Scale selective extended local binary pattern for texture classification

Yuting Hu, Zhiling Long, Ghassan AlRegib

TL;DR

The researchers propose SSELBP (Scale Selective Extended Local Binary Pattern), a new texture descriptor that achieves state-of-the-art accuracy in classifying textures across different scales. By combining multi-scale feature extraction with a Gaussian scale space and maximum pooling, SSELBP matches the performance of existing methods while using only one-third of the feature dimensions.

What Problem Does This Paper Address?

Texture classification is a fundamental challenge in computer vision, with applications ranging from biomedical imaging to object recognition. To be effective, a texture descriptor must be robust to changes in lighting (gray-scale invariance), orientation (rotation invariance), and size (scale invariance).

However, there is a constant tug-of-war between robustness and compactness. While many variants of the Local Binary Pattern (LBP) handle rotation and lighting well, scale invariance remains difficult. Previous state-of-the-art methods like SSLBP achieve high accuracy but result in very high-dimensional feature vectors (e.g., 2400 dimensions), which increases computational costs and the risk of overfitting. The authors set out to create a descriptor that is just as robust but significantly more compact.

Methods

The proposed SSELBP framework consists of four main stages:

  1. Multi-scale ELBP Extraction: The authors build upon the Extended Local Binary Pattern (ELBP), which captures three types of information:
  2. CI (Central Intensity): Comparing the center pixel to the image mean.
  3. NI (Neighboring Intensities): Comparing neighbors to their local average.
  4. RD (Radial Difference): Comparing intensities across two different radii.
  5. Scale Space Generation: To handle scale variations, the input image is smoothed using a series of Gaussian filters to create a "scale space" (four scales were used in this study). This helps the algorithm "see" the texture at different levels of detail.
  6. Joint Histograms: For every scale, the algorithm calculates joint histograms of the CI, NI, and RD components. This captures the co-occurrence of different texture features.
  7. Maximum Pooling: This is the "Scale Selective" part. Instead of concatenating all features from all scales, the algorithm performs maximum pooling across the scales. It selects the highest value from corresponding bins across the different scale histograms. This ensures that the most prominent features are captured regardless of the actual scale of the input image.

Datasets

The authors evaluated SSELBP on two standard public texture databases:

  • KTH-TIPS: Contains 10 texture classes. Each class has 81 samples captured at 9 different scales, under different lighting and poses.
  • UMD: A high-resolution database with 25 classes. It includes significant variations in scale, viewpoint, and illumination.

Key Insights and Ideas

  • Maximum Pooling as a Scale Filter: The most elegant idea here is using maximum pooling across the scale space. By picking the "strongest" response for a specific pattern across multiple blurred versions of an image, the descriptor becomes naturally invariant to the scale at which that pattern originally appeared.
  • ELBP is More Efficient than SSLBP: The authors found that the components of ELBP (specifically the radial differences) provide a richer description of texture than standard LBP variants. This richness allows them to achieve high accuracy with fewer total features.
  • Dimensionality Matters: The paper demonstrates that "more" isn't always "better." By carefully selecting radii and using joint histograms, they proved that an 800-dimensional vector can outperform a 2400-dimensional one, provided the features are more representative of the underlying texture structure.

Key Results

The SSELBP descriptor was tested against several state-of-the-art methods, including CLBP, RP, and the high-dimensional SSLBP.

  • Accuracy on KTH-TIPS: SSELBP achieved 98.11% accuracy, outperforming SSLBP (97.80%) and CLBP (97.19%).
  • Accuracy on UMD: SSELBP reached 98.96% accuracy, again slightly edging out SSLBP (98.84%).
  • Efficiency: The most significant result is the reduction in feature size. SSELBP uses a feature dimension of 800, whereas the previous benchmark (SSLBP) required 2400. This makes SSELBP much more suitable for real-time applications or systems with limited memory.

Limitations and Future Work

While SSELBP is highly effective, the authors note a few areas for improvement:

  • Manual Parameter Tuning: Currently, the scale parameters (the $\sigma$ for Gaussian filters) and the specific radii for the ELBP sampling are chosen empirically (manually).
  • Future Automation: The authors suggest that future research should focus on the automatic generation of the scale space and the automatic selection of the optimal sampling schemes to make the system more "plug-and-play" for different types of image data.

Summary updated

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.
BibTeX
@inproceedings{hu2017scale,
  author = {Yuting Hu and Zhiling Long and Ghassan AlRegib},
  title = {{Scale selective extended local binary pattern for texture classification}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2017},
  doi = {10.1109/icassp.2017.7952389},
}

Loading related content...

Figure