Texture Image Segmentation Using Gray-Level Co-occurrence Matrix and K-Means Clustering

Texture image segmentation is a classic problem in computer vision, where effectively distinguishing different texture regions poses a key challenge. This article presents a practical approach combining Gray-Level Co-occurrence Matrix (GLCM) feature extraction with K-means clustering.

Core Methodology Constructing Gray-Level Co-occurrence Matrix By calculating the co-occurrence probability of pixel pairs at specific directions and distances, GLCM quantifies image texture characteristics. Commonly used statistical measures include contrast, correlation, energy, and homogeneity, which effectively capture texture properties like roughness and regularity. In code implementation, the GLCM is typically computed using a sliding window approach with parameters for distance (e.g., d=1) and angles (0°, 45°, 90°, 135°).

Feature Vector Construction For each pixel, multi-dimensional GLCM features are extracted from its neighborhood (e.g., 4 directions × 4 statistics), forming high-dimensional feature vectors. Principal Component Analysis (PCA) can be applied for dimensionality reduction to improve computational efficiency in subsequent steps. The feature extraction process can be implemented using functions like graycoprops() in MATLAB or similar libraries in Python.

K-Means Clustering The feature vectors are input into the K-means algorithm, which iteratively optimizes cluster centers based on Euclidean distance. The number of clusters K must be predetermined, often determined using the elbow method or silhouette analysis. The final cluster labels correspond to different texture regions. The algorithm implementation typically involves random initialization of centroids followed by iterative assignment and update steps until convergence.

Implementation Considerations The sliding window size affects GLCM feature stability, requiring a balance between local detail preservation and computational load Feature standardization is crucial to prevent clustering bias caused by scale differences Post-processing techniques (such as morphological operations) can optimize segmentation boundary smoothness. In practice, morphological opening/closing operations help remove small artifacts and smooth region boundaries.

This method performs excellently in industrial inspection and remote sensing image analysis, though it shows sensitivity to noise. Robustness can be further improved by incorporating multi-scale feature fusion techniques.