Graphical Image Region Extraction with K-Means Clustering and Watershed

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Image segmentation is crucial for automatic visual systems, but challenges persist with color images due to variations in color, texture, and shape.
• Segmenting trademark graphic images is particularly difficult due to high noise, low resolution, and heterogeneous elements like overlapping objects and varying lighting.

Purpose of the Study:

• To propose a generalized hybrid approach for Image Region Extraction (IRE) tailored for highly variable trademark graphic images.
• To evaluate the effectiveness of combining K-Means Clustering and Watershedding algorithms in a hybrid environment for automated region proposal and segmentation.

Main Methods:

• A multi-stage algorithm processes RGB images, starting with K-Means Clustering to generate a segmented grayscale image.
• Preprocessing includes thresholding for a binary mask, distance map generation, and Watershedding using markers from Connected Component Analysis.
• Object extraction is finalized using a contour-based border following method.

Main Results:

• The proposed hybrid system demonstrates adequate region extraction capabilities across diverse graphical image datasets.
• It successfully distinguishes relevant visual elements in images characterized by noise, low resolution, and significant variations.
• The system requires minimal tweaking to achieve satisfactory segmentation results.

Conclusions:

• The developed generalized implementation effectively addresses the challenges of segmenting heterogeneous trademark graphic images.
• The hybrid K-Means and Watershedding approach offers a robust solution for automated region proposal and segmentation in complex visual data.
• This method provides a versatile tool for image region extraction applicable to a wide range of scenarios.