摘要
同普通视频节目相比,视频广告中的文本具有更为复杂的表现形式.为实现这类文本有效的定位,通过将文本检测视为一种特殊纹理的分类问题,提出一种基于改进的Co-training策略的视频广告文本检测方法,采用两种相对独立的纹理描述子,从多视角来强化文本特性描述.另外,针对Co-training协同学习机制中容易引入噪声样本的问题,提出了一种改进的结合Bootstrap思想的Co-training算法,在两个相对独立的特征空间中交互选择典型样本,以达到提高分类器泛化能力的目的.通过实验,本方法在自建的数据库上获得的正确率与查全率相对于其他方法有10%左右的提高.
The appearance properties of texts in video commercials are more complex than those in general programs. Aiming at locating these texts efficiently, an automatic text detection method based on modified Co-training strategy is proposed in this paper by means of posing text detection as a texture classification problem. Specially, with consideration on the complicated properties of texts in video commercials, two kinds of conditionals independent textual descriptors are extracted for reinforcing the discrimination ability of text from background in multi-view. In addition, to alleviate the problem of noise samples in Co-training, a modified Co-training strategy combining with Bootstrap is presented in this paper. A series of representative samples are selected from those two feature spaces for improving the generalization ability of classifier. The promising experimental results, which are better than the existed method with nearly 10% improvement on precision and recall, show the effectiveness of the proposed method.
出处
《北京交通大学学报》
CAS
CSCD
北大核心
2010年第5期1-7,共7页
JOURNAL OF BEIJING JIAOTONG UNIVERSITY
基金
国家自然科学基金资助项目(60776794)
长江学者与创新团队发展计划项目资助(IRT0707)
中新联合研究计划资助(2010DFA11010)
中央高校基本科研业务费专项资金资助(2009JBZ006-3)
模式识别国家重点实验室开放基金资助项目
关键词
文本检测
协同学习
支持向量机
广告检测
text detection
Co-training
support vector machina(SVM)
commercial detection