Ian Craw

Finding Face Features

We describe a computer program which understands a greyscale image of a face well enough to locate individual face features such as eyes and mouth. The program has two distinct components: modules designed to locate particular face features, usually in a restricted area; and the overall control strategy which activates modules on the basis of the current solution state, and assesses and integrates the results of each module.

A SOM Based Approach to Skin Detection with Application in Real Time Systems

A large body of human image processing techniques use skin detection as a first primitive for subsequent feature extraction. Well established methods of colour modelling, such as histograms and Gaussian mixture models have enabled the construction of suitably accurate skin detectors. However such techniques are not ideal for use in adaptive real time environments. We describe methods of skin detection using a Self-Organising Map or SOM, and show performance comparable (94% accuracy on facial images) to conventional techniques. We also introduce the AXEON Learning Processor as the basis for a hardware skin detector, and outline the potential benefits of using this system in a demanding environment, such as filtering Internet traffic, to which conventional techniques are not best suited.

Spatial Content and Spatial Quantisation Effects in Face Recognition

It has recently become apparent that if face images are degraded by spatial quantisation, or block averaging, there is a nonlinear acceleration of the decline in accuracy of recognition as block size increases. This suggests recognition requires a critical minimum range of object spatial frequencies. Two experiments were performed to clarify the phenomenon. In experiment 1, the speed and accuracy of recognition for six frontoparallel photographs of faces were measured. After familiarisation training sessions, the images were shown for 100 ms with 11, 21, and 42 pixels per face, horizontally measured. Transformations calculated to remove the same range of spatial frequencies were performed by means of quantisation, a Fourier low-pass filter, and Gaussian blurring. Although accuracy declined and speed increased in a significant, nonlinear manner in all cases as the image quality was reduced, it did so at a faster rate for the quantised images. In experiment 2, faces rated as being typical were shown at 9, 12, 23, and 45 pixels per face and with appropriate Fourier low-pass versions. The nonlinear decline was confirmed and it was shown that it could not be attributed to a ceiling effect. A further condition allowed quantised and Fourier low-pass conditions to be compared with an unstructured-noise condition of equal strength to that of the quantised images. These gave comparable, but slightly less impaired, recognition than the quantised images. It can be inferred from these results that the removal of a critical range of at least 8–16 cycles per face of information explains the step decline in recognition seen with quantised images. However, the decline found with quantised images is reinforced by internal masking from pixelisation.

How should we represent faces for automatic recognition

We describe results obtained from a testbed used to investigate different codings for automatic face recognition. An eigenface coding of shape-free faces using manually located landmarks was more effective than the corresponding coding of correctly shaped faces. Configuration also proved an effective method of recognition, with rankings given to incorrect matches relatively uncorrelated with those from shape-free faces. Both sets of information combine to improve significantly the performance of either system. The addition of a system, which directly correlated the intensity values of shape-free images, also significantly increased recognition, suggesting extra information was still available. The recognition advantage for shape-free faces reflected and depended upon high-quality representation of the natural facial variation via a disjoint ensemble of shape-free faces; if the ensemble comprised nonfaces, a shape-free disadvantage was induced. Manipulation within the shape-free coding to emphasize distinctive features of the faces, by caricaturing, allowed further increases in performance; this effect was only noticeable when the independent shape-free and configuration coding was used. Taken together, these results strongly support the suggestion that faces should be considered as lying in a high-dimensional manifold, which is locally linearly approximated by these shapes and textures, possibly with a separate system for local features. Principal components analysis is then seen as a convenient tool in this local approximation.

Face Recognition by Computer

We describe a coding scheme to index face images for subsequent retrieval, which seems effective, under some conditions, at coding the faces themselves, rather than particular face images, and uses typically 100 bytes. We report tests searching a pool of 100 faces, using as cue a different image of a face in the pool, taken 10 years later. In two of three tests with different faces, the target face best matches the corresponding cue.

Parameterising Images for Recognition and Reconstruction

We describe a method based on Principal Component Analysis for extracting a small number of parameters from the whole of an image. These parameters can then be used for characterisation, recognition and reconstruction. The method itself is by no means new, and has a number of obvious flaws. In this paper we suggest improvements, based on purely theoretical considerations, in which the image is preprocessed using prior knowledge of the content. The subsequent Principal Component Analysis (PCA) is both theoretically more attractive, and more effective in practice. We present the work in the context of face recognition, but the method has much wider applicability.

Tracking and measuring drivers eyes

We describe a computer vision system for tracking the eyes of a car driver in order to measure the eyelid separation. This measure is used as part of a larger system designed to detect when a car driver is becoming drowsy. The system runs unattended in a car on modest hardware, does not interfere with the drivers normal driving actions, and requires no co-operation from the driver.

Automatic Face Recognition: What Representation?

A testbed for automatic face recognition shows an eigenface coding of shape-free texture, with manually coded landmarks, was more effective than correctly shaped faces, being dependent upon high-quality representation of the facial variation by a shape-free ensemble. Configuration also allowed recognition, these measures combine to improve performance and allowed automatic measurement of the face-shape. Caricaturing further increased performance. Correlation of contours of shapefree images also increased recognition, suggesting extra information was available. A natural model considers faces as in a manifold, linearly approximated by the two factors, with a separate system for local features.

Masking of faces by facial and non-facial stimuli

Abstract A visual masking technique was used as a tool to study the recognition of well-known faces. Famous faces, but unknown masks were used. The visual relation between target and mask was varied; the effect was measured in terms of the thresholds for reports of “familiarity” and the ability to name the target. Three experiments are reported, all of which employed a backward-masking paradigm. The target and mask had equal display times, and there was no interstimulus interval. Significant masking was found with face masks but none from either object or noise masks. Intermediate levels of masking were found from faces that were inverted, jumbled, or had their inner features removed. The third experiment found that face masks that were similar to the targets had an effect equal to dissimilar ones. In all three experiments no interaction was found between the type of mask and the thresholds for the familiarity and naming criteria. The results are discussed in terms of information-processing models of face...

Finding Image Features Using Deformable Templates And Detailed Prior Statistical Knowledge

Much work in image processing has been devoted to generating filters to detect low level image features, e.g. edges, peaks, valleys. Objects are then located or recognised in the image by using the output from these filters.