基于宽频调谐特征和谱残差分析的显著性目标检测  

Detectingsalient objects based on broadband frequency tuned and spectral residual

作　　者：杨思燕[1] 周国庆[2] 

机构地区：[1]陕西广播电视大学计算机与信息管理系,陕西西安710119 [2]西安电子科技大学计算机学院,陕西西安710071

出　　处：《电子设计工程》2015年第15期177-180,共4页Electronic Design Engineering

基　　金：陕西工商职业学院课题(13G-08-B22);国家自然科学基金(61272280)

摘　　要：根据人眼视觉系统的基本特点,提出一种基于宽频调谐特征和谱残差分析的显著性目标检测算法。该方法通过在上下文信息中提取图像的宽频调谐特征,运用线性叠加融合宽频调谐特征子图得到初步特征图,然后利用多尺度方法构建多分辨率显著图谱子序列,最后利用谱残差分析融合位置特征对显著性子图进行操作运算得到最终的显著图。基于自然图像的显著性目标检测实验结果证明,该方法具有较好的实用性和较强的稳定性,能够获取较为精确的检测结果。According to the visual characteristics of HVS, a salient object detection method based on broadband frequency tuned and spectral residual is proposed. We first extract broadband tuned characteristics of the image from the context to combine broadband tuned features sub-graphs with the principle of linear superposition to obtain feature maps. Then we use multiple scales to construct a multi-resolution saliency maps. Finally we utilize spectral residual to analyze the saliency sub-graphs to get the final saliency graphs. The experimental results on natural images show that the new algorithm is stable and practical, and we achieve satisfied results.

关 键 词：宽频调谐 谱残差 上下文 多尺度 

分 类 号：TN27[电子电信—物理电子学]