Tensorial Multiview Representation for Saliency Detection via Nonconvex Approach

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Multiview features are crucial for salient object detection.
• Current patch-based methods lose spatial structure and view interactions by stacking features.
• Existing approaches fail to fully exploit complementary information from multiview features.

Purpose of the Study:

• To develop a tensorial feature representation framework for salient object detection.
• To better explore complementary information from multiview features.
• To improve the accuracy and robustness of salient object detection.

Main Methods:

• A tensor low-rank constraint is applied to the background.
• Tensor group sparsity regularization is used for the salient part.
• Tensorial sliced Laplacian regularization enhances background-salient object separation.
• A nonconvex tensor Log-determinant function approximates tensor rank for complex backgrounds.

Main Results:

• The proposed method outperforms the latest unsupervised handcrafted features-based methods.
• Experiments on five public datasets validate the effectiveness of the framework.
• The model demonstrates flexibility with various deep features.

Conclusions:

• The tensorial feature representation framework offers a superior approach to salient object detection.
• The method effectively handles complex backgrounds and enhances feature representation.
• The proposed model is competitive with state-of-the-art approaches and adaptable to deep learning features.