The sophistication of real-world commercial systems should not be underestimated.
Many companies have developed innovative methods of face detection and
registration. More importantly for this paper, they have enhanced their matching
techniques, for example by pre-processing images, selecting and in some cases generating
training data, generating spatially localized features, and optimizing classifiers
for compressed subspaces. Sometimes the data being compressed are not face
images at all, but differences of face images [33], Gabor jets [12], or other high-dimensional
data computed from face images. Face recognition systems also employ
a variety of techniques for selecting subspaces. As a result, it can be difficult to assign
credit (or blame) to a particular component of a face recognition system, even when
the details are not proprietary. The purpose of this paper is to compare the performance
of two subspace projection techniques on face recognition tasks in the context
of a simple baseline system. In particular, we compare principal component analysis
(PCA) to independent component analysis (ICA), as implemented by the InfoMax
[8] and FastICA [21] algorithms.
The sophistication of real-world commercial systems should not be underestimated.Many companies have developed innovative methods of face detection andregistration. More importantly for this paper, they have enhanced their matchingtechniques, for example by pre-processing images, selecting and in some cases generatingtraining data, generating spatially localized features, and optimizing classifiersfor compressed subspaces. Sometimes the data being compressed are not faceimages at all, but differences of face images [33], Gabor jets [12], or other high-dimensionaldata computed from face images. Face recognition systems also employa variety of techniques for selecting subspaces. As a result, it can be difficult to assigncredit (or blame) to a particular component of a face recognition system, even whenthe details are not proprietary. The purpose of this paper is to compare the performanceof two subspace projection techniques on face recognition tasks in the contextof a simple baseline system. In particular, we compare principal component analysis(PCA) to independent component analysis (ICA), as implemented by the InfoMax[8] and FastICA [21] algorithms.
การแปล กรุณารอสักครู่..
