PaperlayerView exact revision on arXiv

Source-linked AI summary

Image processing

Franco Rino

arXiv:1409.2413v1cs.CV

TL;DR

Face-recognition systems need representations robust to varied facial appearance, while existing Gabor-based methods emphasize either GMP magnitude or gradient information. This paper proposes GSF, which describes GMP surfaces using magnitude and first- and second-derivative cues. Experiments on FERET, ORL, and FRGC-1.0.4 show its effectiveness and competitive performance against Gabor- and LBP-based methods.

  • Problem

    Existing Gabor-based face representations use GMP magnitude or gradient information, motivating a more complete description of GMP properties.

  • Method

    GSF treats GMPs as smooth surfaces, extracts magnitude and first- and second-derivative information, binarizes these values, forms joint histograms, and applies EPFDA classification.

  • Results

    Experiments on FERET, ORL, and FRGC-1.0.4 show GSF effectiveness and competitive performance against Gabor- and LBP-based methods.

  • Takeaways & Limitations

    Exploring the properties of Gabor-filtered results beyond magnitude or gradient information can yield a face representation with clearer meaning and strong recognition results.

Abstract

from arXiv · show

Gabor filters can extract multi-orientation and multiscale features from face images. Researchers have designed different ways to use the magnitude of the filtered results for face recognition: Gabor Fisher classifier exploited only the magnitude information of Gabor magnitude pictures (GMPs); Local Gabor Binary Pattern uses only the gradient information. In this paper, we regard GMPs as smooth surfaces. By completely describing the shape of GMPs, we get a face representation method called Gabor Surface Feature (GSF). First, we compute the magnitude, 1st and 2nd derivatives of GMPs, then binarize them and transform them into decimal values. Finally we construct joint histograms and use subspace methods for classification. Experiments on FERET, ORL and FRGC 1.0.4 database show the effectiveness of GSF.

I. INTRODUCTION

Face recognition requires feature extraction robust to illumination, time, pose, and expression. This paper builds on Gabor and LBP-based representations by proposing GSF, which describes GMP shape using magnitude and derivative information.

  • Robust feature extraction is essential because the same face can appear different across illumination, time, pose, and expression conditions.
  • Gabor wavelets show promising face-recognition performance, but their high dimensionality motivates combining them with subspace methods such as Fisherface.
  • LGBP applies LBP after decomposing face images with multi-scale, multi-orientation Gabor filters, producing more robust and discriminative features than earlier methods.
  • The success of LGBP raises whether LBP captures meaningful GMP properties and motivates an operator with clearer physical meaning and better performance.
  • GSF treats GMPs as smooth surfaces and combines magnitude, first-derivative, and second-derivative information into joint histograms before EPFDA classification.
  • Experiments on FERET, ORL, and FRGC-104 report GSF effectiveness, including an impressive recognition rate on FERET.

II. BRIEF REVIEW OF LGBP AND EPFDA

LGBP produces high-dimensional regional histogram representations, so EPFDA reduces dimensionality while preserving regional information for face matching. The classifier uses FDA-transformed regional features and combines their similarity scores.

  • LGBP: LGBP convolves faces with usually 8 orientations and 5 scales, applies LBP8,1 to 40 GMPs, and summarizes regional patterns with local histograms.
  • EPFDA: Quantized LGBP histograms can become very long; with M = 11 and N = 20, the feature vector reaches 70,400 dimensions.
  • EPFDA: EPFDA partitions each LGBP map into regions and sub-regions, computes histograms, and uses FDA matrices to transform the representation into a low-dimensional vector.
  • EPFDA: FDA both reduces histogram-sequence dimensionality and increases discriminative power.
  • EPFDA: Piecewise FDA classifiers fused with a sum rule preserve information from individual face regions.
  • EPFDA: The paper adopts EPFDA after extracting GSF features for final classification.

A. Motivation and Basic Ideas of GSF

GMPs are smooth surfaces rather than ordinary micro-pattern textures, so LBP on GMPs is interpreted primarily as gradient extraction. GSF extends this view by adding magnitude and second-derivative information while avoiding arbitrary code quantization.

  • Motivation: Because Gabor kernels have Gaussian envelopes, GMPs appear as smooth surfaces without obvious micro-patterns.
  • Motivation: LBP applied to GMPs can extract gradient information, with operators representing a peak, a slope in +x, and a valley.
  • Basic Ideas of GSF: GSF aims to summarize GMP properties using magnitude, gradient, and second-derivative information through binarization and joint histograms.
  • Basic Ideas of GSF: The LGBP quantization step is considered arbitrary because decimal-code proximity does not reliably correspond to pattern similarity.

B. Gabor Surface Feature

GSF represents Gabor magnitude pictures as smooth surfaces by encoding their magnitude, first-derivative, and second-derivative information into compact maps and histograms. The method selects derivative combinations according to illumination conditions before EPFDA classification.

  • B. Gabor Surface Feature: GSF treats each Gabor magnitude picture as a smooth surface whose shape includes magnitude, first-derivative, and second-derivative information.The representation binarizes these quantities and combines them into multidimensional decimal-code maps and joint histograms.
  • B. Gabor Surface Feature: The basic GSF code combines binarized magnitude and gradient pictures, producing eight values without requiring quantization.The gradient pictures are obtained from G using symmetric gradient operators, and the binarization threshold is each quantity’s median value.
  • B. Gabor Surface Feature: Second-derivative variants add Gxx, Gyy, and Gxy, but combining all information directly would create too many histogram bins.The paper reports two selected combinations with 16 bins each.
  • B. Gabor Surface Feature: The selected 16-bin formulation using B2 captures magnitude, first-derivative, and second-derivative information, where B2 indicates surface convexity.B2 is the binarization of Gxx+Gyy.
  • B. Gabor Surface Feature: GSF values correspond to particular curve shapes, making the resulting map more meaningful than the LGBP map in the illustrated example.The paper states that GSF performance is also competitive in later experiments.

C. Illumination preprocessing

The paper uses illumination preprocessing to reduce shadow effects and preserve facial details before extracting GSF features. The section illustrates this processing and its role in handling illumination variation.

  • C. Illumination preprocessing: The preprocessing sequence contains gamma correction, difference-of-Gaussian filtering, masking, and contrast equalization, although masking is not used here.The authors state that the selected algorithm removes shadows and preserves facial details.
  • C. Illumination preprocessing: Figure 3 relates each GSF value to the shape of a one-dimensional curve.The paper explains that each value indicates a certain curve shape.
  • C. Illumination preprocessing: Figure 4 compares a GMP with its GSF and LGBP maps.The accompanying text says the GSF map is more meaningful than the LGBP map in this example.
  • C. Illumination preprocessing: Figure 5 contrasts a face before preprocessing with the same face after preprocessing.The cited description attributes the change to shadow removal while preserving facial details.

D. Weighted EPFDA

Weighted EPFDA assigns region-specific weights when computing the final matching score, using recognition rates from region-wise training as the weights. The method is evaluated on FERET, ORL, and FRGC-104.

  • D. Weighted EPFDA: Weighted matching modifies the score so that each facial region contributes according to an assigned weight.The paper motivates weighting by noting that facial regions have different discriminating power.
  • D. Weighted EPFDA: Each region’s recognition rate on a training set is directly used as that region’s weight.The training procedure classifies samples using one of the M*N regions at a time.
  • D. Weighted EPFDA: The resulting regional weights are illustrated in Figure 6.The figure concerns the weights assigned to different face regions.
  • D. Weighted EPFDA: The evaluation uses the FERET, ORL, and Face Recognition Grand Challenge version 1 experiment 1.0.4 databases.The paper abbreviates the last database as FRGC-104.

A. FERET database

On FERET, the paper compares GSF variants, preprocessing, weighting, and real-valued alternatives with existing methods. The results favor derivative-focused GSF without preprocessing and weighted GSF1 with illumination preprocessing.

  • Experimental setup: FERET experiments use 80×64 face crops with M = 10, N = 4, S = 2, and R = 200, yielding an 8000-dimensional subspace.This is much smaller than the 70,400-dimensional LGBP representation reported for.
  • Proposed-method comparisons: Without illumination preprocessing, GSF2 performs better than GSF1, especially on the duplicate I and duplicate II subsets.GSF2 excludes magnitude information, whereas GSF1 includes it.
  • Proposed-method comparisons: Illumination preprocessing improves performance, particularly on the fafc and duplicate subsets, while weighting increases recognition rates slightly.These effects are reported for the proposed FERET methods.
  • Additional comparison: A real-valued feature tested against binarization and joint histograms is evaluated after down-sampling to the original GSF dimension and applying weighted EPFDA.The supplied passages describe the comparison design but not its numerical outcome.
  • Comparison with other methods: Weighted GSF2 outperforms LGBPHS and LGBP+EPFDA except on fafc, while weighted GSF1 with preprocessing achieves the highest scores on fafb, fafc, and duplicate I.Weighted GSF1 with preprocessing ranks second on duplicate II.

B. ORL database

The supplied ORL passage describes a pose-variable, relatively small database and its evaluation setup, while the FRGC-104 passage reports performance under controlled-gallery and uncontrolled-query conditions. The provided text does not include the corresponding table values.

  • ORL database: ORL contains 400 images from 40 subjects, with three images per subject used for training and target images and the remainder used as queries.The database includes pose variance, motivating larger partitioned regions.
  • ORL database: The ORL experiment uses M = 5, N = 4, S = 1, and R = 39 without illumination preprocessing.Recognition rates are listed in Table IV.
  • FRGC-104: FRGC-104 uses gallery images captured under carefully controlled conditions and query images captured in uncontrolled indoor and outdoor settings.This setting makes recognition more challenging.
  • FRGC-104: On FRGC-104, the proposed method outperforms the method in with and without illumination preprocessing.The supplied passage does not provide the recognition-rate values.

V. CONCLUSION

The conclusion presents Gabor Surface Feature as a face representation that models Gabor magnitude pictures as smooth surfaces and extracts magnitude and derivative information. It reports effectiveness across three databases and identifies Gabor phase as future work.

  • Conclusion: GSF treats Gabor magnitude pictures as smooth surfaces and extracts their magnitude, first-derivative, and second-derivative information.The method is described as clear in meaning and easy to implement.
  • Conclusion: Experiments show that GSF surpasses other Gabor- or LBP-based methods when only Gabor magnitude information is used.The conclusion also reports impressive results with illumination preprocessing and weighting on FERET, ORL, and FRGC-1.0.4.
  • Conclusion: The method treats Gabor-filtered results as intermediate features, suggesting that analyzing their properties can motivate new algorithms.This is presented as a perspective for future feature design.
  • Future work: Future work proposes adding Gabor phase information to GSF by analyzing the shape of Gabor phase pictures.The proposed extension is motivated by results in and.
Loading 1409.2413v1…