结构化Krylov-SVD分解的显著性目标检测算法

Salient Object Detection Algorithm Based on Structured Krylov-SVD Decomposition

针对低秩矩阵在图像显著性检测中,因凸松弛迭代奇异值分解导致的计算复杂度高及稀疏矩阵元素间潜在结构关系未充分考虑导致的显著图发散或不完整现象,提出了一种结构化低秩矩阵Krylov-SVD分解的显著性目标检测算法。该算法对Arnoldi模型进行了深入研究,在Krylov-Schur重启算法的基础上对Schur分解进行改进,给出了Krylov-SVD奇异值分解算法,通过求其前k个特征值,对稀疏矩阵进行降阶处理,以降低计算复杂度;随后引入了索引树结构化稀疏范数,利用分层稀疏正则化来连接稀疏矩阵中元素之间的空间关系。实验中采用MSRA10K、SOD和ECSSD三个公开数据集、四种评价指标,与现有的十一种算法进行了对比实验。实验结果表明,该显著性目标检测算法在时间性能和精准性方面有着良好表现。

A salient object detection algorithm based on structured Krylov-SVD decomposition of low-rank matrices is proposed due to the high computational complexity caused by the convex relaxation iterative singular value decomposition and the incompleteness of the saliency map caused by the inadequate consideration of potential structural relationships between sparse matrix elements.The algorithm provides an in-depth study of the Arnoldi model,improves the Schur decomposition on the basis of the Krylov-Schur restart algorithm,and gives the Krylov-SVD singular value decomposition algorithm,which reduces the computational complexity by deregulating the sparse matrix by solving its first k eigenvalues.And then the index tree structured sparse parametric is introduced,and hierarchical sparse regularization is used to connect the spatial relationships between elements in the sparse matrix.Three public data sets,MSRA10K,SOD and ECSSD,and four evaluation metrics are used in the experiments for comparison with eleven existing algorithms.The experiment shows that the proposed algorithm has better performance in terms of time performance and accuracy.