MATLAB Implementation of Multi-Scale Retinex Algorithm

Multi-Scale Retinex is a classical image enhancement algorithm primarily used to improve visual quality of images with uneven illumination or low contrast. Its core principle simulates the human visual system's perception mechanism for illumination and reflection components, decomposing images through multi-scale Gaussian filtering to restore more natural colors and details.

The MATLAB implementation of Multi-Scale Retinex typically involves these key steps:

Image Preprocessing Convert input images to logarithmic domain using log() function to facilitate separation of reflection and illumination components in subsequent steps.

Multi-Scale Gaussian Filtering Apply convolution with Gaussian kernels of different scales (typically small, medium, and large) using fspecial('gaussian', [size], sigma) and imfilter() functions. Small sigma values capture fine details while large sigma values handle overall illumination.

Reflection Component Extraction Subtract the Gaussian-filtered results from the logarithmic-domain image using element-wise subtraction (log_image - filtered_image) to obtain approximate reflection components representing detail information without illumination effects.

Multi-Scale Fusion Combine reflection components from different scales through weighted averaging (w1*component1 + w2*component2 + w3*component3) where weights are typically normalized. This preserves detail enhancement while preventing excessive distortion.

Post-processing and Color Restoration Implement color correction strategies like gain compensation (using imadjust() or histeq()) or color balance techniques to address color shifts common in Retinex methods. Apply exponential transformation (exp()) to convert results back to linear domain.

Advanced implementations incorporate adaptive weight adjustment through variance-based weighting or noise estimation functions to prevent noise amplification from small-scale filtering while maintaining large-scale filtering's illumination balancing capability. Key parameters like Gaussian kernel standard deviations (sigma values typically ranging 15-80, 80-250, 250-480) and fusion coefficients can be optimized for different image enhancement scenarios.