Nese Yalabik

Ensemble Classifiers for Medical Diagnosis of Knee Osteoarthritis Using Gait Data

Automated or semi-automated gait analysis systems are important in assisting physicians for diagnosis of various diseases. The objective of this study is to discuss ensemble methods for gait classification as a part of preliminary studies of designing a semi-automated diagnosis system. For this purpose gait data is collected from 110 sick subjects (having knee osteoarthritis (OA)) and 91 age-matched normal subjects. A set of multilayer perceptrons (MLPs) is trained by using joint angle and time-distance parameters of gait as features. Large dimensional feature vector is decomposed into feature subsets and the ones selected by gait expert are used to categorize subjects into two classes; healthy and patient. Ensemble of MLPs is built using these distinct feature subsets and diversification of classifiers is analyzed by cross-validation approach and confusion matrices. High diversifications observed in the confusion matrices suggested that using combining methods would help. Indeed, when a proper combining rule is applied to decomposed sets, more accurate results are obtained. The result suggests that ensemble of MLPs could be applied in the automated diagnosis of gait disorders in a clinical context

Evaluation of textural features for multispectral images

Remote sensing is a field that has wide use, leading to the fact that it has a great importance. Therefore performance of selected features plays a great role. In order to gain some perspective on useful textural features, we have brought together state-of-art textural features in recent literature, yet to be applied in remote sensing field, as well as presenting a comparison with traditional ones. Therefore we selected most commonly used textural features in remote sensing that are grey-level co-occurrence matrix (GLCM) and Gabor features. Other selected features are local binary patterns (LBP), edge orientation features extracted after applying steerable filter, and histogram of oriented gradients (HOG) features. Color histogram feature is also used and compared. Since most of these features are histogram-based, we have compared performance of bin-by-bin comparison with a histogram comparison method named as diffusion distance method. During obtaining performance of each feature, k-nearest neighbor classification method (k-NN) is applied.

Combining neural networks for gait classification

Gait analysis can be defined as the numerical and graphical representation of the mechanical measurements of human walking patterns and is used for two main purposes: human identification, where it is usually applied to security issues, and clinical applications, where it is used for the non-automated and automated diagnosis of various abnormalities and diseases. Automated or semi-automated systems are important in assisting physicians for diagnosis of various diseases. In this study, a semi-automated gait classification system is designed and implemented by using joint angle and time-distance data as features. Multilayer Perceptrons (MLPs) Combination classifiers are used to categorize gait data into two categories; healthy and patient with knee osteoarthritis. Two popular approaches of combining neural networks are experimented and the results are compared according to different output combining rules. In the first one, same set is used to train all networks and afterwards the features are decomposed into five different sets. These two experiments show that using entire data set produces more accurate results than using decomposed data sets, but complexity becomes an important drawback. However, when a proper combining rule is applied to decomposed sets, results are more accurate than entire set. In this experiment sum rule produces better results than majority vote and max rules as an output combining rule.

Unsupervised land use - land cover classification for multispectral images

In this paper, land use/land cover classification of multispectral imagery with unsupervised approaches are presented. Primarily, a pixel based recognition algorithm is applied in three stages. At the first stage, water bodies are classified by using the NIR band histogram. At the second stage, combination of several vegetation indices are used to locate vegetation and at the third stage, by using Gabor filter man-made structures are classified and the unclassified fields are left. Followingly in order to increase the success rate, pixel based classification results are combined with meanshift segmentation results and a homogeneity test is applied for each segment. The segments that passed the homogeneity test are classified to corresponding class and for the rest, pixel based results are assigned. Compared to the similar works, this approach gives successful results.

Object orientation detection and character recognition using optimal feedforward network and Kohonen's feature map

A neural network model, namely, Kohonens Feature Map, together with the optimal feedforward network is used for variable font machine printed character recognition with tolerance to rotation, shift in position, and size errors. The determination of object orientation is found using the many rotated versions of individual symbols. Orientations are detected from printed text, but no knowledge of the context is used. The optimal Bayesian detector is derived, and it is shown that the optimal detector has the form of a feedforward network. This network together with the learning vector quantization (LVQ) approach is able to implement an inspection system which determines the orientation of the fonts. After the size normalization, rotation, and component finding process as a preprocessing step, the text becomes the input for the feature map. The feature map is trained first in an unsupervised manner. The algorithm is then adapted for supervised learning using improved LVQ technique. Rectangular and minimal spanning tree (MST) neighborhood topologies are experimented with. The results are encouraging, 87% of the characters of various fonts are correctly recognized even though the pattern is distorted in shape and transformed in a shift, size, and rotation invariant manner. Experimental results and comparisons are described.

An Activity Planning and Progress Following Tool for Self-Directed Distance Learning

This paper presents an overview of the design and implementation of an Activity Planning and Progress Following Tool (APT) for e-learning. APT is a software based support tool designed to assist learners with their self directed distance learning (SDDL). A literature survey for the additional components of Learning Management Systems (LMS) that are needed for SDDL is conducted. As a result, the importance of “planning”, “feedback” and “resources” properties are revealed. In the continuation of the study, APT was developed by adding “Recommended Study Time Periods”, “Resources”, “Study Planning”, “User Goals” and “Progress Report” modules to the “Content Administration Tool” (CAT), which was developed before in another study.

‘Net-Class’ and a Comparison of the Web Based Course Management Tools

A web based course management tool, ‘Net-Class’, developed in Middle East Technical University is presented. The tool features such as web browsing abilities, instructor and student tools, synchronous and asynchronous sharing is discussed. It is compared to the commercially available tools in terms of these features. The evaluation shows Net-Class is at least as effective as the others in teaching and learning via web.

Connected Word Recognition Using Neural Networks

A connected word recognition system which makes use of two neural network models, namely, Kohonen’s Network and a Multilayer Perceptron is implemented. The digitized speech signal is represented by a sequence of Linear Predictive Coding (LPC) coefficients and segmented into syllables. The Kohonen’s Network is used to perform vector quantization to compress LPC data for the input of the Multilayer Perceptron (MP). MP is used to perform recognition in syllable basis with the Back-Propagation algorithm used for training. The words are constructed by using the sequence of recognized syllables. The system was trained and tested with ten Turkish words with sixteen syllables, with an overall recognition rate of 90 percent.

The Effect of AI Applications on Graduate Education in Turkey

In the last few years, some real-life applications in Turkey created a demand for high level education in subjects like Pattern Recognition, Artificial Intelligence and Expert Systems. The pressing needs of the government provide a good opportunity for the university for conducting M.S. and Ph.D. theses and placing demand in graduate courses in AI. In this study, the nature of some of these applications and their effect on higher education in Middle East Technical University will be discussed.

An efficient algorithm for recognizing isolated Turkish words

In this study, a computationally efficient speaker independent isolated word recognition system for Turkish language is designed and implemented. The approach used is a combination of whole-word matching techniques with segmentation into phonetic units before classification. Linear Predictive Coding (LPC) coefficients for an eight-pole model of the short-time signal are used as feature vectors. Computational costs are reduced by a two-step classification strategy where unlikely words are eliminated in the first step by comparing only the first syllable. The Dynamic Time Warping (DTW) method is used in comparisons at both levels.