Malik Tahir Hassan

CDIM: Document Clustering by Discrimination Information Maximization

Ideally, document clustering methods should produce clusters that are semantically relevant and readily understandable as collections of documents belonging to particular contexts or topics. However, existing popular document clustering methods often ignore term-document corpus-based semantics while relying upon generic measures of similarity. In this paper, we present CDIM, an algorithmic framework for partitional clustering of documents that maximizes the sum of the discrimination information provided by documents. CDIM exploits the semantic that term discrimination information provides better understanding of contextual topics than term-to-term relatedness to yield clusters that are describable by their highly discriminating terms. We evaluate the proposed clustering algorithm using well-known discrimination/semantic measures including Relative Risk (RR), Measurement of Discrimination Information (MDI), Domain Relevance (DR), and Domain Consensus (DC) on twelve data sets to prove that CDIM produces high-quality clusters comparable to the best methods. We also illustrate the understandability and efficiency of CDIM, suggesting its suitability for practical document clustering.

Terms-based discriminative information space for robust text classification

With the popularity of Web 2.0, there has been a phenomenal increase in the utility of text classification in applications like document filtering and sentiment categorization. Many of these applications demand that the classification method be efficient and robust, yet produce accurate categorizations by using the terms in the documents only. In this paper, we propose a novel and efficient method using terms-based discriminative information space for robust text classification. Terms in the documents are assigned weights according to the discrimination information they provide for one category over the others. These weights also serve to partition the terms into category sets. A linear opinion pool is adopted for combining the discrimination information provided by each set of terms to yield a feature space (discriminative information space) having dimensions equal to the number of classes. Subsequently, a discriminant function is learned to categorize the documents in the feature space. This classification methodology relies upon corpus information only, and is robust to distribution shifts and noise. We develop theoretical parallels of our methodology with generative, discriminative, and hybrid classifiers. We evaluate our methodology extensively with five different discriminative term weighting schemes on six data sets from different application areas. We give a side-by-side comparison with four well-known text classification techniques. The results show that our methodology consistently outperforms the rest, especially when there is a distribution shift from training to test sets. Moreover, our methodology is simple and effective for different application domains and training set sizes. It is also fast with a small and tunable memory footprint.

Video summarization: Sports highlights generation

The mechanism to recognize highlights from videos is basic and fundamental problem for indexing and retrieval applications. In this paper we propose techniques to generate sports highlights from cricket video using techniques of optical character recognition. First the score bar is extracted from the frames then the character recognition techniques are used to extract information for events like sixes, fours and wickets. A short video summary is synthesized that includes the frames for the aforementioned significant events termed as Highlights. The process of sports highlights generation is automated resulting in a condensed summary for the viewer that reduces the time and space requirements.

Using Temporal Covariance of Motion and Geometric Features via Boosting for Human Fall Detection

Fall induced damages are serious incidences for aged as well as young persons. A real-time automatic and accurate fall detection system can play a vital role in timely medication care which will ultimately help to decrease the damages and complications. In this paper, we propose a fast and more accurate real-time system which can detect people falling in videos captured by surveillance cameras. Novel temporal and spatial variance-based features are proposed which comprise the discriminatory motion, geometric orientation and location of the person. These features are used along with ensemble learning strategy of boosting with J48 and Adaboost classifiers. Experiments have been conducted on publicly available standard datasets including Multiple Cameras Fall (with 2 classes and 3 classes) and UR Fall Detection achieving percentage accuracies of 99.2, 99.25 and 99.0, respectively. Comparisons with nine state-of-the-art methods demonstrate the effectiveness of the proposed approach on both datasets.

Feature based face recognition using slopes

Face recognition from an image is a popular problem in biometrics research. In the last decade, a lot of research has been done in this area. The advantage of face based identification over other biometrics is its wide acceptability as it does not require any keys, tokens, smart cards, PINs, plastic cards or passwords, etc. In this work, face recognition has been done using several feature based approaches. Two new methods are presented in which simple yet useful new features are proposed and evaluated. The main contribution of this paper is the usage of a slope table along with the other features for face recognition. The slopes of different fiducial points of facial components, i.e., left eye, right eye, nose and lips are computed to fill the slope table. These two methods are compared with the existing approaches based on popular features like principal components and ratios of facial components. The results show that our proposed methods outperform these existing approaches.

Self-Calibration: Enabling Self-Management in Autonomous Systems by Preserving Model Fidelity

Autonomic and autonomous systems exist within a world view of their own. This world view is created from the training data and assumptions that were available at their inception. In most of these systems this world view becomes obsolete over time due to changes in the environment. This brings a level of inaccuracy in the autonomic behavior of the system. When this degradation reaches a certain threshold self-healing or self-optimizing systems generally recreate the world view using current data and assumptions. However, the self-optimization process is akin to kill a fly with a hammer for minor tuning of the world view. Instead we propose the idea of self-calibration for self-managing these systems. We define self-calibration as the ability of the system to perceive the need for and the ability to execute minimal tuning to bridge the gap between systems world view and incoming information about the outside world. In this paper we present a case for considering self-calibration as a self-* enabling property of systems specifically for time-critical systems using data-centric AI technologies. We present our case by discussing three case studies from different domains where self-calibration enables a system to become self-healing or self-optimizing. We then place self-calibration in a generic system and explicitly describe the types of systems in which self-calibration can be implemented and the benefits that one can accrue from its inclusion.