Optimisation of Convolution-Based Image Lightness Processing

Area of Science:

• Computer Vision
• Image Processing
• Statistical Modeling

Background:

• Convolutional retinex methods use center/surround operators to reduce shading and dynamic range.
• Existing methods often tune parameters for visual appeal and include enhancement functions like logarithmic mapping.

Purpose of the Study:

• To introduce and detail a statistical approach to convolutional retinex based on autocorrelation statistics.
• To objectively mitigate shading without subjective image enhancement components.

Main Methods:

• Modeling autocorrelation matrices for image albedo and shading.
• Solving a linear regression to obtain optimal filters in closed form.
• Analyzing the impact of autocorrelation matrix shape on optimal filter shape.

Main Results:

• The statistical approach yields an objectively optimal filter.
• Demonstrated effectiveness in shading removal from text documents.
• Validated performance on a challenging image dataset.

Conclusions:

• The statistical convolutional retinex method provides an objective approach to shading mitigation.
• Autocorrelation statistics are crucial in determining optimal filter characteristics.
• The method shows promise for various image processing applications, including document analysis.