Texture Segmentation

This project implements an unsupervised image segmentation pipeline for texture analysis.
The main goal was to explore classical computer vision approaches without relying on large deep learning models.
Instead, the focus is on feature extraction, clustering, and morphological post-processing.

Methodology

1. Preprocessing

   • Input images were taken directly from the Prague Texture Segmentation Datagenerator and Benchmark.

2. Feature Extraction

   • Gabor filters are applied with multiple orientations and frequencies to capture texture patterns.
   • Local entropy maps are computed to describe texture complexity at a local scale.

3. Clustering

   • Extracted features are reshaped into vectors.
   • K-Means clustering is applied to group pixels into texture-based regions.
   • The number of clusters was set to 6 in this experiment.

4. Post-processing

   • Morphological operations (dilation + erosion) are applied to smooth the segmented regions and remove noise.

5. Evaluation

   • Segmentation results were compared against ground-truth masks provided in the benchmark.

Benchmarking

The pipeline was tested on the Prague Texture Segmentation Benchmark:

• Dataset: Grayscale [normal]
• Evaluation: Mean MS score reported by the benchmark server
• Result: 49.24