Nikolaos Vassilas

Modified self-organizing feature map algorithms for efficient digital hardware implementation

This paper describes two variants of the Kohonens self-organizing feature map (SOFM) algorithm. Both variants update the weights only after presentation of a group of input vectors. In contrast, in the original algorithm the weights are updated after presentation of every input vector. The main advantage of these variants is to make available a finer grain of parallelism, for implementation on machines with a very large number of processors, without compromising the desired properties of the algorithm. In this work it is proved that, for one-dimensional (1-D) maps and 1-D continuous input and weight spaces, the strictly increasing or decreasing weight configuration forms an absorbing class in both variants, exactly as in the original algorithm. Ordering of the maps and convergence to asymptotic values are also proved, again confirming the theoretical results obtained for the original algorithm. Simulations of a real-world application using two-dimensional (2-D) maps on 12-D speech data are presented to back up the theoretical results and show that the performance of one of the variants is in all respects almost as good as the original algorithm. Finally, the practical utility of the finer parallelism made available is confirmed by the description of a massively parallel hardware system that makes effective use of the best variant.

MultiCAD-GA: A System for the Design of 3D Forms Based on Genetic Algorithms and Human Evaluation

The solution engine of MultiCAD-GA, presented in this work, is a part of a new software environment for efficient search for solutions in heavily demanding applications involving the design of three-dimensional forms, such as those of architectural and interior decoration design. MultiCAD-GA starts by using constraint programming techniques in order to find a set (population) of solutions (forms) that satisfy the spatial constraints imposed by the user and create an initial generation. In the sequel, it applies genetic operators to generate new solutions and interacts with the user in order to evaluate the solutions and increase the speed of convergence to those forms that satisfy his/her aesthetics. The forms are coded into chromosomes using the usual binary strings. Visualization of the results is performed through the VRML graphics language.

A New Methodology for Efficient Classification of MultispectralSatellite Images Using Neural Network Techniques

A methodology based on self-organizing feature maps and indexing techniques for time and memory efficient neural network training and classification of large volumes of remotely sensed data is presented. Results on land-cover classification of multispectral satellite images using two popular neural models show orders of magnitude of speedup with respect to both training and classification times. The generality of the proposed methodology is demonstrated with a dramatic improvement of the classification time of the k-nearest neighbors statistical classifier.

Content-Based coin retrieval using invariant features and self-organizing maps

During the last years, Content-Based Image Retrieval (CBIR) has developed to an important research domain within the context of multimodal information retrieval. In the coin retrieval application dealt in this paper, the goal is to retrieve images of coins that are similar to a query coin based on features extracted from color or grayscale images. To assure improved performance at various scales, orientations or in the presence of noise, a set of global and local invariant features is proposed. Experimental results using a Euro coin database show that color moments as well as edge gradient shape features, computed at five concentric equal-area rings, compare favorably to wavelet features. Moreover, combinations of the above features using L1 or L2 similarity measures lead to excellent retrieval capabilities. Finally, color quantization of the database images using self-organizing maps not only leads to memory savings but also it is shown to even improve retrieval accuracy.

Using neural networks and SVMs for automatic medical diagnosis: A comprehensive review

In this paper we make a comprehensive review regarding the use of neural networks in automated medical diagnosis, with a special emphasis in Support Vector Machines (SVMs), which are specialized types of neural functions. Through the study, we see that, in many cases, symptoms and diseases can be efficiently predicted by neural systems, while SVMs are increasingly used in medical diagnosis due to their accurate classification characteristics.

A New Method for Segmenting Newspaper Articles

Digital preservation of old newspapers contributes greatly to the historical register of a country’s social, political and economical events. At the same time, newspaper preservation is an imperative necessity because of the fast paper deterioration and difficulty in tracing the overwhelming amount of information. Lambrakis Press S.A. owns a large collection of newspapers and periodicals that consists of 1,300,000 pages and covers a time period from 1890 up to date. This material is divided into 600,000 A2 pages, 500,000 A3 tabloid and 200,000 A4 pages approximately. Our team is working on all aspects of the transformation procedure from the printed material to an accessible digital archive (verification and quality control, digitization, cataloguing, search and retrieval, design and content presentation). The final digital documents form the foundation of our digital library.

Fusion of aerial images with mean shift-based upsampled elevation data for improved building block classification

Nowadays there is an increasing demand for detailed 3D modeling of buildings using elevation data such as those acquired from LiDAR airborne scanners. The various techniques that have been developed for this purpose typically perform segmentation into homogeneous regions followed by boundary extraction and are based on some combination of LiDAR data, digital maps, satellite images and aerial orthophotographs. In the present work, our dataset includes an aerial RGB orthophoto, a DSM and a DTM with spatial resolutions of 20cm, 1m and 2m respectively. Next, a normalized DSM (nDSM) is generated and fused with the optical data in order to increase its resolution to 20cm. The proposed methodology can be described as a two-step approach. First, a nearest neighbor interpolation is applied on the low resolution nDSM to obtain a low quality, ragged, elevation image. Next, we performed a mean shift-based discontinuity preserving smoothing on the fused data. The outcome is on the one hand a more homogeneous RGB image, with smoothed terrace coloring while at the same time preserving the optical edges and on the other hand an upsampled elevation data with considerable improvement regarding region filling and “straightness” of elevation discontinuities. Besides the apparent visual assessment of the increased accuracy of building boundaries, the effectiveness of the proposed method is demonstrated using the processed dataset as input to five supervised classification methods. The performance of each method is evaluated using a subset of the test area as ground truth. Comparisons with classification results obtained with the original data demonstrate that preprocessing the input dataset using the mean shift algorithm improves significantly the performance of all tested classifiers for building block extraction.

Comparative analysis of classification techniques for building block extraction using aerial imagery and LiDAR data

Building detection has been a prominent area in the area of image classification. Most of the research effort is adapted to the specific application requirements and available datasets. In this paper we present a comparative analysis of different classification techniques for building block extraction. Our dataset includes aerial orthophotos (with spatial resolution 20cm), a DSM generated from LiDAR (with spatial resolution 1m and elevation resolution 20 cm) and DTM (spatial resolution 2m) from an area of Athens, Greece. The classification methods tested are unsupervised (K-Means, Mean Shift), and supervised (Feed Forward Neural Net, Radial-Basis Functions, Support Vector Machines). We evaluated the performance of each method using a subset of the test area. We present the classified images, and statistical measures (confusion matrix, kappa coefficient and overall accuracy). Our results demonstrate that the top unsupervised method is the Mean Shift that performs similarly to the best supervised methods.

Robust Line Detection in Images of Building Facades using Region-based Weighted Hough Transform

A robust region-based weighted Hough Transform method for the detection of straight lines in poor quality images of building facades is presented in this work. Following a typical preprocessing stage that includes color to grayscale transformation, binarization using Otsus automatic threshold selection method, morphological opening and decomposition into connected regions a minimum bounding rectangle is then fitted to each region. A measure of rectangularity is then used in order to filter out regions with low rectangularity scores from further processing. Finally, the contribution of each region to the accumulator array is computed using an appropriately designed kernel that models the uncertainties associated with the geometrical center and orientation of the regions minimum bounding rectangle. Comparisons performed on facade images taken with impaired visual conditions or with low accuracy sensors (e.g. thermal images) between the proposed method and other Hough Transform algorithms, show an improved accuracy of our method in detecting lines and/or linear formations. The proposed method is also used with success for image rectification through vanishing points estimation.

Comparative study of visual feature extraction methods for building retrieval on urban databases

In a great variety of applications, such as real-time robot localization and visual navigation, architectural design, 3D city reconstruction, building recognition in urban environments is required. In this work a comparative study of visual feature extraction methods for building retrieval on urban databases is performed. To this end, a database of 183 building facades taken at two different regions near the center of Athens was created. Five feature extraction methods are used in order to capture visual properties. Features between images are then matched using RANSAC algorithm and a visual transformation is calculated. Comparative retrieval results are provided and discussed.