1. An Introduction to Biometrics
Anil K. Jain, Michigan State University
Arun Ross, West Virginia University
Karthik Nandakumar, Institute for Infocomm Research
Biometrics refers to the automatic identification (or verification) of an individual (or a claimed identity) by using certain physical or behavioral traits associated with the person. By using biometrics it is possible to establish an identity based on `who you are', rather than by `what you possess' (e.g., an ID card) or `what you remember' (e.g., a password). Therefore, biometric systems use fingerprints, hand geometry, iris, retina, face, vasculature patterns, signature, gait, palmprint, or voiceprint to determine a person's identity.
The purpose of this tutorial is two-fold: (a) to introduce the fundamentals of biometric technology from a pattern recognition and signal processing perspective by discussing some of the prominent techniques used in the field; and (b) to convey the recent advances made in this field especially in the context of security, privacy and forensics. To this end, the design of a biometric system will be discussed from the viewpoint of four commonly used biometric modalities - fingerprint, face, hand, and iris. Various algorithms that have been developed for processing these modalities will be presented. Methods to protect the biometric templates of enrolled users will also be outlined. In particular, the possibility of performing biometric matching in the cryptographic domain will be discussed. The tutorial will also introduce concepts in biometric fusion (i.e., multibiometrics) in which multiple sources of biometric information are consolidated. Finally, there will be a discussion on some of the challenges encountered by biometric systems when operating in a real-world environment and some of the methods used to address these challenges.
What is Biometrics
Back to top
Designing a Biometric System
Application in Forensics
Securing Biometric Templates
Challenges in Biometrics
2. Image Computing for Digital Pathology
Shishir Shah, University of Houston
Pathologists and cancer biologists rely on tissue and cellular analysis to study cancer expression, genetic profiles, and cellular morphology to understand the underlying basis for a disease and to grade the level of disease progression. Conventional analysis of tissue histology and sample cytology includes the steps of examination of the stained tissue or cell smear under a microscope, scoring the expression relative to the most highly expressing (densely stained) area on a predefined scale for normal, cancer, stromal regions based on the morphology of the tissue, estimating the percentage area of cancer tissue relative of normal and stroma, and multiplying the score by the percentage area of cancer region and converting to another predefined scale for statistical analyses. Most of this analysis is done manually or with limited tools to aid the scoring process. Over the last 5 years, automated and semi-automated microscope slide scanners have become available in the marketplace. These scanners rely on sophisticated microscopes and allow for the digitization of the entire sample at varying magnifications. This has led to the emergence of digital pathology and a growing amount of image data. Each sample digitized is typically of the order of 2.7GB to 10GB in size depending on the magnification of the digitizing system with an image size of 30,000 x 30,000 pixels or larger. Further, current software and methods for automated scoring of tissue is very limited. This has led to an increased interest in identifying novel solutions to automated histology and cytology analysis.
In order to achieve high computational accuracy with reasonable turnaround times, novel approaches from the data and resource management perspective are also required to address handling of image sizes outlined above. Two developments in computer industry make the current generation of scientists more likely to solve the performance challenges associated with the large image sizes. First, the emergence of multi-core processors allow for parallel processing within a single PC using of-the-shelf components. It is virtually impossible today to buy a PC that does not have at least two computational cores. Furthermore, all major manufacturers have announced processors containing four, eight and even sixteen cores for the next two years, providing an omnipresent potential for parallel processing on every desktop PC. Second, as of today, there is no reasonable size medical or research institution in the US not having a PC cluster in their computing arsenal. The next generation of digital pathology systems will require the ability to share and process data across disparate institutes. Hence, PC clusters would allow for parallelism by analyzing multiple images simultaneously. They would also offer an opportunity to speed up the analysis of a single image. However, exploiting the computational power of multi-core architectures and PC clusters requires modifications to existing, sequential image analysis codes and cautious evaluation of alternative and novel algorithms with respect to their potential for parallelism.
This tutorial will provide an overview of the application domain and present an overview of the challenges. Specifically, opportunities for novel image analysis and pattern recognition algorithms that can leverage frameworks for shared and distributed parallel computing will be discussed along with examples from ongoing research in the labs of the instructors. Topics to be covered will include:
Overview of digital pathology
Back to top
Applications and Challenges (Immunohistochemistry, H&E analysis, FISH, Alternate Image Modalities – Spectral Imaging, Histology, Cytology)
Architectural Developments (Multi-core, Networking, GPU processing, Storage)
Image Analysis Pipelines and Algorithms (Image Segmentation, Nuclei Detection, Morphometric Features, Karyometric Features, …)
Performance implications (data management, exploiting multi-core processors, exploiting PC cluster)
Emerging trends and applications
3. Similarity Searching: Indexing, Nearest Neighbor Finding, Dimensionality Reduction, and Embedding Methods for Applications in Multimedia Databases
Hanan Samet, University of Maryland (see bio)
Similarity searching is usually achieved by means of nearest neighbor finding. Existing methods for handling similarity search in this setting fall into one of two classes. The first is based on mapping to a low-dimensional vector space which is then indexed using representations such as k-d trees, R-trees, quadtrees, etc. The second directly indexes the the objects based on distances using representations such as the vp-tree, M-tree, etc. Mapping from a high-dimensional space into a low-dimensional space is known as dimensionality reduction and is achieved using SVD, DFT, etc. At times, when we just have distance information, the data objects are embedded in a vector space so that the distances of the embedded objects as measured by the distance metric in the embedding space approximate the actual distance. The search in the embedding space uses conventional indexing methods which are often coupled with dimensionality reduction. Some commonly known embedding methods are multidimensional scaling, Lipschitz embeddings, and FastMap. This tutorial is organized into five parts that cover the five basic concepts outlined above: indexing low and high dimensional spaces, distance-based indexing, dimensionality reduction, embedding methods, and nearest neighbor searching.
Back to top
8:30-10:00am First Session
10:00-10:30am Coffee Break
10:30-12:00pm Second Session
4. Semantic Indexing and Retrieval of Video
Marcel Worring and Cees Snoek, University of Amsterdam
The semantic gap between the low level information that can be derived from the visual data and the conceptual view the user has of the same data is a major bottleneck in video retrieval systems. It has dictated that solutions to image and video indexing could only be applied in narrow domains using specific concept detectors, e.g., “sunset” or “face”. This leads to lexica of at most 10-20 concepts. The use of multimodal indexing, advances in machine learning, and the availability of some large, annotated information sources, e.g., the TRECVID benchmark, has paved the way to increase lexicon size by orders of magnitude (now 100 concepts, in a few years 1,000). This brings it within reach of research in ontology engineering, i.e. creating and maintaining large, typically 10,000+ structured sets of shared concepts.
When this goal is reached we could search for videos in our home collection or on the web based on their semantic content, we could develop semantic video editing tools, or develop tools that monitor various video sources and trigger alerts based on semantic events.
This tutorial lays the foundation for these exciting new horizons. It will cover basis video analysis techniques and explain the different methods for video indexing. From there it will explore how users can be given interactive access to the data. For both indexing and interactive access TRECVID evaluations will be considered.
Basic video analysis
Back to top
Concept-based Video Indexing
Large-scale concept detection
Concept-based Video Search
Concept detector selection
Concept detector combination
Learning from the user
5. The Algebraic Approaches and Techniques in Image Analysis
Igor Gurevich and V. Yashina
Automation of image processing, analysis, estimating and understanding is one of the crucial points of theoretical computer science having decisive importance for applications, in particular, for diversification of solvable problem types and for increasing the efficiency of problem solving.
Automation of image mining is one of the most important strategic goals in image analysis, recognition and understanding science and technologies. The main subgoals are developing and applying of mathematical theory for constructing image models accepted by efficient pattern recognition algorithms and for standardized representation and selection of image analysis transforms. Automation of image-mining is possible by combined application techniques for image analysis, understanding and recognition.
The specificity, complexity and difficulties of image analysis and estimation (IAE) problems stem from necessity to achieve some balance between such highly contradictory factors as goals and tasks of a problem solving, the nature of visual perception, ways and means of an image acquisition, formation, reproduction and rendering, and mathematical, computational and technological means allowable for the IAE.
The mathematical theory of image analysis is not finished and is passing through a developing stage. It is only recently came understanding of the fact that only intensive creating of comprehensive mathematical theory of image analysis and recognition (in addition to the mathematical theory of pattern recognition) could bring a real opportunity to solve efficiently application problems via extracting from images the information necessary for intellectual decision making. The transition to practical, reliable and efficient automation of image-mining is directly dependent on introducing and developing of mathematical means for IAE.
During recent years there was accepted that algebraic techniques, in particular different kinds of image algebras, is the most prospective direction of construction of the mathematical theory of image analysis and of development of an universal algebraic language for representing image analysis transforms and image models.
State of the art mathematical theory of image analysis
Back to top
The Algebraic Approaches and Techniques in Image Analysis
Descriptive Approach to Image Analysis and Understanding (DAIA) and its main tools
6. Image Retrieval
Henning Mueller and Thomas Deselaers
Image retrieval has been receiving increasing interest due to the vast amount of images publicly available on the Internet. Most image sharing sites, such as FlickR, allow for text/tag-based image searching. In the research community, content-based image retrieval has been under investigation since the early 1990s. The tutorial will give show approaches to fusing the efforts from content-based image retrieval with the available text-based image searching solutions, including multi-lingual approaches.
Therefore, a short introduction into content-based image retrieval and text-retrieval will be given. Different approaches to fusing these will be discussed and further topics such as user-interaction, retrieval system architecture, and benchmarking issues will be presented.
The tutorial will focus on image retrieval from different perspectives.
Content-based image retrieval, i.e. finding images by their visual content
Back to top
Text-based image retrieval, i.e. finding images using textual information
Combination of the above
7. Component Analysis Methods for Pattern Recognition
Fernando De la Torre
Component Analysis (CA) methods (e.g. Kernel Principal Component
Linear Discriminant Analysis, Spectral Clustering) have been extensively
used as a feature extraction step for modeling, classification and
clustering in numerous pattern recognition, signal processing or social
sciences tasks. The aim of CA is to decompose a signal into relevant
components that explicitly or implicitly (e.g. kernel methods) define
representation of the signal. CA techniques are especially appealing
many can be formulated as eigen-problems, offering great potential for
efficient learning of linear and non-linear representations of the data
without local minima. Although CA methods have been widely used, there
still a need for a better mathematical framework to analyze and extend
techniques. This tutorial reviews previous work and proposes a unified
framework for energy-based learning in CA methods.
The first part of the tutorial will review traditional linear techniques
such as principal Component Analysis (PCA), Oriented Component Analysis
(OCA), Linear Discriminant Analysis (LDA), Canonical Correlation
(CCA), NMF (Non-Negative Matrix Factorization), Independent Component
Analysis (ICA) among other CA methods. In the second part, several
extensions (linear and non-linear) to solve common problems in pattern
recognition (e.g. outliers, lack of training data, geometric invariance,
etc.) will be discussed. In the final part of the tutorial, I will
standard extensions of linear models such as kernels, latent variable
and tensor factorization. The tutorial will discuss how these
can be applied to different pattern recognition problems (e.g. face
recognition and tracking, pedestrian detection, gait recognition,
Back to top
- Generative Models (PCA, NMF,ICA) and extensions
- Discriminative Models (CCA, LDA, OCA) and extensions
- Kernel methods
- Latent variable models
- Tensor factorization
- Applications to pattern recogntion
8. Advanced Techniques for Face-Based Biometrics
Massimo Tistarelli, Universita di Sassari
Face recognition is nowadays one of the most challenging biometric modalities for the identification of individuals. In the last two decades several experimental as well as commercial systems have been developed exploiting different physical properties of the face image. Either being based on processing 2D or 3D information all these methods perform a face classification of the individuals based on some relevant features extracted from the raw image data. The data acquisition, preprocessing and the feature extraction/selection are all topics of the greatest importance to design a good performing recognition system. At the same time, the right choice of the features to be used as the basis for the face representation, which must be based on the uniqueness of such features, as well as most advanced issues such as the incorporation of quality information and the cope for ageing effects, are all of paramount importance.
The tutorial will consists of two sessions (half day of total duration) devoted to the description of both the basic and most advanced techniques related to face recognition. The lectures will provide a comprehensive outline of face-based biometrics, its relation to biological systems (the psychophysics of the human visual system), including the existing applications and commercial systems.
The lectures will provide an in-depth analysis of the state-of-the-art algorithms for face-image analysis including: face detection and tracking, landmark localization, feature extraction, face representation and classification.
The lectures will mainly explore the image processing aspects of the recognition process. As for classification, machine learning algorithms will be also presented, including kernel methods as related to learning and the approximation theory. The most relevant issues and problems will be raised, providing practical solutions and algorithms responding to them. Particular attention will be given to the most advanced and new techniques for face representation and classification, as well as the current approaches presented in the literature. Attention will be also given to the performance evaluation of face recognition systems providing some examples and results from recent competitions and public evaluation contests.
Finally, the tutorial will present three relevant and novel issues: the use of face image sequences for exploiting the time domain, the extension to 3D face analysis, and the how to cope with ageing and data quality.
Back to top
Human Vision System