Face recognition using Histograms of Oriented Gradients

这篇论文的主要内容是将Hog算子应用到人脸识别上。

转载请注明：http://blog.csdn.net/stdcoutzyx/article/details/40757997

1. Main Contribution

• Extract Hog descriptors from a regular grid.
• Fusion of HOG descriptors at different scales allows to capture important structure
• Dimensionality reduction is necessary to make the classification less prone to over-fitting.

2. Feature Conclusion

Features includes geometric or photometric, latter seems to have prevailed in the literature:

• Eigenfaces( Principal Component Analysis)
• Gabor wavelets
• Local Binary Patterns
• Error-Correcting Output Codes
• Independent Component Analysis

3. Improved Hog

Improved Hog for invariance to scale and rotation may be also achieved by extracting descriptors from only salient points (key points) in the scale space following a rotation normalization. The steps involved are:

• Scale-space extrema detection. (achieve scale invariance)
• Orientation assignment. (find the dominant gradient orientation)
• Descriptor extraction.

4. Previous Hog representation for faces

In (Albiol et al., 2008), the authors successfully applied Hog to the face recognition.

• In the 2008 paper, faces are previous normalized in scale and orientation, So the step for scale-space extrema detection were not necessary.
• A set of 25 facial landmarks were localizaed using Active Apperence Models(AAMs).
• Hog descriptors are extracted from the vicinity of each of these 25 landmarks.
• Using nearest neighbor and Enclidean distance to classify.

Drawback of this method:

• Final error may crucially depend on the reliability of the landmark localizations, and the landmarks are not precisely due to occlusions, strong illuminations or pose changes.

5. Improved Method by this paper

• First normalize the face and then extract HOG features from a regular grid. The grid is formed by placing equal side patches around a first cell centered in the image, until the whole image is covered.
• The paper hypothesize that a better result could be obtained by combining information from different patch sizes. And the paper considered a new fusion strategy that is the product combination of the classifiers at patch sizes.
• Several overlapping patches are used, so the final feature representation will be highly redundant, So dimensionality reduction is necessary.

5.1 Detail of the fusion strategy

• R individual classifications c_k (k=1,…,R), each one trained using Hog features with different patch sizes. Each classifier gives one input sample x_k a posterior probability vector:

• The product rule cosists of fusing the final decision as:

6. Experiments

• Effect of the facial feature localization error on the final recognition performance. Large error on the localization of facial features leads to bad classification performance.
  • Evaluate the performance of the method in 4.
  • Calculate the dispersion of coordinates(ellpse fitting) and the total sum of variances of the localized landmarks. Bigger variance, bad performance.
• Extracting regular grids and patch size combination
  • Hog features are processed by PCA and LDA.
  • Nearest classifier with Enclidean and cosine distances is used.
  • Combination of 8×8, 12×12, 16×16, 20×20, 24×24, 28×28 patches.
  • Compare of computation time of the landmark method and the regular grid method.

7. Reference

• [1]. Déniz O, Bueno G, Salido J, et al. Face recognition using histograms of oriented gradients[J]. Pattern Recognition Letters, 2011, 32(12): 1598-1603.