基于自适应幅度谱分析的显著目标检测

Salient Object Detection Based on Adaptive Amplitude Spectrum Analysis

针对HFT显著性检测方法中出现的显著目标突显不均匀以及显著信息丢失问题,结合尺度空间分析的方法,提出一种基于自适应幅度谱分析的显著目标检测算法。该算法利用频域滤波尺度与显著目标尺寸之间的关系,对图像幅度谱的尖刺进行处理,突出显著性区域,保证了显著目标的均匀性。同时引入高斯偏见对多幅显著图进行融合,进一步保留显著信息,改善了显著图效果。在国际公开的数据集和藻种图像数据集上,该算法的检测结果具有更高的精度和更好的召回率,优于现有的其他频域显著性检测算法。

Since HFT model may fail to highlight the salient objects uniformly and lose some meaning- ful saliency information, this paper proposes a salient object detection method via adaptive amplitude spectrum analysis, which combines with the scale-space analysis. In order to pop out the salient region more uniformly, the algorithm smoothes the spikes in the amplitude spectrum using the specific relation between the size of salient region and the scale of filter in frequency domain. We also introduce the Gaussian model of center bias to combine different saliency maps with more meaningful saliency informa- tion. The performance evaluation on two popular benchmark data sets validates that our method gets higher precision and better recall and so outperforms the existing spectral saliency models.