Local dominant orientation feature histograms (LDOFH) for face recognition

Xinfang Cui, Peng Zhou, Wankou YangApplied Informatics 2017 · DOI

Abstract

This paper presents a simple and robust feature descriptor method, namely local dominant orientation feature histogram (LDOFH). In this method, the discriminant histogram contains dominant orientation, and the corresponding relative energy value is obtained by calculating the direction and the amplitude of the gradient of each pixel over a local patch. The LM-NNDA (Yang et al. in Pattern Recognit 44:1387–1402, 2011) method based on the principal component analysis (PCA) method is finally employed to reduce the redundancy information and get the low-dimensional and discriminative features. We apply this descriptor on AR and IMM face image databases. Experimental results demonstrate the effectiveness of the proposed LDOFH method.

Introduction

In the field of computer vision, feature extraction is a critical process for transforming complex facial images into mathematical signatures suitable for comparison. While deep learning models currently provide high accuracy, their computational intensity often limits their use on low-power edge devices. This necessitates the development of efficient "hand-crafted" descriptors that offer robustness without high computational costs.

Traditional descriptors, such as Histograms of Oriented Gradients (HOG) or Local Binary Patterns (LBP), often struggle with lighting variations and noise. The previously developed D-HLDO method addressed these issues by using Singular Value Decomposition (SVD) to identify dominant orientations, but the patch-wise application of SVD is computationally expensive. This paper introduces the Local Dominant Orientation Feature Histogram (LDOFH) to provide similar robustness with significantly improved efficiency.

Methodology

The LDOFH pipeline is designed to be computationally lean while maintaining high discriminative power. It consists of five main stages:

  1. Preprocessing: The input image is smoothed using a Gaussian filter to eliminate high-frequency noise that could interfere with gradient calculations.
  2. Direct Gradient Calculation: Unlike methods that rely on SVD, LDOFH calculates the gradient amplitude ($m$) and direction ($\theta$) for every pixel directly using standard calculus. The dominant orientation is defined as the gradient direction, and the "energy" is defined as the amplitude.
  3. Regional Histograms: The image is divided into overlapping rectangular regions. For each region, a 1D histogram is constructed where the height of each orientation bin is weighted by the corresponding gradient energy (amplitude). This ensures that high-contrast structural areas (eyes, nose, mouth) contribute more to the signature than flat areas.
  4. Concatenation: These local histograms are stitched together into a single high-dimensional feature vector.
  5. Dimensionality Reduction: To manage the high dimensionality and improve recognition, the Local Mean-based Nearest Neighbor Discriminant Analysis (LM-NNDA) is applied. This technique reduces the feature vector size while ensuring that faces of the same individual remain close in the mathematical space, while distinct individuals are separated.

Experimental Results

The researchers evaluated the LDOFH method on the AR and IMM face databases, focusing on accuracy and processing speed.

Recognition Accuracy

On the AR database, which includes variations in expression, lighting, and occlusions (e.g., sunglasses and scarves), LDOFH achieved a recognition rate of 98.6%. This performance outperformed the slower D-HLDO (97.8%) and significantly surpassed traditional methods like LBP (95.3%) and PCA (83.3%).

Computational Speed

The headline result of this study is the efficiency gain. LDOFH was found to be nearly three times faster than the D-HLDO method. In batch processing on the AR dataset, LDOFH required approximately 51 seconds, compared to 121 seconds for D-HLDO. This speedup is primarily due to the elimination of the SVD bottleneck.

Discussion and Conclusion

LDOFH demonstrates that statistical rigor, such as that provided by SVD, is not always necessary for identifying dominant orientations in facial features. Standard gradients, when properly smoothed and energy-weighted, provide a sufficient approximation that is much cheaper to compute.

While LDOFH excelled on the AR database, it showed a slightly lower recognition rate than D-HLDO on the IMM database, suggesting that SVD may still hold a slight advantage in scenarios with extremely subtle orientation changes. Future work will explore feature fusion techniques to combine LDOFH with other descriptors, aiming for a hybrid system that maximizes both speed and precision across all datasets.

Summary updated

Abstract

This paper presents a simple and robust feature descriptor method, namely local dominant orientation feature histogram (LDOFH). In this method, the discriminant histogram contains dominant orientation, and the corresponding relative energy value is obtained by calculating the direction and the amplitude of the gradient of each pixel over a local patch. The LM-NNDA (Yang et al. in Pattern Recognit 44:1387–1402, 2011) method based on the principal component analysis (PCA) method is finally employed to reduce the redundancy information and get the low-dimensional and discriminative features. We apply this descriptor on AR, IMM face image databases. Experimental results demonstrate the effectiveness of the proposed LDOFH method.
BibTeX
@inproceedings{cui2017local,
  author = {Xinfang Cui and Peng Zhou and Wankou Yang},
  title = {{Local dominant orientation feature histograms (LDOFH) for face recognition}},
  booktitle = {Applied Informatics},
  year = {2017},
  doi = {10.1186/s40535-017-0043-4},
}
Figure