基于视觉语言模型的蒙古文古籍图像检索技术研究

发布时间：2018-01-04 21:23

本文关键词：基于视觉语言模型的蒙古文古籍图像检索技术研究　出处：《内蒙古大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着数字化技术的发展,在内蒙古自治区,越来越多的蒙古文古籍被转成数字图像,以便长久保存。但是以图像格式保存的古籍文档没有索引信息,无法实现全文检索,因此制约了古籍文献资料的传播与利用。本文关注蒙古文古籍图像检索技术,这为挖掘与利用蒙古文古籍提供便利。在图像检索领域,视觉词袋模型是近年来图像检索中最常使用的模型。然而视觉词袋模型有两个缺陷:一是视觉单词之间缺乏语义信息,这导致人对图像的相似性判别与计算机之间存在差异,也就是著名的"语义鸿沟"问题;二是视觉单词之间相互独立,忽略了相邻视觉单词之间的空间信息。本文针对这两个缺陷,提出了解决方法,具体如下:(1)针对视觉词袋模型中视觉单词之间缺少语义信息的问题,提出了视觉语言模型。首先,将扫描得到的蒙古文《甘珠尔经》图像切割成独立的单词图像。其次,从每个单词图像中提取局部描述子(即:视觉单词),并将单词图像表示成视觉单词的概率分布。其中,平滑策略被用于处理零概率问题。再次,根据给定的查询单词图像,采用查询似然模型计算查询单词图像与待检索单词图像间的相似度。最终,按照相似度降序可以生成一个单词图像列表,作为检索结果。(2)我们将视觉单词的空间信息和语义信息相结合,提出了空间视觉语言模型用于表示单词图像。首先,将每个单词图像沿着行、列方向划分为多个等大小的子区域。根据蒙古文的书写方式,我们确定了一种适合于蒙古文单词图像的子区域划分方式。其次,在每个子区域中分别构建视觉语言模型。然后,在单词图像匹配过程中,只有对应子区域的视觉语言模型进行相似度计算。最终,将这些子区域间相似度求和得到两个单词图像的相似度。
[Abstract]:With the development of digital technology in the Inner Mongolia Autonomous Region, Mongolian language more and more were converted into a digital image, in order to store. But in order to save the historical documents without index information, unable to realize the full-text retrieval, thus restricting the spread and utilization of ancient literature. This article focuses on ancient Mongolian image retrieval technology, the mining and the ancient Mongolian convenience. In the field of image retrieval, visual bag of words model in recent years is the most commonly used in image retrieval model. However, bag of visual word has two defects: one is the lack of semantic information of visual words, which leads to the differences between people to distinguish the similarity of images with the computer, also known as the "semantic gap"; two is the visual words are mutually independent, ignoring the spatial information between adjacent visual words. Based on the two. In the solution, the details are as follows: (1) aiming at the lack of semantic information between visual words bag of words model in question, proposed the visual language model. First of all, will scan the Mongolian kanjur cut < > images into words independent image. Secondly, the extraction of local descriptors from each word in the image (namely: visual words), and word image expressed as the probability distribution of visual words. The smoothing strategy is used to deal with zero probability problem. Thirdly, according to the query word image is given, the query likelihood model to calculate the query word image and word similarity between images to be retrieved. Finally, according to similarity can generate a descending order a word image list, as search results. (2) we will be visual words space and semantic information by combining spatial visual language model for the word image is proposed. First, the Each word image along the line direction, the column area is divided into a plurality of equal size. According to the Mongolian writing style, we determine a suitable for the Mongolian word image sub region division method. Secondly, construct the visual language model in each sub region. Then, in the words of the image matching process, only the corresponding visual language model sub region similarity calculation. Finally, the sub region similarity between the sum of two words of image similarity.

【学位授予单位】：内蒙古大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】