Single-Scale, Multi-Scale, and Original Retinex Algorithms

The Retinex algorithm is a core technology for image illumination enhancement, rooted in the color constancy theory of the human visual system. By decomposing an image into reflectance and illumination components, it effectively addresses uneven lighting issues and holds significant application value in fields such as medical imaging and autonomous driving. In code implementation, this typically involves logarithmic transformations and Gaussian filtering operations to separate illumination effects from surface properties.

Single-Scale Retinex (SSR) serves as the foundational implementation, using a single Gaussian kernel for convolution operations to process the illumination component. Its advantages lie in straightforward implementation and computational efficiency, though it may produce halo artifacts near strong illumination boundaries. Key implementation considerations include Gaussian kernel parameter selection: smaller σ values preserve more details but yield weaker enhancement, while larger σ values improve illumination uniformity at the cost of texture loss. MATLAB's imgaussfilt() function is commonly used for efficient Gaussian convolution.

Multi-Scale Retinex (MSR) innovatively combines results from multiple Gaussian kernels of different scales. A typical implementation integrates three scales (e.g., σ=15/80/250) through weighted fusion, maintaining local detail contrast while suppressing halo effects. This multi-scale approach enhances robustness in complex lighting scenarios but increases computational complexity. Code implementation often involves parallel processing of different scales followed by weighted averaging using coefficients like [0.33, 0.33, 0.34].

The original Retinex algorithm, as the theoretical foundation, employs random path estimation for illumination components. While mathematically significant, it suffers from computational inefficiency and result instability in practical applications, leading modern improvements to primarily build upon SSR and MSR frameworks. This version demonstrates the theoretical basis for reflectance recovery through random walk iterations.

Critical MATLAB implementation aspects include: optimization techniques for rapid Gaussian kernel computation, numerical stability handling during logarithmic domain conversions, and dynamic range compression methods for output results. The algorithm shows particularly notable improvement for low-light images and backlit photographs, but requires careful parameter tuning to avoid noise amplification from excessive enhancement. Practical code should incorporate histogram equalization or gamma correction for output normalization.