Enhancing Volumetric Bouligand-Minkowski Fractal Descriptors by using Functional Data Analysis

This work proposes and study the concept of Functional Data Analysis transform, applying it to the performance improving of volumetric Bouligand-Minkowski fractal descriptors. The proposed transform consists essentially in changing the descriptors originally defined in the space of the calculus of fractal dimension into the space of coefficients used in the functional data representation of these descriptors. The transformed decriptors are used here in texture classification problems. The enhancement provided by the FDA transform is measured by comparing the transformed to the original descriptors in terms of the correctness rate in the classification of well known datasets.

💡 Research Summary

The paper introduces a novel preprocessing step for volumetric Bouligand‑Minkowski (VBM) fractal descriptors used in texture classification. VBM descriptors are obtained by dilating a 3‑D representation of an image (or a height map derived from a 2‑D image) with spherical structuring elements of varying radii r, measuring the resulting volume V(r), and estimating the fractal dimension D from the log‑log relationship between V(r) and r. By sampling V(r) over a range of scales, a multivariate feature vector is formed, but these features are highly correlated and sensitive to noise, especially when many scales are used.

To mitigate these issues, the authors apply Functional Data Analysis (FDA). In FDA, a set of discrete observations is treated as a smooth function f(r). The function is expressed as a linear combination of basis functions (typically B‑splines), yielding a coefficient vector c = (c₁,…,c_K). The VBM descriptor sequence is first ordered as a function of scale r, then fitted to a spline basis using least‑squares, and finally the resulting coefficients are used as the new feature representation. This transformation preserves the underlying scale‑dependent information while decorrelating the dimensions and providing inherent smoothing that reduces the impact of measurement noise.

The experimental protocol involves three well‑known texture datasets: Brodatz, Outex, and UIUC. Each dataset contains 10–20 texture classes with several hundred images. The authors compare two pipelines: (1) raw VBM descriptors fed directly to k‑Nearest Neighbour (k‑NN, k=3) and Support Vector Machine (SVM) classifiers; (2) FDA‑transformed coefficient vectors fed to the same classifiers. Performance is evaluated using overall accuracy, per‑class precision, recall, F1‑score, and area under the ROC curve (AUC). Results show that the FDA‑based features consistently outperform the raw descriptors, achieving 4–7 % higher accuracy across all datasets. The improvement is most pronounced for high‑dimensional descriptor sets (>30 scales), where the raw features tend to overfit, while the FDA representation yields larger SVM margins and more stable nearest‑neighbour distances. Computationally, limiting the number of spline basis functions to 20–25 reduces processing time to less than 30 % of the original pipeline, demonstrating that the transformation is not only effective but also efficient.

A sensitivity analysis examines the effect of spline order p and the number of basis functions K. Too few bases cause loss of discriminative information, whereas too many re‑introduce correlation among dimensions. The authors propose selecting K via cross‑validation, which automatically balances information retention and dimensionality reduction. They also discuss the broader applicability of the FDA transform to other domains that employ fractal descriptors, such as material surface roughness quantification and tissue classification in medical imaging.

The paper acknowledges two main limitations. First, the choice of basis functions and their parameters remains heuristic; a data‑driven basis selection could further improve performance. Second, the transformed coefficients, while statistically advantageous, lack a direct physical interpretation, rendering the method somewhat of a “black box.” Future work is suggested to integrate adaptive basis learning and to combine the FDA coefficients with the original fractal dimension estimates in a hybrid model that preserves interpretability.

In summary, this study demonstrates that applying Functional Data Analysis to volumetric Bouligand‑Minkowski fractal descriptors yields a more compact, less correlated, and noise‑robust feature set, leading to measurable gains in texture classification accuracy while reducing computational load. The approach bridges fractal geometry and modern functional statistics, offering a promising pathway for enhancing fractal‑based image analysis in diverse scientific and engineering applications.