塑封邮件图像的收件人地址块定位

发布时间：2018-04-24 12:29

本文选题：收件人地址块定位 + BING　；参考：《华东师范大学》2017年硕士论文

【摘要】：邮件图像的收件人地址块定位作为邮政自动化的第一步,是实现自动分拣的先决条件。如何定位邮件图像的收件人地址块变成一个亟待解决的问题。尤其是塑封邮件的背景复杂,使得针对塑封邮件图像的收件人地址块定位具有很大挑战性。本文提出了一种塑封邮件图像的收件人地址块定位方法,研究了候选域提取、特征表示和特征匹配的方法。本文的主要工作包括:1.提出基于改进BING模型的候选域提取方法。通过刻画塑封邮件图像中某块区域的目标显著性水平,自适应的产生一些高质量的收件人地址块候选域。该方法克服了滑动窗口法无法适应收件人地址块尺寸变化的缺点,能够减少候选域的数目,降低塑封邮件图像的搜索空间。2.提出采用稠密SIFT描述子对候选域进行特征描述。在词袋模型的基础上引入金字塔匹配原理,将候选域进行层级网格划分,逐层基于视觉词典对候选域的SIFT特征进行重新表示。该方法考虑了候选域中的视觉单词在不同空间位置上的分布情况,保留了候选域的空间布局信息,是对词袋模型的一种有效扩展。3.提出使用直方图交叉核SVM对提取的特征进行匹配。对候选域提取基于视觉词典的金字塔视觉直方图,将直方图作为特征向量输入到训练好的SVM模型中,得到每个候选域的概率。因为单个候选域无法完全覆盖收件人地址块,并且候选域与候选域之间有重叠,所以合并概率最高的前五个候选域,为最终提取收件人地址块的文字区域提供基础。我们在上海邮政科学研究院提供的塑封邮件图像库上进行了实验,并对本文方法和对比方法的数据进行了分析。实验数据显示本文方法取得了较好的结果,可以有效定位出塑封邮件图像的收件人地址块。
[Abstract]:As the first step of postal automation, the location of recipient address block of mail image is the precondition of automatic sorting. How to locate the recipient address block of mail image becomes an urgent problem. Especially, the background of plastic mail is very complicated, which makes it challenging to locate the recipient address block of plastic mail image. In this paper, a method of location of addressee address block in plastic mail image is proposed, and the methods of candidate domain extraction, feature representation and feature matching are studied. The main work of this paper includes: 1. A candidate domain extraction method based on improved BING model is proposed. By characterizing the target significance level of a certain area in a plastic mail image, some candidate domains of high quality addressee address blocks are generated adaptively. This method overcomes the shortcoming that the sliding window method can not adapt to the size change of addressee address block and can reduce the number of candidate fields and the search space of plastic mail image. A dense SIFT descriptor is proposed to describe candidate domains. Based on the lexical bag model, the pyramid matching principle is introduced, the candidate fields are divided into hierarchical grids, and the SIFT features of the candidate domains are re-represented layer by layer based on visual dictionaries. This method takes into account the distribution of visual words in candidate domains in different spatial locations and retains the spatial layout information of candidate domains. It is an effective extension of the word bag model. A histogram cross kernel SVM is proposed to match the extracted features. The pyramid visual histogram based on visual dictionary is extracted from candidate domain and the histogram is input into the trained SVM model as feature vector to obtain the probability of each candidate domain. Because a single candidate domain can not completely cover the addressee address block and there is overlap between the candidate domain and the candidate domain, the first five candidate domains with the highest merging probability provide the basis for the final extraction of the text area of the recipient address block. We have carried out experiments on the plastic mail image library provided by Shanghai Postal Science Research Institute, and analyzed the data of our method and comparison method. The experimental data show that the proposed method can effectively locate the recipient address block of the plastic mail image.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F250;TP391.41

【参考文献】