Superpixel Segmentation Based on Grid Point Density Peak Clustering

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Superpixel segmentation is crucial for object recognition but suffers from over-segmentation in homogeneous regions, creating complex textures.
• Existing methods often fail to accurately delineate salient object boundaries, impacting recognition accuracy.

Purpose of the Study:

• To develop an improved superpixel segmentation technique using density peak clustering.
• To effectively reduce redundant superpixels and enhance the representation of primary textures and object contours.

Main Methods:

• Feature points (grid pixels) were extracted and their densities calculated.
• Cluster centers were identified using density peaks.
• Feature points were clustered based on these density peaks to form superpixels.

Main Results:

• The method achieved high performance on the BSDS500 dataset with Boundary Recall (BR) of 95.0% and Achievement Segmentation Accuracy (ASA) of 96.3%.
• Demonstrated efficient processing at 30 frames per second (fps).
• Superpixel boundaries showed strong adherence to salient object textures and contours while reducing noise in homogeneous areas.

Conclusions:

• The proposed density peak clustering-based superpixel segmentation effectively addresses over-segmentation issues.
• The method enhances object boundary detection and texture representation, outperforming existing approaches in accuracy and efficiency.