult to achieve a universal method of automatic food classification. In order to do this, automatic analysis of raw materials, ie ingredients, of food was conducted. A real-world Food Image Database (FOOD-SCUT) was compiled and established. This database includes 70 kinds of food ingredients currently on the market in China, for a total of 8 015 pictures . Methods Based on this database, this paper tentatively uses different traditional features and classification methods to automatically classify the foodstuff image database to analyze and compare the performance of various features and classification methods. Selected features in the comparative experiment include: SIFT feature, color histogram feature, gradient histogram, SURF feature, LBP feature and Gabor feature. All the features except the color histogram will be encoded using the bag-of-characters model. The selected classification methods include Support Vector Machine (SVM), Naive Bayes, Random Forest, and K-Nearest Neighbor (KNN) algorithms. In addition, this paper also attempts to use the recent popular method of deep neural network to feature learning and classification of the database. The results verify the experimental performance based on various traditional feature classification methods. Among them, the features include single feature and multi-feature combination. Through continuously adjusting different feature combination and classification identification algorithms and their parameters, the traditional feature classification method The best classification performance. At the same time, the experimental performance of deep convolutional neural network model is validated by experiments. Deep convolutional neural network model uses two different training modes, direct training and pre-training, and obtains the best classification and recognition after adjusting different network layers and weight initialization methods performance. The best classification accuracy of this food database based on traditional feature classification is 88.98%, while the best experimental performance based on deep neural network classification is 95.7%, which is 6.72% higher than traditional feature classification %. Conclusion The statistical results of the database show that the data in this foodstuff image database have great differences and can be used as a basic database for analyzing foodstuffs. This database has very high application value and can provide relevant analysis data for subsequent various food-based analysis applications. And the experimental analysis results in this paper provide suggestions for model and parameter selection for various related applications related to subsequent user development, Save the user to choose the model and parameters of the experimental process.