CS 275: Project Proposal
Face Recognition using Space Variant Image Sensors
Team Members 1
Pranav Sodhani - 804591764 2 Atishay Aggarwal Aggarwal - 704758311 704758311 3 Ameya Kabre - 204732347 204732347 4 Shikhar Malhotra - 504741656 5 Pranav Bhat - 704741684
1
2
3
4
5
{ sodhanipranav, atishay5395, akabre, smalhotra, pranavtbhat}@cs.ucla.edu
Introduction It is well acknowledged that Human Visual System (HVS) has evolved exceptionally well when it comes to the task of face recognition. In order to develop artificial vision systems with human-like ability of face recognition, a lot of research has been done since the last 2 decades at the intersection of machine learning and image processing. However, although the goal has been to match human like performance for face recognition tasks, the approaches followed have been quite disparate. Traditional face recognition algorithms assume the existence of an N x M rectangular image and the presence of a Cartesian grid defining the image structure. In such architectures, resolution (density of pixels per unit area) is uniform across the image. On the other hand, most biological vision systems deploy a space variant sampling with a higher concentration of photoreceptors in the fovea while diminishing concentration towards the periphery. This allows such systems to work with significantly lower bandwidth while still maintaining a high acuity of the visual scene at the fovea. Inspired by human visual system, in this project, we would like to adopt a biologically plausible approach towards the problem of face-recognition system and, to that end, propose and implement algorithms for such tasks.
Related Work Face recognition has a wide range of applications in the fields of biometrics (authentication systems), surveillance (person tracking) and entertainment (VR and video games). Given the huge base of stakeholders, research on face recognition has gained a lot of focus over the last few years [1]. A detailed summary of popular algorithms in this domain can be found in [3]. These algorithms can be broadly classified as belonging to 2 classes - shallow algorithms, which either extract features or match templates for calculating a similarity score between faces and deep algorithms, which deploy neural networks for face identification. Eigenfaces using principal component analysis (PCA) [2], one of the most popular algorithms for face recognition, uses KL transform to reduce the dimensionality of images while retaining most discriminating features. A nearest neighbor approach then helps find the closest matching face. PCA based algorithms are, however, not robust to pose and illumination changes. In [7], Vasilescu et al. develop a multi-linear algebraic framework to extend PCA and make it pose and illumination variant. Such algorithms, however, require the presence of huge training sets for better performance. Images may contain higher order statistical data present in their pixels. An improved algorithm, Independent Component Analysis (ICA) was proposed to find such basis images using the principle of optimal information transfer through sigmoidal neurons. ICA tries to minimize both second-order and higher-order dependencies in the input data and finds the basis along which the image data is statistically independent. Bartlett et al. [9] discussed two such architectures. While the first one statistically independent basis images attempts to discover a spatially local basis for the faces in the images, the second one produces a factorial face code. Performance evaluation demonstrates it to be a better technique than PCA. When both the architectures are combined together, more optimal results are obtained. Parkhi et al. [8], extensively describe face recognition techniques using Convolutional Neural Networks or CNNs. The key to developing a good classifier, according to the authors, is to develop a deep network i.e., one having several layers. The face recognition problem is first modeled as a Nway classification problem, where N is the number of distinct people in the dataset. During the learning phase, the CNN develops a score vector, using a fully connected “classifier” layer. The scores are compared to the ground truth classes by computing the softmax log-loss. Then, the fully connected classifier layer is removed, and the generated scores for each testing image are classified by simply picking the class with the minimum Euclidean distance. The performance of the classifier
2|Page
can be improved further by learning embeddings for score vectors. The authors describe this scheme to be “triplet-loss training”. To the best of our knowledge, there has been no work done in the realm of face recognition using space variant sensors. The most similar work in this category assumes log-polar mapping sensors for the task of object detection [5]. With the help of a frontal f ace detection task using PCA, the authors of [5] demonstrate that such systems have the ability of recognizing faces even at a low resolution. The idea of representing images as graphical structures instead of rectangular grids was presented by Wallace et al. in [6]. Leo Grady in his dissertation [4], further extended the idea to develop low level image processing operations such as image segmentation and interpolation. He has also provided MATLAB toolbox for carrying out simple image operations using graph as the data structure for images.
Proposed Work In this project, we propose to develop space-variant counterparts of existing face-recognition algorithms and evaluate them on standard face databases. Our work will lead to the following contributions: ●
Demonstration of potential low level image operations such as convolution in a space-variant context against traditional kernel based convolutions in rectangular grids
●
Implementation of space variant analogs of traditional face-recognition algorithms such as PCA, ICA and deep-learning based algorithms.
●
Performance evaluation and comparison of the developed algorithms on standard face databases such as ORL (AT&T), CMU Pie and Yale Face Database
This work should be a motivation to carry out future work concentrating on making the proposed algorithms more robust to pose and illumination variations.
References [1] [2] [3] [4] [5] [6] [7] [8] [9]
Jafri, Rabia, and Hamid R. Arabnia. "A survey of face recognition techniques." Jips 5.2 (2009): 41-68. M. Turk, A. Pentland, Eigenfaces for Recognition, Journal of Cognitive Neuroscience, Vol. 3, No. 1, 1991, pp. 71-86 W. Zhao, R. Chellappa, A. Rosenfeld, P.J. Phillips, Face Recognition: A Literature Survey, ACM Computing Surveys, 2003, pp. 399-458 Grady, Leo. "Space-variant computer vision: a graph-theoretic approach." Boston University, Boston, MA (2004). Traver, V., et al. "Appearance-based object detection in space-variant images: a multi-model approach." Image Analysis and Recognition (2004): 538-546. Wallace, Richard S., et al. "Space variant image processing." International Journal of Computer Vision 13.1 (1994): 71-90. M. A. O. Vasilescu, D. Terzopoulos, Proceedings of International Conference on Pattern Recognition (ICPR 2002), Vol. 2, Quebec City, Canada, Aug, 2002, 511-514. Parkhi, Omkar M., Andrea Vedaldi, and Andrew Zisserman. "Deep Face Recognition." BMVC . Vol. 1. No. 3. 2015. M. Bartlett J. Movellan T. Sejnowski "Face recognition by independent component analysis" IEEE Trans. Neural Netw. vol. 13 no. 6 pp. 1450-1464 Nov. 2002.
3|Page
Proposal for Paper Presentation (3) 1.
Deep Face Recognition a. Citation: Parkhi, Omkar M., Andrea Vedaldi, and Andrew Zisserman. "Deep Face Recognition." BMVC . Vol. 1. No. 3. 2015. b. URL: http://www.robots.ox.ac.uk:5000/~vgg/publications/2015/Parkhi15/parkhi15.pdf
2.
Fisher Vector Faces a. Citation: Simonyan, Karen, et al. "Fisher Vector Faces in the Wild." BMVC . Vol. 2. No. 3. 2013. b. URL: https://www.robots.ox.ac.uk/~vgg/publications/2013/Simonyan13/extras/simonyan13_ ext.pdf
3.
Tensor Faces a. Citation: Vasilescu, M. Alex O., and Demetri Terzopoulos. "Multilinear image analysis for facial recognition." Pattern Recognition, 2002. Proceedings. 16th International Conference on. Vol. 2. IEEE, 2002. b. URL: https://www.media.mit.edu/~maov/tensorfaces/icpr02.pdf
4|Page