Improvement in the Between-Class Variance Based on Lognormal Distribution for Accurate Image Segmentation

Area of Science:

• Computer Vision
• Image Processing
• Digital Signal Processing

Background:

• Image histograms represent intensity level frequencies, crucial for segmentation.
• Otsu's method, a standard for optimal thresholding, uses Gaussian distribution and maximizes between-class variance.
• Asymmetric or right-skewed histograms pose challenges for the original Otsu method.

Purpose of the Study:

• To propose an improved Otsu method for image segmentation.
• To address the limitations of Otsu's method with asymmetric and right-skewed image histograms.
• To enhance the accuracy of image segmentation by adapting thresholding techniques.

Main Methods:

• Developed an improved Otsu algorithm incorporating lognormal distribution.
• Utilized the mean and variance of the lognormal distribution to refine between-class variance calculation.
• Tested the proposed model on simulated images and Medical Resonance Imaging (MRI) brain tumor datasets.

Main Results:

• The proposed lognormal-based Otsu method demonstrated superior performance compared to the original Otsu method.
• Segmented images from the improved method showed better threshold estimation accuracy.
• Evaluation using both unsupervised and supervised metrics confirmed the model's effectiveness.

Conclusions:

• The lognormal distribution-based improvement effectively handles right-skewed histograms in image segmentation.
• The enhanced Otsu method offers a more robust solution for segmenting images with asymmetric intensity distributions.
• This approach provides better threshold estimation for complex image datasets, including medical imaging.