Classification-driven watershed segmentation.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Traditional watershed segmentation relies on manually defined or automatically extracted topographical markers.
• Existing methods often struggle with single-channel or multi-channel image data and require specific gradient image processing.

Purpose of the Study:

• To present a novel classification-driven watershed segmentation approach.
• To enhance the creation of topographical function and object markers.
• To improve performance in image-based granulometry and remote sensing.

Main Methods:

• Utilizes two machine-learned pixel classifiers: one for marker generation and another for boundary detection.
• Applies the watershed algorithm using an inverted probability map from a classifier, instead of a gradient image.
• Demonstrates applicability to both single-channel and multi-channel image data.

Main Results:

• The classification-driven watershed segmentation algorithm shows superior performance.
• Achieves improved results in image-based granulometry tasks.
• Demonstrates enhanced capabilities for remote sensing image analysis.

Conclusions:

• The proposed method offers a robust and versatile alternative to traditional watershed segmentation.
• Machine-learned pixel classification significantly enhances marker creation and segmentation accuracy.
• The algorithm's direct applicability to diverse image data types broadens its utility.