图像检索中判别性增强研究

发布时间：2018-07-12 12:45

本文选题：图像检索 + 视觉词袋模型　；参考：《吉林大学》2016年博士论文

【摘要】：本文围绕几何验证和提升全局特征描述能力两大方面对图像检索方法展开研究。研究过程中,沿着提高特征点间区分能力的思想,以融入显著性机制的几何验证方法、提取高层语义信息、高效特征间几何编码和增强聚合特征的判别性四方面提出改进方案,实现对大规模图像数据库进行高效检索。几何验证方法融入显著性机制时,为了快速准确搜索用户感兴趣图像,提出基于分级显著信息的空间编码方法。采用三点构成的角度信息和距离信息分别编码并验证,加强匹配特征验证稳定性。同时引入显著性机制采用分级多点编码方法提高几何验证阶段实时性。为了进一步提取图像高层语义信息,增强特征间区分能力,基于最大熵原则提取图像分级显著区域,同时在这些分级显著区域中,定义一种三角形的空间模型,加强几何编码约束条件,最终融合显著性匹配得分与几何得分。高效特征间几何编码时,为了提高检索性能,本文通过局部邻域约束方法过滤不相关匹配特征,可去除特征匹配时的噪声影响,同时为了加强空间约束能力,对相关匹配特征间相对位置进行编码,并验证其几何一致性。这种方法既减少检索时间,又提高了检索准确率。为了增强聚合特征的判别性,本文首先将特征描述子在每个视觉单词中的熵分布信息嵌入至聚合特征中,同时为了有效将两种特征进行组织,在融合阶段提出了新的指数归一化方法,进而实现检索性能的提升。
[Abstract]:This paper focuses on geometric verification and global feature description of image retrieval methods. In the course of the research, an improved scheme is put forward along with the idea of improving the ability of distinguishing feature points, incorporating the geometric verification method of salience mechanism, extracting high-level semantic information, efficiently coding the geometry between features and enhancing the discriminability of aggregate features. The efficient retrieval of large scale image database is realized. In order to quickly and accurately search users' interesting images, a spatial coding method based on hierarchical salient information is proposed when geometric verification method is incorporated into saliency mechanism. Angle information and distance information are coded and verified separately to enhance the stability of matching feature verification. At the same time, the significance mechanism is introduced to improve the real-time performance of geometric verification stage using hierarchical multipoint coding. In order to further extract high-level semantic information of images and enhance the ability of distinguishing features, image hierarchical salience regions are extracted based on maximum entropy principle, and a triangular spatial model is defined in these hierarchical salient regions. The constraint condition of geometric coding is strengthened, and the significant matching score and geometric score are fused finally. In order to improve the retrieval performance of efficient geometric coding between features, the local neighborhood constraint method is used to filter the uncorrelated matching features, which can remove the noise in feature matching and enhance the spatial constraint ability. The relative positions of correlation matching features are coded and their geometric consistency is verified. This method not only reduces the retrieval time, but also improves the retrieval accuracy. In order to enhance the discriminability of aggregate features, the entropy distribution information of feature descriptors in each visual word is embedded into the aggregate features, and the two features are organized effectively. In the fusion stage, a new exponential normalization method is proposed to improve the retrieval performance.
【学位授予单位】：吉林大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】