基于频域多尺度分析的显著性检测

Saliency Detection Method Based on Multiscale Analysis in Frequency Domain

针对现有显著性检测方法得到的显著区域不完整以及缺乏生物学依据的不足,提出一种基于频域多尺度分析的图像显著性检测方法.首先利用小波变换将输入图像的离散余弦变换(DCT)系数的幅度谱进行多尺度分解,计算得到多尺度下的空间域视觉显著图,然后依据显著性评价函数选出较优显著图,最后以自适应权重合成输入场景的视觉显著图.对不同类型数据集进行实验,包括心理物理学模板数据集、人眼注视轨迹数据集及显著目标分割数据集(包括ASD和ECSSD数据集),该方法对于多类型数据集在P-R曲线、ROC曲线及AUC指标等客观评价标准上均取得较高精确度,且在计算速度统计中计算较快,表明该方法优于其他经典的显著性检测方法.

Since most existing approaches of saliency detection generate incomplete and biologically implausible saliency map,in this paper,we proposed a saliency detection method based on multiscale analysis in frequency domain.Firstly,we used the wavelet transform to decompose the discrete cosine transform(DCT)coefficients of the input image into multi-scale amplitude spectra,and calculated the multi-scale saliency maps in the spatial domain.Then,we selected the better saliency maps according to the saliency evaluation function.Finally,we obtained a final saliency map of the input image by adaptive weighted integration of the selected multi-scale saliency maps.Experiments were conducted on different types of datasets,including psychophysical pattern,human eye fixation datasets,and salient target segmentation datasets(including ASD and ECSSD).The method achieves high accuracy on objective evaluation such as P-R curve,ROC curve,AUC index and calculation speed of these datasets.Experiments results show that the method outperforms other conventional approaches of saliency detection.