基于随机投影的场景文本图像聚类方法研究  

Research on clustering of natural scence images with texts based on random projection

作　　者：徐飞[1] 刘家锋[1] 张博宇[1] 吴锐[1] 赵巍[1] 

机构地区：[1]哈尔滨工业大学模式识别研究中心,哈尔滨150001

出　　处：《计算机应用研究》2011年第12期4730-4733,共4页Application Research of Computers

基　　金：国家自然科学基金重点资助项目(61073128);黑龙江省青年基金资助项目(QC2009C35);哈尔滨工业大学校基金资助项目(NSRIF 2009.158)

摘　　要：图像中的文本字符存在于杂乱的背景之中,拍摄视角的不同使得文本具有较大的几何变形,再加上存在光照变化、字符颜色不统一等现象会导致背景分离和文本识别困难。为此提出一种基于图像文本区域的图像聚类方法。该方法首先对自然场景图像中已定位的文本区域提取局部特征描述,并使用随机投影方法将局部特征矢量集映射为固定维的特征向量,然后对包含图像文本区域的图像进行聚类。这种方法避免了由图像分割与字符识别带来的困难。实验结果表明,该方法可以对包含文字的自然场景图像有效地进行聚类,聚类的准确率能达到86.66%。Text characters in the images are always in a complex background and the different film perspective causes the text characters with large geometric deformation. The illumination or the character color is always not uniform, leading to difficul- ties in the background separating and text recognition. This paper proposed a method which was based on the image text area for image clustering. Firstly, this method extracted local feature description of the targeted text area in the image and used random projection method to map the local feature vectors into a vector which had a specific dimension. Then to cluster these images which contained text characters. The method avoided the difficuhies of image segmentation and character recognition. Experimental results show that the method can be effective clustering for the natural scene images that contain text characters and the accuracy is 86.66%.

关 键 词：图像文本区域 图像聚类 随机投影 局部特征描述 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]