A robust hidden Markov Gauss mixture vector quantizer for a noisy source.

Area of Science:

• Image Processing
• Computer Vision
• Pattern Recognition

Background:

• Image noise significantly degrades the performance of sophisticated algorithms like Hidden Markov Gauss Mixture Models (HMGMM).
• Gaussian and Salt and Pepper noise are common in various imaging applications, including medical MRI and aerial imagery.
• Existing methods often require prior denoising or assume specific noise types, limiting their applicability.

Purpose of the Study:

• To develop a modified HMGMM procedure robust to image noise for improved segmentation.
• To propose a noise-agnostic framework that directly segments noisy images without prior denoising.
• To evaluate the performance of the proposed method against established techniques.

Main Methods:

• Modified HMGMM incorporating adjusted covariance matrices in vector quantizer codebooks.
• Minimization of Minimum Discrimination Information (MDI) distortion by adapting covariance matrices to noisy image characteristics.
• Direct application to noisy image segmentation, bypassing traditional denoising steps.

Main Results:

• The proposed procedure demonstrated superior performance compared to median filter restoration and blind deconvolution methods for noisy aerial images.
• Segmentation accuracy closely matched that of HMGMM applied to clean images, both visually and in error rate.
• The method proved effective for images corrupted by both Salt and Pepper and Gaussian noise without prior assumptions on noise type.

Conclusions:

• The modified HMGMM offers a robust and effective solution for segmenting noisy images directly.
• This noise-agnostic approach provides a significant advancement over traditional image restoration techniques.
• The method achieves high segmentation quality comparable to algorithms operating on clean data.