Superpixel Segmentation Using Gaussian Mixture Model.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Superpixel segmentation is crucial for computer vision, simplifying images into meaningful regions.
• Existing methods often lack efficiency or produce superpixels that do not accurately follow object edges.

Purpose of the Study:

• To introduce a novel pixel-related Gaussian mixture model (GMM) for efficient superpixel segmentation.
• To achieve superpixels with similar sizes and improved boundary adherence.

Main Methods:

• Developed a GMM with constant weights and subset Gaussian functions for linear complexity.
• Implemented an inherently parallel algorithm suitable for multicore systems.
• Incorporated eigenvalue truncation during expectation-maximization to control superpixel regularity.

Main Results:

• The proposed GMM achieves linear complexity with respect to the number of pixels.
• The algorithm demonstrates inherent parallelism for fast execution.
• Experimental results show superior adherence of generated superpixels to object boundaries compared to state-of-the-art methods.

Conclusions:

• The novel GMM offers an efficient and parallel approach to superpixel segmentation.
• This method enhances image analysis by producing more accurate superpixels.
• The technique provides a significant advancement in preprocessing for computer vision tasks.