基于超复数傅里叶变换的自适应显著性检测  

Adaptive saliency detection based on hypercomplex Fourier transform

作　　者：黄侃[1,2,3] 张涌[2] 吕波[1,2] 

机构地区：[1]中国科学院大学,北京100049 [2]中国科学院上海技术物理研究所,上海200083 [3]南洋理工大学电子与电气工程学院,新加坡999002

出　　处：《计算机应用》2017年第A01期149-154,共6页journal of Computer Applications

基　　金：国家863计划项目(2001AA7031002G)

摘　　要：针对已有方法对不同尺寸的显著目标以及对多个显著目标检测效果不佳的问题,提出了一种基于超复数傅里叶变换(HFT)模型的改进算法。通过将HFT模型的后处理过程扩展为中值滤波和自适应高斯模糊处理,解决了对小尺寸目标及多目标的检测问题,取得了更加精确和稳定的检测效果。首先,将原始图像变换到对立色彩空间(RG-BY-I),并进行超复数傅里叶变换;然后,对图像频谱进行多尺度频域滤波、超复数傅里叶反变换得到一组多尺度下的显著图;最后,依据熵准则从显著图组中选取最优显著图,并通过中值滤波和自适应高斯模糊得到最终显著图。在Bruce、Judd、Img Sal三个数据集上与9种方法进行对比实验,受试者接收曲线下面积(AUC)在3个数据集上分别较目前最佳模型提高了2.5%、5.3%和1.2%。实验结果表明,所提方法能够很好地检测不同尺寸的显著性目标,清晰地区分多个显著性目标。Concerning that existing mothods cannot efficiently detect salient objects of different sizes and multiple salient objects, an improved method based on HFT（ Hypercomplex Fourier Transform） model was proposed. In this method, median filtering and adaptive Gaussian blurring were incorporate into the post-processing of HFT model, solving the problem of detecting salient objects of small size and multiple salient objects, thus achieving more accurate and stable detection. First,the original image was converted to opponent color space representation, then HFT was performed on the opponent color representation. Second, a set of saliency maps were obtained by subsequent multi-scale frequency filtering and inverse HFT.Finally, a optimal map was selected among the set of saliency maps according to entropy criterion, and then processed by median filtering and adaptive Gaussian blurring to get the final saliency map. The comparison experiments with 9 other methods were conducted on Bruce, Judd, Img Sal datasets, ang the AUC scores of the proposed method improved 2. 5%,5. 3%, 1. 2% respectively compared to the existing state-of-the-art method HFT. The experimental results show that the proposed method can extract salient regions of different sizes more robustly and separate multiple salient objects more accurately.

关 键 词：显著性检测 超复数傅里叶变换 自适应检测 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]