Incorporating patch subspace model in Mumford-Shah type active contours.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Segmentation of non-smooth image structures like textures is challenging.
• Existing methods often rely on predefined texture features, limiting their generalizability.

Purpose of the Study:

• To propose a unified energy minimization model for unsupervised segmentation of non-smooth image structures.
• To establish a computational framework based on patch subspaces and linear patch models.

Main Methods:

• Developed a model based on the Mumford-Shah functional and a linear patch model.
• Formulated segmentation as minimizing a piecewise linear patch reconstruction energy.
• Derived an algorithm for globally optimal linear patch reconstruction with linear convergence.

Main Results:

• Segmentation error is bounded by the linear patch reconstruction error.
• The method achieves unsupervised segmentation without predefined features.
• Generated optimized orthonormal descriptors for segmented regions.

Conclusions:

• The proposed method offers superior performance for general texture segmentation compared to existing approaches.
• Improving linear patch reconstruction directly reduces segmentation error.
• The model is suitable for handling diverse textures in an unsupervised manner.