Session spéciale traitement de l'information médicale co-organisé avec la SFGBM
P. Dubois, Inserm, Lille, France
Plus de détails »
........................................................
2ème appel à communication
Plus de détails »
........................................................
Ouverture des soumissions
26 Janvier 2011
........................................................
Date limite de soumission
15 Mars 2011 22 Avril 2011
........................................................
Notification d’acceptation
23 Mai 2011 24 Juin 2011
........................................................
Version finale
31 Mai 2011 10 Août 2011
........................................................
Title of the Session: 3D Face Analysis and Recognition
Organizers: Mohamed Daoudi (TELECOM Lille1/LIFL), Liming Chen (Ecole Centrale de Lyon)
Motivation for the Session: Face analysis and recognition has been actively researched in recent years due to a very largenumber of possible applications like biometrics, videosurveillance, advanced human computer interaction or image and video indexing. Various techniques using ideas from 2D image analysis have been presented. Although asignificant progress has been made, the task of automated, robust face recognition is
still adistant goal. 2D Image-based methods are inherently limited by variability in imaging factorssuch as illumination and pose. An emerging solution is to use laser scanners for capturing surfaces of human faces, and use this data in performing face recognition. As the technology for measuring facial surfaces becomes simpler and cheaper, the use of 3Dfacial scans will be increasingly prominent.The three-dimensional face recognition is in a period of rapid expansion.
In this session, we will invite active researchers in this area to showcase latesttechnology advances and to discuss future possibilities.
Paper 1:
Title: 2D Face Recognition Across Pose with 3D-Active Shape Model (3D-ASM) Landmark Detection and 3D Morphable Model (3DMM),
Authors: Dianle Zhou, Dijana Petrovska-Delacrétaz, Berndette Dorizzi, IT-Sudparis, 9 Rue Charles Fourier, Evry, France
Abstract: A big challenge in the field of 2D face recognition is the development of techniques that are robust to variations in pose on face images. The direct solution for bypassing this challenge has been to take face recognition to the 3D domain, since the shape of an object is invariant to changes in illumination or pose. Although the results obtained by 3D face recognition are very encouraging, the majority of the technology already available for security and surveillance works only on 2D face recognition with frontal view images. It is therefore still necessary to develop techniques that help 2D face recognition to cope better with changes in pose.
One of most successful face recognition system across pose is using 3D Morphable Model based on image-based reconstruction and prior knowledge of human faces. The prior knowledge of face shapes and textures is learned from a set of 3D face scans. The morphable model is then fitted into a single face image in an arbitrary pose by iteratively minimising pixel differences of image intensities and reconstructed virtual intensities using the set of parameters controlling the variations of shape, texture, illumination, camera parameters, etc. The process first makes use of several facial landmarks defined on both image and 3D model to find a rough alignment.
The principal components of shape model and texture model were obtained in this processwhich was then used to reconstruct personalised 3D models. In this paper, we follow the framework of the 3D Morphable Model but using 3D Active Shape Modellandmark detector to automatically initialize the model. After 3D face reconstruction, the shape and texture parameters for the targeted subject are extracted,which can be used for face recognition.
For face recognition, different algorithm and information from reconstructed 3D face are used, including:(a) 3D shape-basedICP distancemeasure; (b) Shape and texture coefficients-based comparison; (c) Viewpoint normalization approach.
Experiments on the PIE database showed that the approaches proposed for pose correction improved the performance of a state of the art 2D face recognition algorithm when non frontal images were used on a system trained with near frontal images only. During the experiment, we found that the facial texture is more important for the face recognition and the 3D shapebasedapproach don’t work well. The best method in our experiment is the viewpoint normalization approach, where 3D Morphable Model is used as a preprocessing tool for generating frontal views from non-frontal face images. Those generated images are then fed in the 2D face recognition system who works only on frontal face images. Our results are compared with published results on the PIE database, and show encouraging results.
Paper 2:
Title: Facial Surface Sections for Recognition of 3D Faces with Missing Parts
Authors: Stefano Berretti, Alberto Del Bimbo, Pietro Pala (University of Florence)
Abstract: In this work, we propose and experiment an original solution to 3D face recognition that supports accurate face matching also in cases where just some parts of probe scans are available. In the proposed approach, distinguishing traits of the face are captured by first extractingkeypoints of the 3D depth image and then measuring how the face depth changes along facial curves connecting pairs of keypoints. Face similarity is evaluated by comparing facial curves across inlier pairs of keypoints that match between probe and gallery scans. In doing so, facial curves of the gallery scans are associated with a saliency measure in order to distinguish curves that model characterizing traits of some subjects from curves that are frequently observed in the face of many different subjects. The recognition accuracy of the approach is experimented using the Face Recognition Grand Challenge v2.0 dataset."
Paper 3:
Title: Modèle de fusion de classifieurs basé sur le recuit simulé et son application à la vérification de visages en 3D
Authors: Wael Ben Soltana, Di Huang, Mohsen Ardabilian, Liming Chen, Chokri Ben Amar
Abstract: Classifier fusion is considered as one of the best strategies for improving performancesand overcoming the limitations imposed by a single classification system. It is most used in thebiometrics related domain. Meanwhile, the search for an optimal fusion scheme remainsextraordinarily complex because the cardinality of the space of possible combination strategies isexponentially proportional to the number of competing classifiers. This paper proposes a mixtureof gated experts for the application of 3D face verification using an ensemble of 24 differentscores. The mixture of gated experts is optimized by a Simulated Annealing (SA) based algorithm,and it automatically selects and fuses the most relevant similarity measures. Experimental results of 3D face recognition on the FRGC v2.0 database demonstrate and examine the performance and stability of the proposed method.
Paper 4:
Title: Local 3D shape analysis for facial expression recognition
Authors: Ahmed Maalej, Boulbaba Ben Amor, Mohamed Daoudi (TELECOM Lille1/LIFL)
Abstract: In this paper we propose a novel approach for identity-independent 3D facial expression recognition. Our approach is based on shape analysis of local patches extracted from 3D facial shape model. A Riemannian framework is applied to compute geodesic distances between correspondent patches belonging to different faces of the BU-3DFE database and conveying different expressions. Quantitative measures of similarity are obtained and then used as inputs to several classification methods. Using Multi-boosting and Support Vector Machines (SVM) classifiers, we achieved average recognition rates respectively equal to 98.81% and 97.75%.
