Reconhecimento facial 3D utilizando o Simulated Annealing com as medidas Surface Interpenetration Measure e M-estimator Sample Consensus
Resumo
Resumo: Este trabalho apresenta uma nova abordagem para o problema de reconhecimento facial 3D. Muitos trabalhos presentes na literatura empregam o Iterative Closest Point (ICP) como técnica de alinhamento das faces e, portanto, baseia seus resultados na similaridade provida pelas impróprias medidas Mean Squared Error e Root Mean Squared Error. Este trabalho propõe a utilização do algoritmo estocástico Simulated Annealing baseado em duas medidas robustas, M-estimator Sample Consensus e Surface Interpenetration Measure (SIM), para obter um alinhamento preciso das faces. A conclusão acerca da similaridade das faces é feita utilizando a SIM, que provê uma maior discriminação entre imagens da face de um mesmo indivíduo e de indivíduos diferentes. Neste trabalho utilizou-se diferentes técnicas de alinhamento, como por exemplo o ICP e algoritmo genético com a SIM, para uma definir um guia de referência com a intenção de comprovar a eficácia da abordagem proposta. Os experimentos foram realizados em uma conhecida base de dados de faces 3D disponível no Biometric Experimentation Environment. Além da face como um todo, também foram utilizadas outras regiões da face (como a do nariz) para atingir melhores resultados e garantir a insensibilidade do algoritmo na presença de expressões faciais suaves. Com a abordagem proposta, foram obtidas altas taxas de verificação, que estão além daquelas sugeridas no Face Recognition Grand Challenge 2006 e tida como um dos principais objetivos da competição. Os resultados comprovam a eficácia da abordagem para o problema do reconhecimento facial 3D. Resumo: This work presents a novel approach for the 3D face recognition problem. Many works found in the literature employ the Iterative Closest Point (ICP) as the registration technique for the faces and, therefore, base their results in the similarity provided by the improper measures Mean Squared Error and Root Mean Squared Error. This work proposes the utilization of the Simulated Annealing stochastic algorithm based on two robust measures, M-estimator Sample Consensus and Surface Interpenetration Measure (SIM), to achieve a precise alignment of the faces. The conclusion concerning the similarity of the faces is done using the SIM, which provides a larger discrimination between face images from the same subject and face images from different ones. In this work, different registration techniques were used, like ICP and genetic algorithm using the SIM, to define a baseline in order to verify the efficiency of the proposed approach. The experiments were performed on a well-known 3D face database available on the Biometric Experimentation Environment. Besides the whole face, other regions of the face (like the nose) were also used to achieve a better result and to ensure the insensitivity of the algorithm in the presence of minor facial expressions. With the proposed approach, high verification rate scores were achieved that are beyond the ones suggested by the Face Recognition Grand Challenge 2006 and are considered one of the main goals of the competition. The results prove the efficiency of the proposed approach for the 3D face recognition problem.
Collections
- Teses & Dissertações [10016]