多面的顔空間における強調効果の予測

概要

論文の詳細を見る
A popular method of machine-based recognition codes the expected objects using statistical knowIedge of their shape and sometimes grey-level values, usually expressed in terms of Principal Components Analysis. A natural question is whether such systems behave similarly to humans. The general model of recognition involves a probe to be identified; a representation is selected in order to search a gallery for a match, irrespective of variations. One possible transformation exaggerates the distinctive features of the probe to from caricatures. Human subjects judge modest caricatures (less than about 110%) to be better likenesses than a veridical image and they are recognised with greater accuracy. A discussion of caricaturing can be given in terms of 'face space'. Although often modeled as Euclidean, a manifold model seems more natural as it both emphasizes the local nature of any parameterisation and accommodates transformations under which recognition is invariant. Local co-ordinates for the manifold are chosen to give a paIameterisations or chart centred on an average of similar faces (the ensemble); the chart then codes deviations from this mean. This modeI leads to a machine-based face recognition systezn, similar to practical systems, capable of reproducing the human caricature effect. The search for a chart is crucial. One candidate separates the shape of the face, forming a vector of facial landmarks, and the texture vector obtained by warping the face image to a fixed shape and using a raster scan of the grey-levels. Such a shape-free representation is both voluminous and redundant; a reduction is obtained by restricting to an appropriate sub-manifold, chosen to describe the ensemble by Principal Components Analysis. A shape-free linearisation arises naturally in the computer graphics technique of 'morphing' between faces which can be described as moving along the line segment joining their shape-free representations. The geometric interpretation of the observation that the mean, or mid-point of the morph, between two faces appears more average, or typical, then either ingredient is that the mid-point of a line segment between two points in the chart is (usually) nearer the origin than either end, something which holds provided the distribution 1s not too heavily peaked about the origin. Caricaturing in this representation is simply projecting away from the origin and should remove the increase in typicality. These variations in typicality should be reproduced by the location of the faces within the Principal Components Analysis, both at a global and individual-image level, as should the similarity between the source-images aud the morphs.
社団法人電子情報通信学会の論文
1996-09-18

多面的顔空間における強調効果の予測

スポンサーリンク

概要

著者

関連論文

スポンサーリンク