ID 原文 译文
22245 现有的大部分基于扩散理论的显著性物体检测方法只用了图像的底层特征来构造图和扩散矩阵,并且忽视了显著性物体在图像边缘的可能性。 Most existing salient object detection methods based on diffusion theory usually only use one feature of image to construct graph and diffusion matrix, and ignore the possibility that salient objects appear at the border regions of the image.
22246 针对此,该文提出一种基于图像的多层特征的扩散方法进行显著性物体检测。 In this paper, a diffusion method based on the multi-layer features of image is proposed to detect salient objects.
22247 首先,采用由背景先验、颜色先验、位置先验组成的高层先验方法选取种子节点。 Firstly, the seed nodes are selected by adopting the high-level prior method, which is composed of background prior, color prior, and location prior.
22248 其次,将选取的种子节点的显著性信息通过由图像的底层特征构建的扩散矩阵传播到每个节点得到初始显著图,并将其作为图像的中层特征。 Then, the initial saliency map is obtained by propagating the saliency information carried by the selected seed nodes to each nodes via the diffusion matrix constructed by the low-level feature of the image, and used as the middle-level feature of image.
22249 然后结合图像的高层特征分别构建扩散矩阵,再次运用扩散方法分别获得中层显著图、高层显著图。 The diffusion matrices are re-synthesized again by the middle-level feature and the high-level feature of the image, and the middle-level saliency map and the high-level saliency map are obtained by the diffusion-based method respectively.
22250 最后,非线性融合中层显著图和高层显著图得到最终显著图。 The final saliency map is obtained by nonlinearly combining the the middle-level and high-level saliency map.
22251 该算法在 3 个数据集 MSRA10K,DUT-OMRON ECSSD 上,用 3种量化评价指标与现有 4 种流行算法进行实验结果对比,均取得最好的效果。 Results on three datasets, MSRA10K, DUT-OMRON and ECSSD, show that the proposed method achieves superior performance compared with the four state-of-art methods in terms of three evaluation metrics.
22252 在非相干分布式非圆信号波达方向(DOA)估计中,针对利用信号非圆特性后输出矩阵维数扩展带来的较大运算量问题,该文提出一种基于互相关抽样分解的 DOA 快速估计算法。 In the Direction Of Arrival (DOA) estimation of incoherently distributed noncircular sources, the increase of dimension caused by array output matrix extension can cause a large computational complexity. To solve this problem, a rapid DOA estimation algorithm is proposed based on cross-correlation sampling decomposition.
22253 该算法仅需要从子阵间的扩展互相关矩阵中抽样出少量行元素和列元素,构成两个低维子矩阵,进而通过低秩近似分解便可快速地同时求出左右奇异矢量,即分别对应两个子阵的信号子空间,避免了计算整个互相关矩阵及其奇异值分解运算; It only needs to calculate two low-dimensional sub-matrices, which are formed by a small number of rows and columns in the extended Cross-Correlation (CC) matrix. On the premise of the sub-matrices, the right and left singular vectors corresponding to two signal subspaces can be simultaneously obtained by the low-rank approximation decomposition, which avoids the calculation of the whole covariance matrix and its singular value decomposition.
22254 最后利用两个子阵信号子空间的旋转不变性通过最小二乘得到 DOA 估计。 Finally, the DOA estimation can be obtained by the least squares with the rotation invariance of the signal subspaces.