Abstract
In this paper, we propose a new texture descriptor, scale selective extended local binary pattern (SSELBP), to characterize texture images with scale variations. We first utilize multi-scale extended local binary patterns (ELBP) with rotation-invariant and uniform mappings to capture robust local micro- and macro-features. Then, we build a scale space using Gaussian filters and calculate the histogram of multi-scale ELBPs for the image at each scale. Finally, we select the maximum values from the corresponding bins of multi-scale ELBP histograms at different scales as scale-invariant features. A comprehensive evaluation on public texture databases (KTH-TIPS and UMD) shows that the proposed SSELBP has high accuracy comparable to state-of-the-art texture descriptors on gray-scale-, rotation-, and scale-invariant texture classification but uses only one-third of the feature dimension.
Introduction
Texture classification is a fundamental challenge in computer vision, essential for applications ranging from biomedical imaging to object recognition. An effective texture descriptor must maintain robustness across variations in lighting (gray-scale invariance), orientation (rotation invariance), and size (scale invariance).
However, researchers often face a trade-off between robustness and compactness. While many variants of the Local Binary Pattern (LBP) handle rotation and lighting effectively, achieving scale invariance typically results in high-dimensional feature vectors. For instance, the state-of-the-art SSLBP method achieves high accuracy but requires 2400 dimensions, increasing computational costs and the risk of overfitting. SSELBP was developed to bridge this gap, offering a robust descriptor that is significantly more compact.
The SSELBP Framework
The SSELBP framework is designed to capture texture features across multiple levels of detail while maintaining a manageable feature size. The process consists of four primary stages:
- Multi-scale ELBP Extraction: The method utilizes Extended Local Binary Pattern (ELBP), which captures three distinct types of information: Central Intensity (CI), Neighboring Intensities (NI), and Radial Difference (RD). This provides a richer description than standard LBP.
- Scale Space Generation: To address scale variations, the input image is smoothed using a series of Gaussian filters to create a scale space (typically using four scales, $s_1$ to $s_4$).
- Joint Histograms: For each scale, joint histograms of the CI, NI, and RD components are calculated to capture the co-occurrence of texture features.
- Maximum Pooling: This "Scale Selective" step is the core innovation. Instead of simple concatenation, the algorithm performs maximum pooling across the scale space, selecting the highest value from corresponding bins across different scale histograms. This ensures the most prominent features are captured regardless of the input image's scale.
Experimental Results
The SSELBP descriptor was evaluated against several state-of-the-art methods, including CLBP, RP, and the high-dimensional SSLBP, using two standard public texture databases: KTH-TIPS (10 classes with scale and lighting variations) and UMD (25 high-resolution classes with significant viewpoint changes).
- KTH-TIPS: SSELBP achieved a maximum accuracy of 98.11%, outperforming SSLBP (97.80%) and CLBP (97.19%).
- UMD: SSELBP reached 98.96% accuracy, slightly edging out SSLBP (98.84%).
As shown in the performance analysis, the optimal configuration for SSELBP uses four radii ($N=4$), which yields the peak accuracy of 98.11% with a feature dimension of only 800. This is exactly one-third of the 2400 dimensions required by the previous benchmark, SSLBP.
Discussion and Conclusion
The success of SSELBP highlights two key insights: maximum pooling serves as an effective scale filter, and the radial differences in ELBP provide a more efficient texture description than standard LBP variants. By carefully selecting radii and using joint histograms, the researchers proved that an 800-dimensional vector can outperform much larger vectors if the features are more representative of the underlying texture structure.
While highly effective, current limitations include the manual tuning of Gaussian filter parameters and ELBP sampling radii. Future work will focus on automating these selections to create a more adaptive, "plug-and-play" system for diverse image datasets.
Summary updated