Source-linked AI summary
Gray Level Co-Occurrence Matrices: Generalisation and Some New Features
Bino Sebastian, A. Unnikrishnan, Kannan Balakrishnan
TL;DR
The paper addresses how to extend GLCM-based texture analysis and obtain additional features for image retrieval. It generalizes co-occurrence matrices to n-dimensional grayscale images and introduces trace from the normalized matrix. Trace features outperform Haralick features in the reported CBIR experiment, while combining trace with Haralick features provides better results.
Problem
GLCM is an established texture-analysis method, but the paper seeks generalized co-occurrence matrices and additional features for CBIR.
Method
The paper extends co-occurrence matrices to n-dimensional grayscale images and defines trace from the normalized co-occurrence matrix.
Results
Trace features outperform Haralick features in CBIR; trace-based retrieval achieves average precision 0.8194 versus 0.7222 for Haralick features.
Takeaways & Limitations
Trace can serve as a texture feature for CBIR, and combining trace with Haralick features provides better results.
Abstract
from arXiv · showhide
Gray Level Co-occurrence Matrices (GLCM) are one of the earliest techniques used for image texture analysis. In this paper we defined a new feature called trace extracted from the GLCM and its implications in texture analysis are discussed in the context of Content Based Image Retrieval (CBIR). The theoretical extension of GLCM to n-dimensional gray scale images are also discussed. The results indicate that trace features outperform Haralick features when applied to CBIR.
1. INTRODUCTION
The paper generalizes GLCM-based texture analysis and introduces additional features for CBIR, motivated by the established importance of texture in identifying image regions.
- 1. INTRODUCTION: GLCM is an early and widely used method for extracting image texture features.Haralick extracted fourteen texture features from GLCMs.
- 1. INTRODUCTION: The paper generalizes co-occurrence matrices to n-dimensional Euclidean spaces.It also extracts new features from the generalized matrix.
- 1. INTRODUCTION: The newly defined features are evaluated for their usefulness in content-based image retrieval.
2. THEORETICAL BACKGROUND
The paper defines traditional and generalized co-occurrence matrices, extends them to n-dimensional images, and introduces trace as a feature derived from normalized GLCMs.
- 2.1. Construction of the Traditional Co-occurrence Matrices: A traditional GLCM records how often intensity i occurs adjacent to intensity j in a grayscale image.Matrices are formed for four adjacency directions, and Haralick features are calculated for each direction.
- 2.1. Construction of the Traditional Co-occurrence Matrices: Texture features are averaged over the four directional co-occurrence matrices.
- 2.2. Generalized Gray Scale Images: An n-dimensional grayscale image is modeled as a function from a subset of Z^n to integer-valued intensities.The model represents each image point as an n-tuple in integer space.
- 2.3. Generalized Co-occurrence Matrices: The generalized co-occurrence matrix counts intensity pairs separated by distance k in a specified direction d.The direction is represented by an n-component vector whose entries can be 0, k, or -k.
- 2.3. Generalized Co-occurrence Matrices: Independent directions determine distinct co-occurrence matrices, while opposite directions produce transpose-related matrices.The paper derives the number of independent directions in n-dimensional integer space.
- 2.4. Trace: Trace is defined as the sum of the main diagonal of the normalized co-occurrence matrix.Diagonal entries represent pairs of pixels with the same intensity, so higher trace indicates more constant image regions.
3. METHODOLOGY
The study evaluates trace-based image retrieval on a Brodatz texture database and compares its results with Haralick features. Trace-based retrieval achieves higher average precision than the Haralick comparison.
- The retrieval database contains 333 Brodatz texture images across 36 classes, with eight images retrieved in each run.
- Trace features use a four-dimensional vector formed by summing four equal parts of the normalized co-occurrence matrix main diagonal.
- 0.8194 average precision is obtained with the trace-based retrieval experiment.
- The figure compares output screenshots for the same query image using trace features on the left and Haralick features on the right.
- 0.7222 average precision is obtained using the Haralick Contrast, Correlation, Energy, and Homogeneity features with a four-dimensional query vector.
4. CONCLUSION
The conclusion presents trace as a promising GLCM feature that outperforms Haralick features in the reported retrieval setting, while identifying broader evaluation as future work.
- Trace can be used as a texture feature and outperforms the Haralick features in the reported content-based image retrieval setting.
- Combining trace with Haralick features provides better results than the individual feature approaches described.
- Only one third of the Brodatz texture database images were used for testing.
- Future work includes testing the complete database, investigating trace from three-dimensional images, and extending the theory to n-dimensional Euclidean space.
Authors
The paper includes biographical information about its authors, covering their education, academic appointments, research interests, and professional activities.
- Bino Sebastian V is an Assistant Professor of Mathematics at Mar Athanasius College and is pursuing a Ph.D. in Computer Applications.
- A. Unnikrishnan is Associate Director of the Naval Physical and Oceanographical Laboratory and works in sonar signal processing, image processing, and soft computing.
- Kannan Balakrishnan is an Associate Professor in Computer Applications at Cochin University of Science & Technology and a co-investigator of an Indo-Slovenian research project.