1 Introduction

  Personal identification (authentication, verification of identity) is an important issue in many security applications. In this paper we focus on personal identification from frontal face images. Identification is closely related to recognition, but differs in at least three fundamental aspects. Firstly, a client - an authorised user of a personal identification system - is assumed to be co-operative and makes an identity claim. Computationally this means that it is not necessary to consult the complete set of models (reference images in our case) in order to verify a claim. A test image is thus compared to a small number of reference images of the person whose identity is claimed and not, as in the recognition scenario, with every image (or some descriptor of an image) in a potentially large database. Secondly, an automatic authentication system must operate in near-real time to be acceptable to users. And finally, in recognition experiments only images of people from the training database are presented to the system, whereas the case of an imposter (most likely a previously unseen person) is of utmost importance for authentication.

In this paper we propose an identification method based on optimised robust correlation. We show that in the context of the identification task it offers some advantages over standard face recognition approaches, eg. the dynamic link architecture [ 11 ] and methods based on principal component analysis [ 12 , 17 , 13 ]. High recognition rates for methods based on correlation of grey-level distributions in selected areas of the face have been reported [ 4 , 5 ]. But since direct correlation is sensitive to changes in scale, rotation, and illumination conditions, these methods have to rely on pre-normalisation and pre-segmentation. The segmentation and the normalisation process typically depends on detectors of facial features. With this strategy, recognition performance depends critically on the reliability of the detector, because a failure of the detector almost certainly implies recognition failure.

In the method proposed in this paper we avoid this dependence by using an integrated approach, where localisation, normalisation (geometric and photometric) as well as identification is achieved simultaneously. To that end, a robust form of correlation is evaluated inside an optimisation loop. In the optimisation, we search the space of all affine transformations between the test image and reference images augmented to the space of all linear mappings between their corresponding grey-level values. Such direct approach clearly must evaluate hundreds of correlations per verification. This seems to be extremely computationally inefficient and therefore bound to fail the near real-time requirement of practical identification systems. However, this is not so. By evaluating the correlation in a Monte-Carlo fashion, ie. by estimating the correlation from a small sample of suitably chosen points, we are able to speed up the evaluation of the cost function inside the optimisation loop. In our experiments, it was sufficient to take a sample of 2-4% of pixels to converge to essentially the same solution that would have been obtained had a full image correlation been evaluated. The complete minimisation process that simultaneously achieves intensity normalisation, registration of the test and reference images and detection of outliers (occluded parts of the face, hair and beard changes) terminates in the time it would take to evaluate approximately ten full correlations (ie. computed using every pixel) of the face image. For the image resolution used in our experiments (appr. 280 350), the optimisation takes on average only a fraction of a second (see Section 3.2 ). Comparing this with the principle component approach, we see that the computational effort involved is similar to computing about ten projections onto eigenvectors.

The framework has a number of attractive features. It does neither require pre-registration nor does it depend on the success of a face detector (localisation). If some a priori knowledge is available, eg. estimates of scale or head positions, the optimisation process can take advantage by starting close to the optimum in the search space and thus run faster. Unlike the eigen-face methods, no manual model-building is necessary. Normally, a video stream, rather than a single image, is available to the identification system. The speed of the robust optimised correlation allows to repeat the identification process and thus achieve higher reliability.

The rest of the paper is organised as follows. In Section  2 details of the formulation of the optimisation problem are given, including the definition of the search space and the cost function (Section  2.1 ), the description of the search algorithm (Section  2.2 ) and the randomised sampling technique used for fast evaluation of correlation (Section  2.3 ). Experiments on recognition performance and run-time efficiency are presented in Sections  3 and, finally, results are summarised in Section  4 .