Florin Sebastian TELCEAN Fatih KAHRAMAN
Transkript
Florin Sebastian TELCEAN Fatih KAHRAMAN
Illumination Invariant Face Alignment Fatih KAHRAMAN Florin Sebastian TELCEAN Informatics and Mathematical Modelling Technical University of Denmark Email: s053812@student.dtu.dk Informatics and Mathematical Modelling Technical University of Denmark Email: s041386@student.dtu.dk Introduction Method Face recognition systems are typically required to work under highly varying illumination conditions. This leads to complex effects imposed on the acquired face image that pertains little to the actual identity. As face recognition techniques advance, more researchers have focused on challenging issues arising from illumination. Summary of the searching algorithm Warp candidate faces Calculate residue btw. synthetized and restored faces textures initial shape Converge? N Y Fine tuning using Ratio-image normalization and then performing a normal AAM search Restored and aligned face used as initialization for fine tuning module Ratio-image light normalization it is the quotient between a face image whose lighting condition is to be normalized and a reference face image. The method requires the faces to be aligned. It is not a good candidate to be used in the AAM [4,5] searching. In our approach, it is used only for fine tuning alignment. Aim The main idea of this method is that two faces under the same lighting will be similar to each other after blurring. A restored image can be computed from the original one captured under an arbitrary lighting direction, the blurred reference image, and the blurred original image: I restored = I original The aim of this study is to enhance the AAM face alignment accuracy by using some face illumination normalization/correction methods [1,2,3]. Figure 2: Light normalization results using ratio-image method. Top: the input images, Bottom: the normalized images using ratio-image method. Breference Boriginal An iterative procedure is used to obtain a final restored image with frontal illumination. During the iterative procedure, the reference image is updated with the new reconstructed image. The iterative procedure continues until a stopping criterion is met. The initial reference image is the mean face [1]. consists in fitting the histogram of a test face to the histogram of a well-lit face (the mean face of the model) (a) b) Figure 1: Face alignment using standard AAM under good and extreme illumination. a) Normal illumination,(b) Extreme illumination In this project our aim is to enhance the AAM face alignment accuracy by using two different illumination normalization/correction methods and we analyze the ways how can we use them in order to improve AAM face alignment. We constructed a 8-dimensional appearance space to represent 95% of the total variation observed in the combined coefficients. In our experiments, we observed that the Ratio-image restoration method [1] is not suitable for AAM searching. The main problem of the Ratio-image method is that when it is applied to a region of an image that is NOT face-like, the normalization result will have a lot of information of the mean face. Thus the error will be much smaller than the real one, and it will introduce false alarm in the searching process. The histogram based normalization method [2,3] will never change the general aspect of an image. Thus the false alarms are reduced using this normalization method. We choose to start the searching algorithm using histogram based normalization method. Then for this initial result we apply the normalization using Ratioimage method, and search again. This part can be seen as a fine tuning step of the searching algorithm. (b) (c) Figure 3: Example of light normalization using histogram fitting method. (a) mean face, (b) test face, (c) light normalized test face, and their histograms . Conclusions and future works In this study, a novel method is proposed to automatically align a face image from an image captured under arbitrary lighting conditions. The restored image can then be used for face recognition. • The method requires only one image with frontal illumination of each person for training, which means it is very practical. • There is no need to build complex models for illumination. Histogram fitting light normalization The face is split in two windows (left and right); Two mapping functions to the mean face histogram are separately computed for the left and right windows. Face regions that can be covered by hair are avoided. a) proposed AAM alignment and illumination restoration result Proposed AAM Typically, the Active Appearance Models are built using only one training image for each person, taken under frontal illumination; thus the AAM doesn’t model the illumination variations. On the other side, face recognition systems are required to work with illumination conditions much different than the ideal cases. In these situations the AAM based faces alignment method normally fails. proposed AAM alignment result Update model parameters Standard AAM Deform shape: scale, translation, rotation 10 frontal images of 10 different human faces from Yale Face Dataset (the total database consists of 200 face images) used for training the others are used for testing. Each face is annotated with 73 corresponding points. We selected the light directions between +/-35 degrees in the azimuth angle and +/-45 degrees in the elevation angle. input image initial texture iter #=3 iter #=6 Standard AAM This study combines the concept of face restoration approach and Active Appearance Model [4,5] based face alignment and develops illumination invariant AAM method for fine face alignment. Illumination normalization using histogram fitting method Training data The mapping functions are applied by gradually favor one mapping function to another while traveling across the image from left to right. Proposed AAM Varying illumination is one of the most difficult problems and has received much attention in recent years. It is known that image variation due to lighting changes is larger than that due to different personal identity. Because lighting direction changes alter the relative gray scale distribution of face image. Consequently, illumination normalization is required to reach acceptable recognition rates in face recognition systems. Initialize AAM Experimental Results initial shape initial texture iter #=1 iter #=6 • The main advantage of the method consists in the simplicity of light normalization algorithms used. The AAM search method is the same, the normalization is added as a separated module. • The method was evaluated with Yale B Face Database and IMM Face Database. Experimental results show that the performance of the proposed method is independent of the database being used. Currently, the method can be applied to face images under the upright frontal view only. For faces with different poses, further research is necessary to solve the combined effect of the pose and the lighting conditions on face images. References Final Alignment [1] Dang-Hui Liu, Kin-Man Lam, Lan-Sun Shen, Illumination invariant face recognition. Pattern Recognition, 38(10): 1705-1716, 2005. Two simple but efficient illumination normalization techniques are proposed to be integrated in the standard AAM face alignment algorithm, in order to increase its accuracy for different illumination conditions. [2] R.C. Gonzalez and R.E. Woods. Digital Image Processing. Addison-Wesley Publishing Company, 1992. Figure 4: Light normalization results using histogram-fitting method. Top: the input images, Bottom: the normalized images using histogram fitting method. [3] T. Jebara, "3D Pose Estimation and Normalization for Face Recognition", Bachelor's Thesis, McGill Centre for Intelligent Machines, 1996. [4] Cootes, T. F. and Edwards, G. J. and Taylor, C. J., Active Appearance Models, Proc. European Conf. On Computer Vision, Vol. 2, pp. 484-498, 1998. input image initial shape Aligned and restored face (ratio-image) Aligned and restored face (histogram fitting) [5] Mikkel B. Stegmann: The AAM-API: An Open Source Active Appearance Model Implementation. MICCAI (2) 2003: 951-952.