ARISTOTLE UNIVERSITY OF THESSALONIKI DEPARTMENT OF INFORMATICS FACE

ARISTOTLE UNIVERSITY OF THESSALONIKI DEPARTMENT OF INFORMATICS FACE AUTHENTICATION BASED ON MATHEMATICAL MORPHOLOGY A. Tefas, C. Kotropoulos, I. Pitas FRONTIERS OF MATHEMATICAL MORPHOLOGY April 17 -20, 2000, Strasbourg, France Aristotle University of Thessaloniki, Department of Informatics

OUTLINE 4 Introduction 4 Morphological techniques in elastic graph matching 4 Morphological signal decomposition 4 Experimental Results 4 Conclusions Aristotle University of Thessaloniki, Department of Informatics

INTRODUCTION 4 Face recognition has exhibited a tremendous growth for more than two decades. 4 Face verification: “Given a reference facial image or images and a test one, decide whether the test face corresponds to the reference one”. 4 Multi-modal person verification. Aristotle University of Thessaloniki, Department of Informatics

INTRODUCTION 4 Elastic graph matching (EGM) exploits both the gray- level information and shape information. 4 The response of a set of 2 D Gabor filters tuned to different orientations and scales is measured at the grid nodes in EGM. 4 Morphological elastic graph matching (MEGM) and morphological signal decomposition elastic graph matching (MSD-EGM) use the multi-scale morphological dilation-erosion and the morphological signal decomposition instead of Gabor filters. Aristotle University of Thessaloniki, Department of Informatics

MORPHOLOGICAL TECHNIQUES IN ELASTIC GRAPH MATCHING Elastic Graph Matching 4 Local descriptors extracted at the nodes of a sparse grid: 4 Signal Similarity measure: 4 The objective is to minimize the cost function: Aristotle University of Thessaloniki, Department of Informatics

MORPHOLOGICAL TECHNIQUES IN ELASTIC GRAPH MATCHING Definitions Aristotle University of Thessaloniki, Department of Informatics

MORPHOLOGICAL TECHNIQUES IN ELASTIC GRAPH MATCHING Multi-scale dilation-erosion of an image by a structuring function: Feature vector located at a grid node: Aristotle University of Thessaloniki, Department of Informatics

MORPHOLOGICAL TECHNIQUES IN ELASTIC GRAPH MATCHING Suitable structuring functions 4 Scaled hemisphere: 4 Flat: 4 Circular paraboloid: Aristotle University of Thessaloniki, Department of Informatics

MORPHOLOGICAL TECHNIQUES IN ELASTIC GRAPH MATCHING Aristotle University of Thessaloniki, Department of Informatics

MORPHOLOGICAL TECHNIQUES IN ELASTIC GRAPH MATCHING Output of multi-scale dilation-erosion for nine scales. The first nine pictures are dilated images and the remaining nine are eroded images. Aristotle University of Thessaloniki, Department of Informatics

MORPHOLOGICAL TECHNIQUES IN ELASTIC GRAPH MATCHING MORPHOLOGICAL SIGNAL DECOMPOSITION 4 objective: 4 i-th component: and 4 spine: 4 maximal function: 4 first spine: Aristotle University of Thessaloniki, Department of Informatics

MORPHOLOGICAL TECHNIQUES IN ELASTIC GRAPH MATCHING MSD algorithm 4 Step 1: initialization. 4 Step 2: i-th level of decomposition. 4 Step 3: calculate i-th component. 4 Step 4: calculate reconstructed image. Aristotle University of Thessaloniki, Department of Informatics

MORPHOLOGICAL TECHNIQUES IN ELASTIC GRAPH MATCHING 4 Feature vector for MSD-EGM: 4 Reconstructed images at nineteen levels of decomposition Aristotle University of Thessaloniki, Department of Informatics

GRID MATCHING PROCEDURE (a) (b) (c) Grid matching procedure: (a) Model grid for person BP. (b) Best grid for test person BP after elastic graph matching with the model grid. (c) Best grid for test person BS after elastic graph matching with the model grid for person BP. Aristotle University of Thessaloniki, Department of Informatics

DISCRIMINANT ANALYSIS TECHNIQUES 4 Principal component analysis for feature dimension reduction. 4 Linear discriminant analysis for feature selection. 4 Discriminatory power coefficients based on Fisher linear discriminant function for node weighting. 4 Support vector machines for node weighting. Aristotle University of Thessaloniki, Department of Informatics

EXPERIMENTAL RESULTS M 2 VTS database 4 The database contains 37 persons’ video data, which include speech consisting of uttering digits and image sequences or rotated heads. 4 Four recordings (i. e. , shots) of the 37 persons have been collected. 4 Frontal facial images with uniform background were used for the experiments. Experimental protocol 4“Leave one out” principle. 45328 impostor and 5328 client claims. Aristotle University of Thessaloniki, Department of Informatics

EXPERIMENTAL RESULTS Experimental protocol Aristotle University of Thessaloniki, Department of Informatics

EXPERIMENTAL RESULTS Performance evaluation 4 False acceptance (FA) occurs when an impostor claim is accepted. 4 False rejection (FR) occurs when a client claim is rejected. 4 Equal error rate (EER) is the operating state where FA rate=FR rate. 4 Receiver operating characteristics (ROC) is the plot of FA rate versus FR rate. Aristotle University of Thessaloniki, Department of Informatics

EXPERIMENTAL RESULTS Comparison of equal error rates for several authentication techniques in the M 2 VTS database. Aristotle University of Thessaloniki, Department of Informatics

EXPERIMENTAL RESULTS Aristotle University of Thessaloniki, Department of Informatics

PERFORMANCE OF MEGM IN XM 2 VTSdb XM 2 VTS database 4295 persons (8 images person) 4 uniform background Experimental protocol Configuration I 4 Training: 39800 impostor and 200 client claims 4 Evaluation: 40000 impostor and 600 client claims 4 Testing: 112000 impostor and 400 client claims Experimental protocol Configuration II 4 Four training images for each client 4 Evaluation: 40000 impostor and 400 client claims 4 Testing: 112000 impostor and 400 client claims Aristotle University of Thessaloniki, Department of Informatics

PERFORMANCE OF MEGM IN XM 2 VTSdb Receiver Operating Characteristics MEGM Configuration II Aristotle University of Thessaloniki, Department of Informatics

PERFORMANCE OF MEGM IN XM 2 VTSdb Rates at several FAR on XM 2 VTSdb in the two configurations of the experimental protocol. All rates are in %. Aristotle University of Thessaloniki, Department of Informatics

CONCLUSIONS åNovel methods for image analysis into the elastic graph matching have been proposed. åThey are based on multi-scale erosion dilation and morphological signal decomposition of the facial image. åDiscriminant analysis was applied in order to enhance the performance of the proposed methods. åThe experimental results indicated the success of the proposed methods in frontal face authentication. Aristotle University of Thessaloniki, Department of Informatics