CASIA OpenIR  > 毕业生  > 博士学位论文
Alternative TitleNatural Scene Semantic Recognition and Annotation for Image Retrieval
Thesis Advisor王春恒
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline模式识别与智能系统
Keyword自然场景图像语义分类 视觉目标类识别 图像多标签标注 基于内容的图像检索 图像理解 Natural Scene Image Semantic Classification Visual Object Classes Categorization Multi-label Image Annotation Content Based Image Retrieval Image Understanding
Abstract本论文围绕图像理解与检索的三个重要子课题:自然场景图像的语义分类、视觉目标类识别算法、多标签图像标注算法展开研究。主要内容包括: 1、逐步深入、循序渐进地提出了两种场景图像特征抽取及建模框架。基于Global Gist + Local Topic的场景模型,以认知心理学的相关研究为出发点,兼顾考虑了全局结构性信息和局部话题分布信息,对图像内容进行建模;基于GaborSIFT+NNScSPM的图像抽取算法,有机结合了HMAX思想和非负稀疏编码思想,较为合理地模拟了生物视觉处理的过程。合理借鉴生物机制指导计算机视觉进行特征抽取算法研究,是一个非常有吸引力的研究方向。我们的工作可以为此提供有意义的参考。 2、 在视觉目标类识别研究中,本文提出一种基于互补信息的特征点集融合策略,可以实现使用较为单一的局部特征,就可以达到同期相当的性能。对于实现充分挖掘同一种特征的信息,具有一定的参考意义。 3、针对多标签图像标注问题,提出了一种多标签分类与排序的改进算法。通过引入所谓的“虚拟标签”作为基准标签,对经典的RankSVM算法进行改进,将多标签的分类和排序问题有机融合到算法的统一优化过程,从而算法非常自然地、较好地解决了图像多标签标注问题,有效地弥补了传统RankSVM的重要缺陷。
Other AbstractIn this thesis, we concentrate the research on three subtopics of image understanding and retrieval: natural image semantic classification, visual object categorization, multi-label image annotation. The main contributions include: First, we have proposed two kinds of image feature extraction and modeling frameworks. The scene model based on “Global Gist + Local Topic” has considered both the global structural visual information and local topic distribution information, borrowing some research work in cognitive psychology. Image feature extraction method based on “GaborSIFT + NNScSPM”, try to combine HMAX model and non-negative sparse coding to mimic the information process in V1 area in visual cortex. Biological inspired feature extraction scheme has been demonstrated to be very attractive in computer vision. And our work can provide some reference value for deeper research. Second, in visual object categorization, we have suggested a complementary information fusion strategy for local feature set. For one simple feature set, we suggest to use two representations: traditional bag of features and fisher score representation to make full use of complementary information. Third, we have proposed an improved version of the famous algorithm-RankSVM, named Calibrated RankSVM for multi-label ranking and classification, through incorporating a "virtual label" as a calibration, which acts as a natural zero point to split label set into relevant and irrelevant labels. Multi-label classification and ranking are holistically fused into one optimal problem, which can improve the classification performance while preserve the ranking performance of traditional RankSVM.
Other Identifier200718014628039
Document Type学位论文
Recommended Citation
GB/T 7714
江爱文. 自然场景图像语义识别及标注算法研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2010.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20071801462803(3286KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[江爱文]'s Articles
Baidu academic
Similar articles in Baidu academic
[江爱文]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[江爱文]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.