Attention-Guided Low-Rank Tensor Completion.

Area of Science:

• Computer Vision
• Data Science
• Machine Learning

Background:

• Low-rank tensor completion (LRTC) methods struggle to preserve original data structures and are computationally expensive.
• Existing LRTC algorithms often yield inaccurate restoration results due to limitations in preserving tensor structures.

Purpose of the Study:

• To develop an attention-guided low-rank tensor completion (AGTC) algorithm for accurate and efficient data recovery.
• To enhance the preservation of original data tensor structures during the completion process.
• To reduce the computational cost associated with LRTC.

Main Methods:

• Formulating LRTC as a robust factorization problem with low-rank and sparse error assumptions.
• Employing an attention mechanism to guide low-rank tensor recovery and preserve original data structures.
• Developing implicit regularizers to address modeling inaccuracies.
• Solving the optimization problem using an iterative technique and unfolding it into a multistage deep network.

Main Results:

• The proposed AGTC algorithm demonstrates superior performance in restoring original data tensor structures.
• Experimental results show that AGTC outperforms state-of-the-art algorithms in high dynamic range imaging and hyperspectral image restoration.
• The deep unfolding network effectively updates optimization variables and learned regularizers at each stage.

Conclusions:

• AGTC offers a significant advancement in low-rank tensor completion by effectively preserving data structures.
• The attention-guided approach and deep unfolding network contribute to improved accuracy and efficiency in data restoration.
• AGTC shows strong potential for applications in image processing and other high-dimensional data recovery tasks.