TL;DR
This paper introduces the Robust Local Binary Pattern (RLBP), a novel image descriptor that enhances the robustness of the widely-used Local Binary Pattern (LBP) to image noise. RLBP achieves superior performance in texture classification and face recognition tasks, particularly in the presence of noise, by intelligently re-coding certain LBP patterns that are likely corrupted by noise.
What Problem Does This Paper Address?
The Local Binary Pattern (LBP) is a highly successful and computationally efficient image descriptor widely employed in texture classification, human detection, and face recognition. However, a significant drawback of LBP is its susceptibility to noise. When noise introduces non-monotonic gray-level changes, even minor ones, the LBP descriptor's performance can degrade substantially. This paper aims to address this limitation by developing a more robust version of LBP that maintains its discriminative power even when images are corrupted by noise.
Methods
The core innovation of RLBP lies in its method of handling LBP patterns that are likely affected by noise. The authors identify specific three-bit substrings within the LBP code that are prone to noise-induced changes. For instance, they focus on patterns like y3=(010) and y6=(101). The RLBP algorithm then re-codes these potentially noisy patterns into more stable ones. Specifically, y3 can be re-coded as y'3=(000) and y6 as y'6=(111). This re-coding process is applied to the LBP histogram. Instead of a direct one-to-one mapping, a single LBP bin might map to multiple RLBP bins, allowing for a more nuanced representation that accounts for noise. The authors emphasize that this re-coding is designed to preserve the discriminative power of the descriptor, especially by re-considering non-uniform patterns that might arise from noise. The mapping function for RLBP is pre-computed, ensuring its computational efficiency is comparable to LBP.
Datasets
The paper evaluates RLBP on the following datasets:
- Brodatz: A standard texture dataset used for classification. The authors use a split of 2048 samples with 64 samples per category for 32 texture categories.
- UIUC: Another texture dataset with 25 classes, featuring images with variations in illumination, albedo, and shape. For UIUC, ten samples per class are used for training and the rest for testing.
- FRGC (Face Recognition Grand Challenge) version 2.0: A dataset for face recognition. The authors focus on Experiment 1, which involves 16,028 frontal facial images from 222 subjects, using Fisher Discriminant Analysis (FDA) as the classifier.
The authors also introduce noise (white Gaussian, salt & pepper, and Rayleigh) to the Brodatz and FRGC datasets to specifically test the robustness of RLBP.
Key Insights and Ideas
- Noise is a critical vulnerability for LBP: The paper clearly demonstrates that even small amounts of noise can significantly degrade LBP's performance, particularly when noise causes non-monotonic changes in pixel values.
- Targeted re-coding is effective: Instead of a blanket approach, RLBP's success stems from identifying specific LBP patterns that are most likely to be corrupted by noise and intelligently re-coding them. This targeted approach preserves the descriptor's discriminative power.
- RLBP maintains high uniform pattern ratio: The authors show that RLBP significantly increases the ratio of uniform patterns in the histogram, especially under noisy conditions, suggesting that it effectively "corrects" patterns that would otherwise become non-uniform due to noise.
- Simplicity can be powerful: The core idea of RLBP is conceptually simple – re-coding specific LBP bits – yet it yields substantial improvements in robustness.
Key Results
- Texture Classification (Brodatz): RLBP achieves an accuracy of 98.9% on the original Brodatz dataset and 98.6% with high levels of noise. This significantly outperforms LBP, which drops to 23.1% accuracy with high noise. RLBP also outperforms other established descriptors like SIFT, Gabor, and MR8, especially in noisy conditions.
- Texture Classification (UIUC): RLBP also shows impressive performance on the UIUC texture database, outperforming LBP and other methods.
- Face Recognition (FRGC): On the FRGC dataset, RLBP achieves the best results, outperforming LBP, CLBP, and even Gabor-based methods. With added white Gaussian noise, RLBP demonstrates superior performance, maintaining accuracy at higher noise levels where other methods falter. For instance, at log(1/SNR) = 0.04, RLBP achieves 70% accuracy, compared to 63% for Gabor, 48% for CLBP, and 44% for LBP.
- Computational Efficiency: RLBP is computationally efficient, with feature extraction times comparable to LBP (0.6 ms per image) and significantly faster than Gabor (197 ms per image).
Limitations and Future Work
- Limitations: The paper acknowledges that one potential risk of their method is mapping a natural non-uniform pattern to a uniform pattern. However, they argue this does not significantly impact discrimination as all non-uniform patterns are mapped to a single, less discriminative bin in the uniform LBP histogram.
- Future Work: The authors suggest several interesting avenues for future research:
- Developing a multiscale RLBP to capture features at different resolutions.
- Employing RLBP for depth images, which are often noisy (e.g., from Kinect sensors).
Summary updated