基于时-频结合的显著性区域检测

Saliency Detection Based on Integration of Time-frequency Domains

模拟人的视觉系统对图像中的显著性区域进行检测是计算机视觉领域重要研究内容之一。本文提出了一种新颖的基于时-频结合的显著性区域检测算法,通过结合显著目标在时域的局部对比度和在频域的全局频谱特性,从而有效增强显著目标与背景区域的区分度。为提高时域局部对比度的刻画能力,反应人类视觉"颜色双对立"理论的多通道特性得到了充分考虑。此外,在频域利用奇异值分解(SVD)方法对幅度谱系数矩阵做低秩逼近,通过自适应控制逼近矩阵秩的大小,实现去除高频信息与杂波的目的,而后仅利用相位信息,得到基于频域全局频谱特性的显著图像。在公开数据库上的性能测试结果表明,同目前代表性显著性检测技术相比,本文提出的基于时-频结合的显著性区域检测算法具有更优的检测性能。

Detecting visually salient regions of an image by simulating the visual nervous system of mankind has become one of key issues in the community of computer vision.In this paper,we proposed a novel saliency detection method based on the integration of time-frequency domains.By taking advantage of the local contrast cue from the time domain and the global frequency spectrum cue from the frequency domain,the discrimination between the salient obj ect and background can be effectively enhanced.To boost the characterization of local contrasts in the time domain,multi-channel properties in accordance with the human visual“color double-opponent”system were fully considered.In addition,in the frequency domain we applied SVD to provide a low rank approximation to the magnitude spectrum coefficient matrix and by adaptively tuning the rank of the approximation matrix after SVD,some clutter like noises and high frequency information can be eliminated,then the saliency map in the frequency domain can be obtained only by using the phase information.Compared with some typical saliency detection methods,the proposed algorithm shows competitive performance on several publicly available datasets.