An Exploration of the Introduction of Warping to the Eigenfaces
Autore
Luca Sorbello - University of Leeds - [1998-99]
Documenti
Abstract
The aim of this work is to provide a robust and reliable model of the face by exploring the benefits of the introduction of warping as a pre-processing technique to improve the result of the eigen faces statistical analysis. We prove the advantages of such an approach both from a theoretical and experimental perspective.
This model is based on all the previous works listed but it has a different approach to the problem of modelling the face space : on the theoretical side we try to eradicate the sources of noise and variance before performing the face analysis rather than trying to minimise them during the process. From the practical point of view
we introduce a complete modularity of the processes therefore we reduce the risk of side effects and correlation that might reduce the reliability of the results. Moreover we can add other pre-processing or post processing techniques to improve the results without affecting what we have already achieved.
The key issue is the eigen face as everything we studied and all this work has been done to improve the performance of this technique but we prove that this technique is a valid and reliable approach to the problem.
We approached both the shape variance and the lighting variance from a slightly different perspective to create a technique that is widely usable and can be adapted to any particular set of data by changing one of the two algorithms involved leaving the other one unaffected.
We have applied the "Douglas Smythe" ( 1990 ) algorithm to warp a training set of 21 colour images as a pre-processing technique of the eigen-faces analysis.
This approach is motivated by the intrinsic nature of the eigen-faces where all the important features of the faces are supposed to be aligned. In real life this is quite unlikely to happen and the result is blurred. But, if we apply a warping technique to all the images of our training set and we warp them to any arbitrary image, within this set, before performing the eigen faces; the images will have exactly the same features in the same spatial coordinates. This should dramatically improve the performance of the eigen technique.
Moreover we have a set of landmarks and offsets that can be used to perform a statistical analysis to describe how these point vary in our training set that describes the spatial variance of the main features of the faces in the training set we use. The applications are various: we can use this program as a platform to achieve image recognition or to create a large number of new faces from the training set or we can use it to add and remove facial expressions, locate a face in an image and to achieve a more "natural" animation
This model is based on all the previous works listed but it has a different approach to the problem of modelling the face space : on the theoretical side we try to eradicate the sources of noise and variance before performing the face analysis rather than trying to minimise them during the process. From the practical point of view
we introduce a complete modularity of the processes therefore we reduce the risk of side effects and correlation that might reduce the reliability of the results. Moreover we can add other pre-processing or post processing techniques to improve the results without affecting what we have already achieved.
The key issue is the eigen face as everything we studied and all this work has been done to improve the performance of this technique but we prove that this technique is a valid and reliable approach to the problem.
We approached both the shape variance and the lighting variance from a slightly different perspective to create a technique that is widely usable and can be adapted to any particular set of data by changing one of the two algorithms involved leaving the other one unaffected.
We have applied the "Douglas Smythe" ( 1990 ) algorithm to warp a training set of 21 colour images as a pre-processing technique of the eigen-faces analysis.
This approach is motivated by the intrinsic nature of the eigen-faces where all the important features of the faces are supposed to be aligned. In real life this is quite unlikely to happen and the result is blurred. But, if we apply a warping technique to all the images of our training set and we warp them to any arbitrary image, within this set, before performing the eigen faces; the images will have exactly the same features in the same spatial coordinates. This should dramatically improve the performance of the eigen technique.
Moreover we have a set of landmarks and offsets that can be used to perform a statistical analysis to describe how these point vary in our training set that describes the spatial variance of the main features of the faces in the training set we use. The applications are various: we can use this program as a platform to achieve image recognition or to create a large number of new faces from the training set or we can use it to add and remove facial expressions, locate a face in an image and to achieve a more "natural" animation
Questa tesi è correlata alle categorie
segnala questa pagina ::.