Low-Rank Subspace Learning of Multikernel Based on Weighted Truncated Nuclear Norm for Image Segmentation

Previous natural image segmentation algorithms through subspace learning method have over-segmentation issues in the pre-segmentation process, which will compromise the edge information, and the subspace learning model cannot effectively utilize the nonlinear structure in the image data and has weak resistance to multiple noises. To address these problems, a multi-kernel subspace learning method based on weight truncated Schatten-p norm for image segmentation is designed in this paper. First, the original natural image pre-processing operation, which is conducting adaptive morphological reconstruction watershed transformation on the image, then the original pixels are aggregated to form a superpixel image, of which the obtained superpixel block would retain more comprehensive local information; Secondly, perform feature extraction for each superpixel block, and stack the obtained feature vectors into the desired feature matrix; Then, it is input into the weighted truncated Schatten-p low-rank multi-kernel subspace learning model to obtain a similarity matrix with cluster structure on the diagonal; Finally, the similarity matrix is used as the adjacency matrix in the spectral clustering model, and the final feature data clustering and image segmentation results are obtained by the optimization solution. The final experimental results demonstrate that contrasts to existing clustering models, the proposed method accomplishes the best clustering property on two public datasets; Compared with seven segmentation algorithms on the BSDS500 dataset, and achieved the best segmentation effect on two evaluation metrics.