Tools for recognizing faces

Abstract

Face Image Analysis by Unsupervised Learningby Marian Stewart Bartlett, (Kluwer International Series in Engineering and Computer Science, Vol. 612), Kluwer Academic Publishers, 2001. £66.50 / $95.00 (192 pages) ISBN 0 792 37348 0Face Image Analysis by Unsupervised Learning consists mainly of comparative studies of several popular techniques in the context of face identification and expression analysis. A reproduction of a paper on using neural networks to generate pose-invariant representations of faces is also included. In the rest of the book, neural networks are presented as a tool among many others. The material covered is such that it could have a wide audience, but unfortunately, the style of presentation is so concise that only those who have a background in both neural networks and advanced image processing techniques are likely to understand it without recourse to other materials. However, if the reader is prepared to make the effort, this book provides an up-to-date overview of most of the field and presents some interesting results (Bartlett's own) and conclusions.The book comprises three sections:1.A comparison of Principal Component Analysis (PCA) against Independent Component Analysis (ICA) in the context of facial identification where changes of expression are allowed. PCA is a very popular technique in image processing, whereas ICA is little used. The experiments described here showed that, for this particular application, ICA outperforms PCA.2.A comparison of numerous techniques to identify facial actions from the set defined by the ‘Facial Action Coding System’, a system developed by experimental psychologists to study facial behaviour.3.A description of a neural-network technique to identify people from pictures taken from different views. This method is unusual because it generates a pose-independent representation, that is not a 3-D model of a face. This part consists solely of a reproduced journal paper and seems to be an addendum to the rest of the book.The experiments on facial actions, outlined above, are split into two parts. The first compares PCA against a feature-measurement technique that finds wrinkles, and a simple optic flow algorithm. The results show that both PCA and the optic flow algorithm outperform the feature-measurement technique. However, the best performance is from a hybrid system that combines information from all three sources. The reason given for this is that the errors made by the feature-measurement technique and the other two methods are uncorrelated (the errors made by PCA and optic flow are correlated).The second part compares several more advanced techniques for recognizing individual facial actions: ICA (the only neural-network technique), PCA (both global and local versions), an improved optic flow algorithm, Local Feature Analysis (derived from PCA), Fisher's linear discriminants (a supervised learning technique), and Gabor wavelets. Of these methods ICA and Gabor wavelets gave the best performance, and in fact performed as well as human experts. Optic flow gave intermediate performance between these and all the others. Interestingly, optic flow performed much better with smoothing than without.Note that some popular methods were not tested, such as physical models of the face, elastic matching and methods based on hidden Markov models. Also, there was no registration (alignment) of the faces in consecutive frames to correct for head movements. Bartlett claims that her data was such that this wasn't necessary because people made their facial actions without rigid head movement. However, in a non-experimental situation, this is unlikely to be the case, and registration issues will need to be addressed.Bartlett draws links between PCA, ICA and Gabor filters. Gabor filters were designed to model the behaviour of visual cortical cells. The relationships between these methods places the two most successful techniques firmly in a biological context, something that Bartlett explores throughout her book. The final set of experiments in the book is also placed firmly in a biological context and presented as a method that our brains might in fact use to build pose-invariant representations of the people we can recognize.Generally facial expression analysis is done by matching images to canonical expressions. The approach taken by Bartlett of identifying facial actions is likely to be more flexible, in that it will cope with varying degrees of expression and mixtures of expression. However, an important advantage is that it could prove useful for distinguishing between faked and spontaneous expressions, as people tend not to move all the appropriate muscles in faked expressions. The work presented here is by no means a complete system for analysing expressions. Bartlett's system was trained to identify only 12 out of 46 actions; and experiments were set up so that only one facial action was present in each image. Further work will need to be done to make sure that combinations of facial actions can also be identified.