Evaluations of both academic face recognition algorithms and commercial systems have shown that the recognition performance degrades significantly due to the variation of illumination. Previous methods for illumination robust face recognition usually involve computationally expensive 3D model transformations or optimization base reconstruction using multiple gallery face images, making them infeasible in practical large scale face identification applications. In this paper, we propose an alternative face identification framework, in which one image per person is used for enrollment as is commonly practiced in real life applications. Several probe images captured under different illumination conditions are synthesized to imitate the illumination condition of the enrolled gallery face image. We assume Lambertian reflectance of human faces and use the harmonic representations of lighting. We demonstrate satisfactory performance on the Yale B database, both visually and quantitatively. The proposed method is of very low complexity when linear facial feature are used, and is therefore scalable for large scale applications.
Proc. SPIE. 8005, MIPPR 2011: Parallel Processing of Images and Optimization and Medical Imaging Processing
KEYWORDS: Image processing algorithms and systems, Signal to noise ratio, Lithium, Magnetic resonance imaging, Image segmentation, Image processing, Computing systems, Medical imaging, Numerical analysis, Data processing
Medical images have the characteristics of high noise and blurred edges, which makes them difficult to segment using
traditional segmentation methods. The level set algorithm, which is a commonly used method for medical image
segmentation, is restricted in use mainly due to the extremely intensive computation during the iterative contour
evolution. The paper proposes some criteria of loop iteration break for the level set algorithm, making it possible to
adaptively adjust the number of iterations to the specific characteristics of various medical images, so that the contour
evolution can be terminated appropriately. Meanwhile, we change the step length of the iteration according to the
previous loop iteration result, making it possible to decrease the number of iterations needed. To decrease the
computational workload, we also restrict the iteration to a certain part of the image instead of the whole image.
Conference Committee Involvement (1)
Parallel Processing for Imaging Applications II
23 January 2012 | Burlingame, California, United States