基于区域注意力机制的遥感图像检索

遥感图像存在大量语义对象,相同的语义对象视觉差异较大,针对卷积神经网络(CNN)提取的全局特征不能准确描述图像内容的问题,提出了一种使用区域注意力机制的遥感图像检索方法。首先去除CNN的全连接层,将高层特征作为区域注意力网络的输入;然后在遥感图像数据集上分别训练CNN和区域注意力网络,提取具有区域关注度的图像特征;最后构建了一种多距离相似性度量矩阵并采用扩展查询以提高检索性能。实验结果表明,相比基于全局特征的遥感图像检索方法,本方法能有效抑制遥感图像背景和不相关的图像区域,在两大遥感实验数据集上的检索性能更好。

Fast vectorized distance matrix computation for multiple sequence alignment on multi-cores

Although high quality multiple sequence alignment is an essential task in bioinforma- tics, it becomes a big dilemma nowadays due to the gigantic explosion in the amount of molecular data. The most consuming time and space phase is the distance matrix computation. This paper addresses this issue by proposing a vectorized parallel method that accomplishes the huge number of similarity comparisons faster in less space. Per- formance tests on real biological datasets using core-iT show superior results in terms of time and space.