Tensor Singular Value Decomposition-Based Multiple View Spectral Segmentation

This research presents a novel approach called “Multiple View Spectral Segmentation based on Tensor Singular Value Decomposition” for the segmentation of multi-view data. The algorithm utilizes three-rank tensors and constructs a probability transfer matrix for all view data. By exploiting the low-rank nature of tensors in the lateral, longitudinal, and vertical directions, the proposed procedure characterizes the tensor's low-rank properties in each dimension using a multi-rank approach based on tensor singular value decomposition (Tensor-SVD). Tensor-SVD decomposition, being based on tube convolution, enables the model to capture spatial correlations more effectively compared to other tensor resolution techniques and procedures based on two-dimensional structure relationships. Furthermore, the use of Fourier transformation allows for efficient calculations, thereby improving computational efficiency. Experimental results demonstrate that the proposed tensor resolution model based on Tensor-SVD achieves improved segmentation performance for multiple-view data.