Gabor wavelet networks for object representation

The choice of an object representation is crucial for the effective performance of cognitive tasks such as object recognition, fixation, etc. because how robustly and efficiently vision tasks can be performed depends on the choice of the representation. In this work we introduce Gabor Wavelet Networks as an effective and efficient object representation. Gabor Wavelet Networks represent objects with sets of weighted Gabor wavelets that are specifically chosen to reflect the properties of the represented objects. The degrees of freedom of each Gabor wavelet are allowed to vary continuously. This is in contrast to the wellknown bunch graph approach, also based on Gabor wavelets, where the wavelet parameters are chosen according to a specific discrete scheme that is based on the discrete wavelet transform. The optimized parameter choice of the Gabor Wavelet Networks allows the representation to be very sparse and specific to the represented objects. We will show experimentally that the specificity of the parameters can be exploited for the recognition of faces. Recognition rates are shown to be as high as 97%. The degrees of freedom of wavelets allow any affine deformation that does not involve shearing. Adding shearing to the degrees of freedom, Gabor Wavelet Networks can easily be deformed affinely. This makes tracking applications very easy. Gabor Wavelet Networks represent objects through linear combinations of Gabor wavelets. Changing the dimensionality of the linear combination changes the complexity and precision of the representation. Computations based on the representation also vary in their complexity and precision. Controlling the dimensionality of the linear combinations used in vision tasks allows desired degrees of precision or speed to be achieved. This will be referred to as progressive attention. Affine variability and progressive attention will be tested in an affine real-time face tracking experiment. The scalar weights in the linear combination of wavelets can be computed by applying each Gabor wavelet as a filter. The filter is applied to (projected onto) the image only at the position indicated by the wavelet parameters. The relation between the filter responses and the weights is linear, and the responses contain the same visual information as the weights. Therefore, the optimized Gabor Wavelets of a network can be used not only for representation of an object but also for optimized filtering. We have exploited this in a head-pose estimation experiment. Our experiments have shown that the optimized filtering scheme is superior to a filtering scheme in which the filters are homogeneously distributed. The pose-estimation error was as low as 0:20Æ.