Abstract

Face recognition is a very important task in many applications such as biometric authentication or for content-based indexing photo and video retrieval systems. In recent years, considerable progress has been made on the problems of face detection and recognition, using different methods divided into two groups of geometrical measures and template matching. However, as computation is very expensive and require a great amount of storage for the earlier methods based on correlation, several more recent methods have then been based on principal component analysis, neural network classification and deformable model of templates of features. The first topic of the work reported in this thesis is the experimental evaluation of face recognition methods based on template approaches. Our aim is to test different approaches of template matching: using cross-correlation with Fast Fourier transform, using features obtained from filtering with Gabor wavelet transform or Daubechies wavelet transform, with both rigid grid matching and deformable graph matching. Then, in the second part of the thesis, we propose an implementation of face recognition based on the Daubechies wavelet transform with the matching of series of corresponding graphs while being both speed and storage friendly. The experiments performed on the entire image database of AT&T Laboratories Cambridge show that while the training phase from our proposed face recognition system outperforms in terms of speed other previously described methods such as those based on Fast Fourier transform, Gabor wavelet transform, and even Eigenfaces, its recognition rate is 81% for the entire raw database and even reaches 91% when images are not distorted by strong facial expressions or accessories.