Consensus Sparsity: Multi-Context Sparse Image Representation via L_∞-Induced Matrix Variate.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Sparsity is crucial for image processing tasks like representation, compression, and analysis.
• Existing sparse representation (SR) methods utilize L0-norm or L1-norm regularization but do not fully exploit multi-context sparsity.

Purpose of the Study:

• To introduce the concept of consensus sparsity (Con-sparsity) for a more comprehensive exploitation of sparsity in image representation.
• To develop a multi-context sparse image representation (MCSIR) framework that simultaneously learns within-sample, between-sample, and structural sparsity.

Main Methods:

• The proposed consensus sparsity is theoretically achieved using L-infinity-induced matrix variates within a Bayesian inference framework.
• A novel multi-context sparse image representation (MCSIR) framework is developed to integrate these sparsity properties.

Main Results:

• The MCSIR framework effectively integrates and learns multiple sparsity contexts simultaneously.
• Experimental results demonstrate superior performance compared to state-of-the-art methods, including deep learning approaches.

Conclusions:

• Consensus sparsity offers a more complete exploitation of sparsity properties for enhanced image representation.
• The proposed MCSIR framework provides a powerful and versatile tool for advanced image processing tasks.