非正交二值子空间模板表示的并行生成

非正交二值子空间是一种重要的图像模板表示方法,能够高效地进行模板匹配。针对其生成过程耗时多难以实时应用的问题,提出了一种基于统一计算设备架构(CUDA)的非正交二值子空间模板表示的快速并行生成算法。该算法将模板生成过程划分为3个步骤,并根据每个步骤数据处理的特点,进行并行任务分解,极大地缩短了生成时间。实验结果表明,相比基于CPU生成相同的模板表示,该算法的速度提升了60～120倍。在应用于模板匹配时,该并行的模板表示生成算法能够很好的改进匹配效率。

A QUASI-REFINED ITERATIVE ALGORITHM BASED ON THE LANCZOS BIORTHOGONALIZATION PROCEDURE FOR LARGE EIGENPROBLEMS

The two-sided Lanczos method is popular for computing a few selected eigentriplets of large non-Hermitian matrices. However, it has been revealed that the Ritz vectors gained by this method may not converge even if the subspaces are good enough and the associated eigenvalues converge. In order to remedy this drawback, a novel method is proposed which is based on the refined strategy, the quasi-refined idea and the Lanczos biothogonalization procedure, the resulting algorithm is presented. The relationship between the new method and the classical oblique projection technique is also established. We report some numerical experiments and compare the new algorithm with the conventional one, the results show that the former is often more powerful than the latter.

多特征综合的视频拷贝检测

基于内容的视频拷贝检测是多媒体领域的一个研究热点。由于拷贝变换的多样性和综合性,单一特征难以获得很好的检测效果。提出一种多特征综合的方法来提高视频拷贝检测的效果。除了使用传统的局部和全局视觉特征外,还使用非正交二值子空间(NBS)方法来表示视频内容,并在其基础上使用归一化互相关(NCC)来提高拷贝视频内容相似度计算的效果。在此基础上,还采用多种措施对拷贝视频的判定结果进行精化。实验结果表明,该套方案对多种拷贝变换具有很强的鲁棒性,并且能够得到很好的检测精度。