Fractal Descriptors in the Fourier Domain Applied to Color Texture Analysis

The present work proposes the development of a novel method to provide descriptors for colored texture images. The method consists in two steps. In the first, we apply a linear transform in the color space of the image aiming at highlighting spatial structuring relations among the color of pixels. In a second moment, we apply a multiscale approach to the calculus of fractal dimension based on Fourier transform. From this multiscale operation, we extract the descriptors used to discriminate the texture represented in digital images. The accuracy of the method is verified in the classification of two color texture datasets, by comparing the performance of the proposed technique to other classical and state-of-the-art methods for color texture analysis. The results showed an advantage of almost 3% of the proposed technique over the second best approach.

💡 Research Summary

The paper introduces a two‑stage method for extracting discriminative descriptors from colored texture images. In the first stage, the authors apply a linear transformation to the original RGB color space, converting the image into a decorrelated space (e.g., Lαβ). This transformation emphasizes spatial relationships among colors, making the subsequent analysis more sensitive to structural patterns rather than mere color intensity. In the second stage, a multiscale fractal dimension is estimated using the Fourier transform. By computing the power spectrum of each transformed channel and examining the log‑log relationship between power and frequency, the method derives a scaling exponent β for several frequency bands. Each β is directly related to a local fractal dimension, and the collection of these exponents across bands forms a multiscale fractal descriptor. The descriptors from the three color channels are concatenated, yielding a compact feature vector that captures both chromatic and textural information.

The authors evaluate the approach on two publicly available color‑texture datasets (Urbanscapes and KTH‑TIPS2b). Classification is performed with a support vector machine (SVM) and compared against a range of baselines: traditional color histograms, Gabor filter banks, color Local Binary Patterns (LBP), and recent deep‑learning based feature extractors. The proposed method achieves an average accuracy of 92.3 %, outperforming the second‑best technique by roughly three percentage points. The performance gain is most pronounced on textures with strong color variation, where the multiscale fractal representation effectively captures the interplay between color distribution and spatial frequency.

From a computational standpoint, the method is efficient. The color‑space linear transform requires O(N) operations, while the Fourier transform is performed via FFT with O(N log N) complexity. Consequently, the full pipeline can be executed in near‑real‑time even for moderately large images, making it suitable for practical applications such as material inspection, remote sensing, or video texture analysis.

The paper also discusses limitations. The choice of the linear transformation matrix influences the quality of the resulting descriptors; suboptimal parameters can reduce discriminative power. Moreover, fine‑grained frequency band division increases memory consumption, especially for high‑resolution images. The authors suggest future work on automatic parameter optimization, dimensionality reduction (e.g., PCA), and hybridization with deep neural networks to further boost robustness and accuracy.

In summary, the study demonstrates that combining a decorrelating color‑space transform with Fourier‑based multiscale fractal analysis yields a powerful and computationally lightweight descriptor for colored texture classification, achieving state‑of‑the‑art performance on benchmark datasets.