摘要： 贴条信封具有背景复杂、贴条地址块不固定等特点,传统的基于规则的地址块定位方法难以准确识别信封的背景和目标地址块。针对该问题,提出一种基于视觉显著性的贴条信函地址块定位方法。采用二值化归一梯度方法快速检测出图像中与训练地址块具有相似局 部显著性分布的块状区域,作为候选区域,抽取候选区域的位置、灰度、梯度、纹理等基于外观的特征,使用协方差进行非线性融合,生成区域协方差描述子,利用支持向量机实现训练和分类。为了精确定位地址块区域中的文字,用图像签名技术计算稀疏显著性并通过高斯滤 波器进行平滑。实验结果表明,该方法能快速准确定位目标地址块,平均准确率达到85.8%,比基于条件随机场的地址块定位方法高33.5%。

关键词: 视觉显著性, 地址块定位, 区域协方差, 二值化归一梯度, 支持向量机

Abstract: Traditional address localization methods based on rules are difficult to accurately identify the background and target address block in envelope with cluster background due to its complicated background,not fixed position,etc.This paper presents an address block localization method based on vision saliency.It uses the Binarized Normed Gradient(BING) method to identify the candidate regions,whose local saliency maps are similar with the training blocks.Region covariance descriptors are adopted to nonlinerly fuse various low-level features,for example,pixel location,intensity,gradient and texture features.Support Vector Machine(SVM) are applied to classify the cadidate regions into address or non-address block.In order to further pinpoint the address block’s text region in envelopes,it uses the image signature to compute sparsity saliency and Gaussian filter to smooth the saliency map.Experimental results show that this method can accurately locate the target address block with the accuracy of 85.8%,which is 33.5% higher than methods based on conditional random fields.

Key words: vision saliency, address block localization, region covariance, Binarized Normed Gradient(BING), Support Vector Machine(SVM)

中图分类号: 

• TP18