基于多特征融合的面部表情识别方法研究

发布时间：2018-03-19 09:12

本文选题：表情识别　切入点：特征融合　出处：《西安科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：表情识别的研究由来已久,产生了众多特征提取和分类识别算法,且性能良好,但识别率仍然制约着表情识别的实际应用。针对面部表情识别仍然存在的问题,对表情识别中的特征提取和表情分类识别进行了研究,主要内容及创新点如下:(1)提出了一种多特征加权融合(Multi-feature Weight Fusion,MFWF)的特征提取方法。单一特征信息不能准确描述图像,影响图像的识别率的问题。为了用特征信息有效表征图像,提高图像识别率,根据信息融合的理念,将特征加权融合方法用于表情特征的提取。首先,确定人脸的局部感兴趣区域;其次,采用多种特征提取方法,获得人脸的多个特征信息;最后,根据不同面部区域对表情的表现力不同,对不同特征进行加权处理,并采用串联融合的方法,将所提取到的特征连接成一行,作为图像的最终特征信息。为了验证MFWF方法的有效性,采用JAFFE数据库作为数据源,将融合特征和非融合特征的分类识别结果作对比。实验结果表明:融合特征能够准确表征图像信息,普遍提高了表情的识别率,特别是在LDA、DT和DDA三个分类器中的效果最明显。(2)提出了一种二次优化选择性(Quadratic Optimization Choice,QOC)集成分类模型,并用于面部表情识别方法。针对单一分类器的缺陷,引入集成学习的思想,生成集成分类器模型用于面部表情识别。首先,对于众多的基分类器,依据各自的分类识别性能进行从大到小的排序,并按照相应的停止准则确定待集成的基分类器;其次,对于已确定的待集成分类器,按照集成规则生成集成模型簇;最后,对集成模型簇的性能进行测试,并从大到小进行集成模型的排序,选择排序首位的集成模型作为最终的集成分类器模型。为了验证QOC方法的性能,以JAFFE数据库作为测试数据集,并采用最大值、最小值和均值集成规则的集成作为对比模型。实验结果表明,相比较于基分类器,集成分类器的分类识别性能显著提高,且SRM集成模型将悲伤表情的平均识别率从78.89%提高到96%;相对于非选择集成模型,QOC方法的LFER集成规则好于其他集成规则;相对于选择集成模型,QOC方法中的选择策略可以有效提高集成模型的识别率。
[Abstract]:The research of facial expression recognition has a long history, and has produced many feature extraction and classification algorithms, and the performance is good, but the recognition rate still restricts the practical application of facial expression recognition. In this paper, feature extraction and facial expression classification recognition in facial expression recognition are studied. The main contents and innovations are as follows: 1) A multi-feature weighted fusion multi-feature Weight fusion (MFWF) method is proposed. In order to use feature information to represent the image effectively and improve the recognition rate, according to the idea of information fusion, the method of feature weighted fusion is used to extract facial features. Determine the region of local interest of the face; secondly, using a variety of feature extraction methods to obtain multiple features of the face information; finally, according to the expression of different facial regions to the expression of different features, weighted processing, In order to verify the validity of the MFWF method, the JAFFE database is used as the data source. The experimental results show that the fusion features can accurately represent the image information and generally improve the recognition rate of facial expressions. Especially in the three classifiers of LDA-DT and DDA, this paper presents a quadratic optimized selective Optimization ChoiceQOC-based ensemble classification model, which is used in facial expression recognition. In view of the shortcomings of the single classifier, the idea of ensemble learning is introduced. An integrated classifier model is generated for facial expression recognition. Firstly, according to their classification performance, the classifier is sorted from large to small, and the basis classifier to be integrated is determined according to the corresponding stopping criteria. Finally, the performance of the ensemble model cluster is tested according to the integration rules, and the integration model is sorted from large to small. In order to verify the performance of the QOC method, the JAFFE database is used as the test data set, and the maximum value is adopted. The experimental results show that compared with the base classifier, the ensemble recognition performance of the ensemble classifier is significantly improved. The average recognition rate of sad expression is increased from 78.89% to 96 by the SRM integration model, and the LFER integration rule is better than other integration rules compared with the non-selective integration model. Compared with the selection strategy in the QOC method, the recognition rate of the integrated model can be improved effectively.
【学位授予单位】：西安科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】