A unified tensor level set for image segmentation

Area of Science:

• Computer Vision
• Image Processing
• Computational Mathematics

Background:

• Traditional region-based level set methods often struggle with noise and accurately capturing complex image features.
• Representing pixel information solely through scalar values limits segmentation performance in diverse image types.

Purpose of the Study:

• To introduce a novel region-based unified tensor level set model for enhanced image segmentation.
• To improve robustness against noise and boundary detection accuracy in image segmentation tasks.
• To generalize region-based level set methods to handle high-order tensor data.

Main Methods:

• A three-order tensor is employed to represent pixel features, including gray value and local geometry (orientation, gradient).
• A weighted distance metric is defined to generalize scalar-based region-based level set methods to tensor representations.
• Gaussian filter banks are incorporated for noise robustness, particularly against salt-and-pepper noise.

Main Results:

• The proposed tensor-based model demonstrates superior performance in segmenting synthetic, medical, and natural images.
• The model exhibits enhanced accuracy and natural segmentation by utilizing unified tensor pixel representations.
• Improved boundary adherence is achieved by considering local geometrical features like orientation and gradient.

Conclusions:

• The unified tensor level set model offers a significant advancement in image segmentation, outperforming existing methods.
• The model's ability to handle diverse data types (scalar to tensor) makes it highly versatile.
• This approach provides a robust and accurate solution for complex image segmentation challenges.