Interest Operator versus Gabor filtering for facial imagery classification

Previous work has shown that Gabor feature extraction is one of the most effective techniques employed for the human face recognition problem. However, the selection of a particular set of Gabor filters is often problematic and, also the computational requirements are considerable. We propose an alternative feature extraction method – the Interest Operator – to be applied for the facial recognition problem. This method has already been successfully used in the mobile robots navigation, stereoscopic vision and automatic target recognition. Experimental results presented in this paper indicate that classifiers, both neural (Multi-layer Perceptron) and statistical (k Nearest Neighbour), using the Interest Operator – based feature extraction, are capable to achieve almost the same classification rate as the Gabor-wavelet-based methods but in one order of magnitude lower processing time. A special care has been put on the selection of the feature extraction filters and classifiers parameters. Then, on AT&T public facial database, the system has achieved an average recognition rate of 95.2% using Gabor Approach and 94.7% using the Interest Operator.