STAR

STAR performs structure- and texture-aware Retinex decomposition of images to separate illumination and reflectance components using exponentiated local derivatives for improved illumination/reflectance estimation and downstream tasks such as low-light enhancement and color correction.

Key Features:

• Retinex-based decomposition: Implements separation of illumination and reflectance components based on Retinex principles and analysis of local image derivatives.
• Exponentiated local derivatives (γ): Uses exponentiated local derivatives with an adjustable exponent γ to modulate emphasis on structure (γ>1) versus texture (γ<1).
• Structure and texture maps: Generates distinct structure and texture maps that emphasize prominent features and fine details respectively.
• Exponential filters: Applies exponential filters to local derivatives to produce precise structure and texture maps.
• Regularization in decomposition: Incorporates the generated structure and texture maps as regularizers within the Retinex decomposition framework to improve separation accuracy.
• Optimization and solvers: Solves the model using an alternating optimization algorithm with subproblems formulated as vectorized least squares regression problems admitting closed-form solutions.
• Empirical validation: Demonstrates quantitative and qualitative improvements over prior methods on commonly used datasets for illumination/reflectance decomposition, low-light enhancement, and color correction.

Scientific Applications:

• Illumination and reflectance decomposition: Separates and estimates illumination and reflectance components for image analysis and correction.
• Low-light image enhancement: Enhances visibility and contrast in low-light images via improved decomposition of illumination and reflectance.
• Color correction: Improves color fidelity by correcting reflectance estimates.
• Structure and texture analysis: Extracts structure and texture maps for feature-focused image analysis and processing.

Methodology:

The model is solved by alternating optimization; each subproblem is transformed into a vectorized least squares regression problem with closed-form solutions, and exponentiated local derivatives combined with exponential filters are used to compute structure and texture maps that serve as regularizers in the Retinex decomposition.