Identification of faces by multimodal information fusion of depth and color
Résumé
Recognition of traditional optical faces of color images or intensity presents many challenges, such as variations in lighting, pose and expression. In fact, the human face not only generates 2D texture information, but also 3D shape information. In this paper, we examine what information the contributions of depth and color to make facial recognition when the variation in lighting and expression are taken into account. We present three methods of feature extraction based on reduction of one-dimensional space: the Linear Discriminant Analysis (LDA), Enhanced Fisher Linear Discriminant Model (EFM) and the Direct LDA (DLDA). A theoretical presentation of these approaches and their applications on the depth images and color is made. It is also a comparative study on information fusion of depth and color for both levels: characteristics and scores to select the most effective features and robust and thus build a strong classifier. The concatenation of feature vectors and fusion of the pixels of the image depth and color: the average, the product, the minimum and maximum are used in the case of fusion characteristics. For the merger to level scores, we used the fuzzy Sugeno integral and Choquet and support vector machines (SVM). The experiments are performed on the database CASIA 3D Face, complex data sets with variations, including variations in lighting, expression and longstanding failures between two scans. The experimental results show the promising performance of the proposed system. Note that in our system, all processes are performed automatically.