An ADMM Approach to Masked Signal Decomposition Using Subspace Representation

Area of Science:

• Signal Processing
• Optimization
• Computer Vision

Background:

• Signal decomposition separates observed signals into components with distinct properties.
• Traditional methods focus on additive components, but real-world signals are often non-additive.
• Non-additive decomposition requires solving complex binary optimization problems.

Purpose of the Study:

• To develop a novel algorithm for signal decomposition that handles non-additive components.
• To address the challenges posed by binary optimization in signal separation.
• To demonstrate the effectiveness of the proposed method across diverse applications.

Main Methods:

• Relaxing intractable binary optimization problems to approximated continuous problems.
• Employing an alternating optimization technique to solve the relaxed problem.
• Applying the algorithm to image, video, and 1-D signal decomposition tasks.

Main Results:

• The proposed algorithm successfully separates non-additive signal components.
• Significant improvements were observed in separating text from image backgrounds.
• Effective separation of moving objects in videos and components in 1-D signals was achieved.
• The non-additive approach demonstrated superior performance over traditional methods.

Conclusions:

• Considering the non-additive nature of signals is crucial for accurate decomposition.
• The developed alternating optimization technique provides an effective solution for non-additive signal separation.
• The method shows broad applicability and significant performance gains in practical scenarios.