基于MCA与判别字典学习的场景图文字检测方法

Text detection from natural-scene images using MCA and discriminative dictionary learning

传统的文字检测方法在场景图像复杂背景、噪声污染和文字的多种形态特征的干扰下,检测的准确率很低,漏检、误检非常严重。针对这些问题,提出了基于形态成分分析(MCA)与判别字典学习的场景图像文字检测的方法。通过学习过完备字典将文字检测问题转化成稀疏和鲁棒表示的问题。利用MCA与改进的Fisher判别准则学习一个过完备字典,求解待检测图像文字部分的稀疏系数,重建待检测图像中的文字图像,进行文字检测。通过在ICDAR2003/2005/2011和MSRA-TD500数据库中的大量的实验证明了与其他文字检测方法相比,该方法能有效提高检测准确率。

It is very difficult to locate and recognize text in natural-scene images by interference of complex background, noise pollution and multiple morphological of text using traditional text detection method. Propose a novel method for detecting text in natural-scene images using MCA and discriminative dictionary learning method. Text-detection problems are converted to sparse and robust representations by learning redundant dictionary. An over-complete dictionary is learned using MCA and an improved version of Fisher＇ s discriminant law, the sparse- representation coefficients of text components in the query image are obtained using the learned dictionary. Text image is reconstructed in image to be test, and text test is carried not. The proposed method is extensively evaluated using International Conference on Document Analysis and Recognition （ICDAR）2003/2005/2011 datasets and MSRA-TD500 datasets, and it can effectively improve accurary of detection.