跨媒体旅游大数据的语义学习与内容识别的研究

发布时间：2018-04-20 11:09

本文选题：跨媒体 + 语义建模　；参考：《北京邮电大学》2016年硕士论文

【摘要】：目前,互联网的飞速发展给智慧旅游带来了机遇和挑战,机遇是人们在旅游过程中生产了丰富的跨媒体数据,挑战是不同模态跨媒体数据之间的“语义鸿沟”为智慧旅游的发展造成了很大的障碍。本文针对三个问题进行了研究:跨媒体旅游大数据的语义分析和建模、基于旅游领域本体知识库推理的图像语义内容自动标注以及融合GIST特征和人群微观行为特征的拥挤旅游场景内容识别。论文完成的主要工作如下:(1)提出了一种基于PLSA主题模型的对称的建模方法,建立了不同模态跨媒体数据之间的潜在语义关联模型,克服了传统方法只能表现不同模态数据之间显式关系的缺点。通过本文的建模方法,建立了一个包含文本词和视觉词之间映射关系的数据模型。(2)提出了一种基于旅游领域本体知识库推理机制的图像语义自动标注算法。在已经建立的跨媒体数据语义模型的基础上,采用融合主题模型的图像标注算法对图像内容进行初步标注。同时,结合旅游领域本体知识库进行进一步的推理,从而更精确地识别图像中的内容,得到了更加具体的描述图像内容的词,提高了标注效果。通过旅游领域本体知识库的使用,使标注的结果与具体的景点关联。与传统的非对称算法相比,融合语义主题的对称图像标注算法的图像标注正确率得到了提高。(3)提出了一种融合GIST特征和微观行为特征的拥挤场景识别方法。对待识别的视频进行镜头分割,对每一个镜头进行背景提取,采用基十GIST特征的场景识别算法对场景进行初步判断,得到初步的识别结果后,根据人群移动的规律提取人群的微观行为特征,根据不同场景中人群移动规律的差异,进行再次场景判断,从而进一步提高场景识别的精确程度。该方法实现了针对拥挤旅游场景的有效的场景识别。与传统的基于GIST特征的场景识别算法相比,融入人群的微观行为特征以后,实验结果的准确率和召回率得到了进一步提高。(4)设计和实现了跨媒体旅游大数据语义学习和内容识别系统,包括跨媒体旅游大数据的语义分析和建模、基于旅游领域本体知识库推理的图像标注、融合GIST特征和微观行为特征的拥挤场景识别三个模块。通过该系统对上述每一部分的实验结果进行了验证。实验表明上述算法在旅游跨媒体数据语义学习和内容识别方面具有较好的效果。
[Abstract]:At present, the rapid development of the Internet has brought opportunities and challenges to intelligent tourism. The opportunity is that people produce rich cross-media data in the process of tourism. The challenge is that the semantic gap between different modes and media data causes a great obstacle to the development of intelligent tourism. This paper focuses on three problems: the semantic analysis and modeling of cross-media tourism big data. Image semantic content automatic tagging based on tourism domain ontology knowledge base reasoning and content recognition of crowded tourism scene based on GIST feature and crowd micro behavior feature. The main work of this paper is as follows: (1) A symmetric modeling method based on PLSA topic model is proposed, and the latent semantic association model between different modes of cross-media data is established. It overcomes the shortcoming that the traditional method can only express the explicit relationship between different modal data. Through the modeling method in this paper, a data model containing the mapping relationship between text words and visual words is established. (2) an image semantic automatic annotation algorithm based on the reasoning mechanism of ontology knowledge base in tourism domain is proposed. Based on the established cross-media data semantic model, the image tagging algorithm based on the fusion topic model is used to label the image content. At the same time, combined with the tourism domain ontology knowledge base to further reasoning, thus more accurate recognition of the image content, more specific words to describe the image content, improve the tagging effect. Through the use of ontology knowledge base in tourism field, the results of labeling are associated with specific scenic spots. Compared with the traditional asymmetric algorithm, the accuracy of symmetric image tagging algorithm based on semantic topic is improved. (3) A congestion scene recognition method based on GIST feature and micro behavior feature is proposed. The scene recognition algorithm based on the base ten GIST feature is used to judge the scene, and the initial recognition result is obtained after the scene is segmented by shot segmentation, each shot is extracted from each shot, and the scene recognition algorithm based on the base 10 GIST feature is used to judge the scene. According to the law of crowd movement, the micro-behavior characteristics of the crowd are extracted, and the accuracy of scene recognition is further improved by judging the scene again according to the difference of the law of crowd movement in different scenes. This method realizes the effective scene recognition for the crowded tourist scene. Compared with the traditional scene recognition algorithm based on GIST feature, after the microcosmic behavior of the crowd is merged, The accuracy and recall rate of the experimental results are further improved. (4) A cross-media tourism big data semantic learning and content recognition system is designed and implemented, including semantic analysis and modeling of cross-media tourism big data. There are three modules: image annotation based on ontology knowledge base reasoning in tourism domain, and congestion scene recognition based on GIST feature and micro behavior feature. The experimental results of each part are verified by the system. Experiments show that the algorithm is effective in semantic learning and content recognition of travel data across media.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】