4005-759-01: Pattern Recognition (RIT CS, 20081)

Department of Computer Science
Pattern Recognition (Topics in Artificial Intelligence, Fall 2008)

4005-759-01 (Calendar Description)
Home --- Syllabus --- Schedule & Slides --- Assignments --- Resources

Instructor: Richard Zanibbi, Office hours: 2-3:50pm Tuesdays and Thursdays, Room 70-3551 (Golisano College)
Lectures: 10-11:50am Tuesdays and Thursdays, Room 70-2590 (Golisano College)

Research Paper Presentations

Tuesday, Nov. 4

1. Semantic Analysis of Sports Video
Brian Rezenbrink
Abstract: This research will be a survey into the various components of semantic analysis of sports video. The research will identify underlying foundations of the field as well as conventional methods and successful deviations from those methods. The three major components of semantic analysis are temporal segmentation, visual tracking, and action classification. The sports that will analyzed will be grouped as team sports and individual sports, with subdivision into continuous motion sports and turn-based competition.

2. Content-based Image Retrieval for Documents
Li Yu
Abstract: For maintaining and accessing information in large databases of document images, it's not applicable for people to completely convert these documents images to electronic representations due to many factors such as cost of time and difficulties of converting rare and strange symbols. My research first briefly discusses traditional text indexing techniques on imperfect data and the retrieval of partially converted documents. After that, a more comprehensive review of techniques for the direct characterization, manipulation and retrieval, of documents images is provided and the strengths and limitations of existing technologies are discussed.

3. Feature Extraction in Face Recognition
Amit Pillay
Abstract: Of many biometric recognition systems, face recognition system have attracted much attention because of its varied unobtrusive applications, such as in airport security and access control, building surveillance and monitoring, human-computer intelligent interaction and so on. While there have been some successful face recognition systems build over 3 decades, most face recognition system fail when the facial pose is not fixed at a full frontal view, when the lighting is not controlled, when the facial expressions is varied, or any combination of the above. Variations in imaging condition have big impact on the performance of face recognition system. So far, no revolutionary and practical solutions are available for the problems cause by varying imaging conditions. These problems can be best possibly reduced in face recognition process if the features selected for recognition process are properly selected and preprocessed. This paper is trying to look into various methods that have been employed in recent past years for features selection and extraction from facial images which are under varying imaging conditions for face recognition and try to analyze different methods and possibly state some suggestions about the approaches.

4. Multi-Agent Learning
Owen Roberts
Abstract: Single agent learning has been studied for many years and is well developed. Their uses are limited though. Many environments require multiple agents to deal with it in an informative way. The methods of single agent learning are not very effective in dealing with multiple agents. Most often multiple agents are dealing with dynamic environments with many unknowns. Single agents classically deal with fixed environments with little to no unknowns. Cooperative agents have been pursued to answer this challenge. This talk will be about some of the methods being used to try and solve the multiagent learning problem.

5. Shape and Appearance Models in Image Recognition
Ishwarryah Ramanathan
Abstract: Image Recognition has been a field of research for decades. Applications of image recognition is not restricted to Bio-Medical images, as there are further more approaches that builds toward authentication systems using face and fingerprint algorithms. Biomedical images usually contain complex objects, which will vary in appearance significantly from one image to another. Many industrial applications involve assemblies with moving parts or components whose appearance can vary. In such cases flexible models or flexible templates can be used to allow some degree of variability in the shape of objects in the image. Various models have been proposed to construct the flexible model. This paper is trying to look into various techniques for image recognition which specifically deals with shape and texture properties of the object to be recognized.

6. Facial Expression Recognition in Motion Capture Systems
Yuqiong Joan Wang
Abstract: In this study we carried out a survey on existing human motion capture systems that are able to capture and recognize facial expressions. We found very few human motion capture systems are able to capture and classify facial expressions. Then we summarize recent facial motion capture and recognition methods, and analyze their potential to be applied in human motion capture systems. Most methods have both opportunities and setbacks to be integrated into motion capture systems. Finally we suggest that the integration of facial expression recognition into motion capture system will be an active and fruitful research area in the near future.

7. Neural Networks: Face Recognition
Doug Roberts
Abstract: The research will focus on neural networks and their application to face recognition. A survey of several different research papers has been conducted. The paper will give an overview of the process used in each of the papers. Finally a neural network for face detection will be discussed that will include the best features of each of the neural networks that were discussed in the researched papers.

Thursday, Nov. 6

1. Energy-based Models for Active Contour Evolution
Jeffrey Robble
Abstract: Contour evolution is a technique primarily used to identify regions of interest in images. Contours are active in the sense that they expand, contract, move, stretch, and bend in response to image features, internal transformation rules, and constraints imposed by an automated system or interactive user. One application of active contours is for solving the shortest path problem. The basic theory is that a contour behaves as a wave front that starts at the source point in the image and expands throughout a series of iterations until it encounters the known destination point. We discuss the various energies and forces that dictate the behavior of the wave and discuss the how this approach can be applied to different image environments.

2. Different Approaches to Text Classification
Ganesh Sugunan
Abstract: This survey will give a brief summary of three pattern recognition techniques used for text categorization, and explain the basics of how they work. These techniques are used because the subject creates a large feature space with a huge quantity of dimensions, there is an inherent level of uncertainty among humans as to the correct classification of some inputs, and all data must be classified into a set of predetermined classes. The techniques I cover here are: Support vector machines, Bayesian decision networks, and Neural nets.

3. Foreground-Background Separation in Natural Scenes
Zachary Busser
Abstract: Image matting is the process of extracting an alpha matte from a given image. This alpha matte segments the image into a background and foreground, allowing the foreground to be extracted or the background replaced. This paper is a survey of several techniques to extract such a matte from natural images, with a special focus on methods with foundations in spectral graph theory. Poisson and Bayesian approaches are examined as well, and the theory and practical results of each are compared.

4. Geometric Active Contour Models with shape priors for Cardiac Image Segmentation
Hongda Mao
Abstract: Active contour models, also known as deformable models, has brought tremendous impact to image segmentation. However, these classical models can't be directly used in many medical images with physical meanings. In this presentation, we will introduce some recent geometric active contour models with shape priors for cardiac image segmentation. First, we will introduce geometric active contour models, which can be roughly categorized into three classes: region-based, edge-based and combined models. Then we will focus our attention on geometric segmentation. Finally, we will discuss some other prior knowledge used in medical image segmentation.

5. Wavelet-Based Analysis of Microarray Data
James Defelice
Abstract: The analysis of microarray experiment data is a complex problem in biology. Common statistical techniques can sometimes fail to discover important relationships among groups of genes, indicating a need for alternative approaches. Wavelet transforms, often used in the imaging science and signal analysis realms, provide an alternative non-parametric technique for the analysis of microarray experiment data. This paper will examine the role of wavelet transforms in the analysis of microarray experiment data, in particular with regard to gene clustering, and feature selection for gene classification.

6. Feature Selection for Text Classification
Nicky Nicolosi
Abstract: Text classification is an important and well studied area of pattern recognition, with a variety of modern applications. Effective spam email filtering systems, automated document organization and management, and improved information retrieval systems all benefit from techniques within this field. The problem of feature selection, or choosing the most relevant features out of what can be an incredibly large set of data, is particularly important for accurate text classification. This paper provides a brief overview of the area of text classification, followed by a survey of several popular feature selection methods commonly used for text classification. Pruning techniques are also briefly discussed as a way to further reduce the set of possible features (typically words) within a document prior to applying a method of feature selection.

7. Scale and Rotation Invariant Recognition using Linear Combinations
Jacob Hays
Abstract: A particular problem in recognition of visual objects is that they become difficult to recognize if they are in spatial views that they were not trained at. An object trying to be recognized may be rotated, scaled, brightened, dimmed, or partially occluded from its trained view. An approach that is very effective against rotation and scaling of objects, is representing an 3-d object as a linear combination of 2-d images of that object at differing views. For this to be practical, there needs to be a way to corespondent the same features amongst 2 or more views of that object. Several methods of this have been developed such as SIFT (Scale Invarient Feature Transform) and SURF (Speeded up Robust Features), which both aim to find relevant key points in images, and match them across different images. They look for features that are easy to recognize regardless of rotation, scaling, and lighting of the object, and that when taken all together, can be used to uniquely identify that object.