تقرير نهائي لمشروع بحث Research project final report / Rapport final du projet de recherche 1 4102 برنامج دعم البحوث العلمية Grant Program for Scientific Research in Lebanon – 2014 Programme de subvention à la recherche scientifique au Liban – 2014 مستند إداري Administrative Document -------- Administrative information / المعلومات اإلداري :المرجع Project Title - )عنوان المشروع (عربي وأجنبي التعرف على الوجوه وتقدير التوجه الثالثي األبعاد لها:تحليل الوجوه Face Analysis: Face recognition & 3D pose estimation of faces. Principal Investigator - الباحث الرئيسي طرابلس – القبة العنوان شارع الجيش a.assoum@ul.edu.lb 03/069591 االسم والشهرة Address عمار أسوم ّ العنوان االلكتروني الجامعة اللبنانية e-mail كلية العلوم المؤسسة Institution رقم الهاتف باحث/أستاذ Telephone Name & surname الوظيفة Post Co-investigators - الباحثون المشاركون العنوان االلكتروني e-mail fadi.dornaika@ehu.es youssoftraboulsi@gmail.com abedfg1@gmail.com المؤسسة Institution IKERBASQUE – San Sebastian (Spain) LU – Doctorate School of Science and Technology LU – Faculty of Science االسم والشهرة Name and surname فادي درنيقة يوسف الطرابلسي عبد هللا الغازي 2 4102 برنامج دعم البحوث العلمية Grant Program for Scientific Research in Lebanon – 2014 Programme de subvention à la recherche scientifique au Liban – 2014 .1 Duration and starting date of the research / المدة التعاقدي للمشروع وتاريخ بدء البحث 4 Duration (year) / المدة التعاقدية للمشروع 4104/4/0 Starting date of the research /وتاريخ بدء البحث Scientific Information / العلمي المعلومات.2 ّ Objectives - الهدف (mandatory field to fill 5-8 lines) – ) أسطر8-5 : ( معلومات إلزامية The main goal of the project is to enhance the performance of vision-based applications. Indeed, the proposed research aims to make current vision-based algorithms more reliable and more autonomous. The objective is to enhance the performance of the end-user applications (e.g., video surveillance, identity recognition, and gesture based control). Achievements - أالنجازات المحقق (mandatory field to fill 5-8 lines) – ) أسطر8-5 : ( معلومات إلزامية Improving the performance of Local Discriminant Embedding method (LDE) by adding a parameter-less self-optimization possibility and using a two-phase coding schemes for classification purpose (based on based on classic Regularized Least Square) Applying the improved approach to the problem of model-less 3D head pose estimation Perspectives - آفاق البحث (mandatory field to fill 5-8 lines) – ) أسطر8-5 : ( معلومات إلزامية Preparing a local face database for 3D pose estimation Applying the proposed improvements as well as new paradigms (semi-supervised nonlinear dimensionality reduction and multi-observation and multi-view learning) on the new database and on a public domain free database. 3 4102 برنامج دعم البحوث العلمية Grant Program for Scientific Research in Lebanon – 2014 Programme de subvention à la recherche scientifique au Liban – 2014 Publications & Communications - المنشورات والمساهمات في المؤتمرات Two papers have been published in international conferences Model-less 3D Head Pose Estimation using Self-optimized Local Discriminant Embedding (International Joint Conference on Computer Vision Theory and Applications – VISAPP 2013, Barcelona - Spain) Adaptive Two Phase Sparse Representation Classifier For Face Recognition (Advanced Concepts for Intelligent Vision Systems – ACIVS 2013, Posnan - Poland) Abstract - موجز عن نتائج البحث (mandatory field to fill 5-8 lines) – ) أسطر8-5 : ( معلومات إلزامية A bibliographic study has been performed about the publicly available databases used in the domain of computer vision especially in face recognition and 3D pose estimation. The state of the art about supervised dimensionality reduction that can be used in image classification and 3D face pose estimation has also been done. The Local Discriminant Embedding method (LDE) has been tested then improved by adding a parameter-less selfoptimization possibility and using a two-phase coding schemes for classification purpose. Two papers have been published in two international conferences توقيع الباحث 4 4102 برنامج دعم البحوث العلمية Grant Program for Scientific Research in Lebanon – 2014 Programme de subvention à la recherche scientifique au Liban – 2014 Final report / 1. Principal investigator / Chercheur principal Name and surname / Nom et prénoms Ammar Assoum Institution of affiliation / Institution d'affiliation Lebanese University – Faculty of Science 2. Title of the project as proposed in the original application / Titre du projet tel qu'il a été proposé dans la demande originale (English and French / Anglais et Français) Face Analysis: Face recognition & 3D pose estimation of faces. Analyse des visages: Reconnaissance faciale et estimation de l’orientation 3D des visages 3. Purpose of the project / Objectifs du projet (1page) The automatic face recognition and 3D face pose estimation from visual data such as video sequences and individual images has become the cornerstone of many applications such as Human Computer Interaction (HCI), video surveillance, biometrics, and multimedia. The automatic facial image analysis is a major challenge for researchers and engineers since the goal is to create artificial intelligent systems capable of inferring high level knowledge from raw visual data. The proposed research attempts to enhance the performance of vision-based applications by making current vision-based algorithms more reliable and more autonomous. The objective is to enhance the performance of the end-user applications (e.g., video surveillance, identity recognition, and gesture based control). 4. Expected outputs / Résultats attendus 5 4102 برنامج دعم البحوث العلمية Grant Program for Scientific Research in Lebanon – 2014 Programme de subvention à la recherche scientifique au Liban – 2014 1. Novel and robust Linear Dimensionality Reduction algorithms that are useful for handling face data with high variability. The algorithms are LDE for which the graph and the transformation are simultaneously estimated. 2. Non-linear supervised embedding methods. The proposed method will be named Supervised LE (SLE). 3. Semi-supervised paradigms (linear and non-linear methods) that will be applied for both problems: face recognition and coarse 3D face pose estimation. The main methods that will be upgraded to the semi-supervised case are Average Neighborhood Margin Maximaization (ANMM) and Supervised LE (SLE). 4. Algorithm for providing fine 3D face poses from raw face images without any 3D model. A novel strategy that incorporates face appearance and poses in the similarity matrices. 5. Recognition method that exploits multiple observations on face manifolds. 5. Résultats obtenus / Obtained results 6 4102 برنامج دعم البحوث العلمية Grant Program for Scientific Research in Lebanon – 2014 Programme de subvention à la recherche scientifique au Liban – 2014 5 to 10 pages / 5 à 10 pages Appendices can be added a the end of this document / Des annexes peuvent être ajoutées à la fin de ce document The main tasks that have been achieved during the first period of the project can be summarized as follows: 1. Bibliographic study about the publicly available databases used in the domain of computer vision especially in face recognition and 3D pose estimation. This study resulted in choosing the following databases ORL, Yale, Extended Yale, PIE, Feret, Facepix and Pointing'04. 2. Bibliographic study about supervised dimensionality reduction that can be used in image classification and 3D face pose estimation: several methods have been studied and implemented and tested on some of the above mentioned databases, among them some linear ones such like the classical Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Locality Preserving Projections (LPP) and MultiDimensional Scaling (MDS) as well as non-linear ones such like Locally Linear Embedding (LLE), Local Discriminant Embedding (LDE), Laplacian Eigenmaps (LE) and Isomap. The aim of the study is to investigate the performance of each of the methods (execution time, accuracy, complexity). The accuracy of each of the methods has been measured through the corresponding recognition rate versus the number of dimension retained from the input pattern (image). 3. Manifold learning methods for fine face pose estimation: the technique of manifold learning has recently become more and more attractive in the domain of computer vision and machine learning. This technique is based on an algorithm that may learn an internal model of the data through an N-dimensional graph construction. The obtained model can be used to map points unavailable at training time into the embedding in a process often called out-of-sample extension. All graph-based manifold learning techniques (supervised and unsupervised) depend on several parameters that require manual tuning. In our work we studied a particular case which is the Local Discriminant Embedding (LDE) method and we added some modification to it that made the graph determination algorithm parameter-free. 4. New local database construction: in order to leave a fingerprint of our work for the scientific community and since the number of databases publicly available for 3D pose estimation purposes is very limited we proposed to construct our proper local Lebanese database. A platform has been developed in order to take photos of subject persons (a rotary chair attached to an image/angle acquisition system). Acquisition process is in progress and code needed to process the photos (matching, face detection, cropping) is under preparation. 6. Summary table of expected and obtained results / Tableau récapitulatif des résultats attendus et des résultats obtenus Expected outputs / Résultats attendus Obtained results / Résultats obtenus 7 4102 برنامج دعم البحوث العلمية Grant Program for Scientific Research in Lebanon – 2014 Programme de subvention à la recherche scientifique au Liban – 2014 1. Novel and robust Linear Dimensionality Reduction algorithms that are useful for handling face data with high variability. The algorithms are LDE for which the graph and the transformation are simultaneously estimated. 2. Non-linear supervised embedding methods. The proposed method will be named Supervised LE (SLE). 3. Semi-supervised paradigms (linear and non-linear methods) that will be applied for both problems: face recognition and coarse 3D face pose estimation. The main methods that will be upgraded to the semi-supervised case are Average Neighborhood Margin Maximaization (ANMM) and Supervised LE (SLE). 1. Novel and robust Linear Dimensionality Reduction algorithms that are useful for handling face data with high variability. The algorithms are LDE for which the graph and the transformation are simultaneously estimated. 2. Non-linear supervised embedding methods. The proposed method will be named Supervised LE (SLE). 3. Algorithm for providing fine 3D face poses from raw face images without any 3D model. A novel strategy that incorporates face appearance and poses in the similarity matrices. 4. Algorithm for providing fine 3D face poses from raw face images without any 3D model. A novel strategy that incorporates face appearance and poses in the similarity matrices. 5. Recognition method that exploits multiple observations on face manifolds. 7. Possible encountered difficulties / Difficultés éventuelles rencontrées 8. Scientific publications )articles in peer review journals, books, communications, etc …) / Publications scientifiques (articles dans des revues à comité de lecture, livres, communications, etc …) Attach a copy of each publication as it appeared in the journal) / (Joindre une copie de chaque publication telle qu'elle a paru dans la revue) 9. Oral presentations or posters in national, regional and international conferences / Présentations orales ou affichées à des congrès nationaux, régionaux ou internationaux. 8 4102 برنامج دعم البحوث العلمية Grant Program for Scientific Research in Lebanon – 2014 Programme de subvention à la recherche scientifique au Liban – 2014 (Attach a copy of each presentation as it was presented or published in refereed conference proceedings)/ (Joindre une copie de chaque présentation telle qu'elle a été affichée ou publiée dans les comptes rendus des congrès) [1] F. Dornaika, A. Bosaghzadeh and A. Assoum, Model-less 3D Head Pose Estimation using Self-optimized Local Discriminant Embedding, International Joint Conference on Computer Vision Theory and Applications – VISAPP 2013, Barcelona – Spain, February 2013. [2] F. Dornaika, Y. El Traboulsi and A. Assoum, Adaptive Two Phase Sparse Representation Classifier For Face Recognition, Advanced Concepts for Intelligent Vision Systems – ACIVS 2013, Posnan – Poland, October 2013. Financial report Expenditures made in the first year of project Description Amount in Lebanese Pounds PhD student (Y. El Traboulsi) 6,000,000 LBP Total 6,000,000 LBP 9 4102 برنامج دعم البحوث العلمية Grant Program for Scientific Research in Lebanon – 2014 Programme de subvention à la recherche scientifique au Liban – 2014