摘要
基于部件的树结构模型(TSM)使用的底层特征梯度直方图(HOG)对文字特征表达性不强,且降维时易造成信息丢失。针对该问题,构建以稀疏编码直方图(HSC)为底层特征的基于部件的树结构模型(HSC-TSM)识别场景文本。将K-SVD学习字典用于计算稀疏编码,逐像素地将稀疏编码聚合成HSC,描述文字的局部外观信息;通过奇异值分解对HSC进行降维,避免信息丢失。HSC-TSM在数据集ICDAR 2003上对各类文字的识别率比TSM高3.08%-10.28%,在数据集ICDAR 2003和SVT上的单词识别率分别提升了5.30%和3.62%。
The histograms of gradient(HOG)as low-level feature of the part-based tree-structured model(TSM)is not representative for characters,and it can easily lead to the loss of information when reducing the dimensions.To solve the problem,histograms of sparse codes(HSC)as low-level feature of the part-based tree-structured model(HSC-TSM)was constructed to recognize scene text.Sparse codes were computed with dictionaries learnt from data using K-SVD,and per-pixel sparse codes were aggregated into HSC,the local appearance information was better described.The dimensions of HSC were reduced by singular value decomposition to avoid the loss of information.The recognition rates of HSC-TSM recognizing various categories of characters on ICDAR 2003 dataset are 3.08%-10.28% higher than that of TSM,the word recognition rates on ICDAR 2003 and SVT dataset are respectively increased by 5.30% and 3.62%.
出处
《计算机工程与设计》
北大核心
2016年第4期988-992,1090,共6页
Computer Engineering and Design
基金
2014年国家科技支撑计划基金项目(2014BAH30B01)
关键词
场景文本识别
基于部件的树结构模型
奇异值分解
稀疏编码直方图
scene text recognition
part-based tree-structured model
singular value decomposition
histograms of sparse codes