Mohamad T. Musavi

Original Contribution: On the training of radial basis function classifiers

An approach for the implementation of the Radial Basis Function (RBF) technique is presented and applied to a network of the appropriate architecture. The paper explores a methodology for selecting kernel function parameters and the extent to which the number of RBF nodes can be reduced without significantly affecting the overall training error. These objectives are accomplished through algorithms that shall be described in detail. Emphasis is also placed on the problems faced by a technique that has been proved superior to the more traditional training algorithms, particularly in terms of processing speed and solvability of nonlinear patterns. Solutions are consequently proposed in view of making RBF a more efficient method for interpolation and classification purposes.

Self-organised clustering for road extraction in classified imagery

Abstract The extraction of road networks from digital imagery is a fundamental image analysis operation. Common problems encountered in automated road extraction include high sensitivity to typical scene clutter in high-resolution imagery, and inefficiency to meaningfully exploit multispectral imagery (MSI). With a ground sample distance (GSD) of less than 2 m per pixel, roads can be broadly described as elongated regions. We propose an approach of elongated region-based analysis for 2D road extraction from high-resolution imagery, which is suitable for MSI, and is insensitive to conventional edge definition. A self-organising road map (SORM) algorithm is presented, inspired from a specialised variation of Kohonens self-organising map (SOM) neural network algorithm. A spectrally classified high-resolution image is assumed to be the input for our analysis. Our approach proceeds by performing spatial cluster analysis as a mid-level processing technique. This allows us to improve tolerance to road clutter in high-resolution images, and to minimise the effect on road extraction of common classification errors. This approach is designed in consideration of the emerging trend towards high-resolution multispectral sensors. Preliminary results demonstrate robust road extraction ability due to the non-local approach, when presented with noisy input.

Localization using neural networks in wireless sensor networks

Noisy distance measurements are a pervasive problem in localization in wireless sensor networks. Neural networks are not commonly used in localization, however, our experiments in this paper indicate neural networks are a viable option for solving localization problems. In this paper we qualitatively compare the performance of three different families of neural networks: Multi-Layer Perceptron (MLP), Radial Basis Function (RBF), and Recurrent Neural Networks (RNN). The performance of these networks will also be compared with two variants of the Kalman Filter which are traditionally used for localization. The resource requirements in term of computational and memory resources will also be compared. In this paper, we show that the RBF neural network has the best accuracy in localizing, however it also has the worst computational and memory resource requirements. The MLP neural network, on the other hand, has the best computational and memory resource requirements.

Inversion of ocean color observations using particle swarm optimization

Inversion of ocean color reflectance measurements can be cast as an optimization problem, where particular parameters of a forward model are optimized in order to make the forward-modeled spectral reflectance match the spectral reflectance of a given in situ sample. Here, a simulated ocean color dataset is used to test the capability of a recently introduced global optimization process, particle swarm optimization (PSO), in the retrieval of optical properties from ocean color. The performance of the PSO method was compared with the more common genetic algorithms (GA) in terms of model accuracy and computation time. The PSO method has been shown to outperform the GA in terms of model error. Of particular importance to ocean color remote sensing is the speed advantage that PSO affords over GA.

A vision based method to automate map processing

Abstract This paper presents a method for the digitization, vectorization and storage of land record maps. Map data is input into a vision system via a CCD camera as raw data to be digitized. The digitized data is accompanied by various types of noise mainly due to the maps symbols and the binarization process. This data is passed to an intelligent program for noise removal. The processed data is converted to vector data by an efficient algorithm. This point data that presents the key points can be stored on peripheral storage devices for future reconstruction and modification of the map. Experimental results for a test map have been presented. This method has been implemented on an IBM-PC compatible machine vision system keeping in mind its office related applicability.

Automated Extraction of Linear Features from Aerial Imagery Using Kohonen Learning and GIS Data

An approach to semi-automated linear feature extraction from aerial imagery is introduced in which Kohonens self-organizing map (SOM) algorithm is integrated with existing GIS data. The SOM belongs to a distinct class of neural networks which is characterized by competitive and unsupervised learning. Using radiometrically classified image pixels as input, appropriate SOM network topologies are modeled to extract underlying spatial structures contained in the input patterns. Coarse-resolution GIS vector data is used for network weight and topology initialization when extracting specific feature components. The Kohonen learning rule updates the synaptic weight vectors of winning neural units that represent 2-D vector shape vertices. Experiments with high-resolution hyperspectral imagery demonstrate a robust ability to extract centerline information when presented with coarse input.

A probabilistic model for evaluation of neural network classifiers

Abstract A technique for evaluation of the generalization ability in artificial neural network (ANN) classifiers is presented. A probabilistic input model is proposed to account for all possible input ranges. The expected value of a square error function over the defined input range is taken as a measure of generalization ability. The minimization of the error function outlines the boundary of the decision region for a minimum error neural network (MENN) classifier. Two essential elements for carrying out the proposed technique are the estimation of the input density and numerical integration. A non-parametric method is used to locally estimate the distribution around each training pattern. The Monte Carlo method has been used for numerical integration. The evaluation technique was tested for measuring the generalization ability of back propagation (BP), radial basis function (RBF), probabilistic neural network (PNN) and MENN classifiers for different problems.

Adaptive detection of small sinusoidal signals in non-Gaussian noise using an RBF neural network

This paper addresses the application of locally optimum (LO) signal detection techniques to environments in which the noise density is not known a priori. For small signal levels, the LO detection rule is shown to involve a nonlinearity which depends on the noise density. The estimation of the noise density is a major part of the computational burden of LO detection rules. In this paper, adaptive estimation of the noise density is implemented using a radial basis function neural network. Unlike existing algorithms, the present technique places few assumptions on the properties of the noise, and performs well under a wide variety of circumstances. Experimental results are shown which illustrate the system performance as a variety of noise densities are encountered.

A minimum error neural network (MNN)

Abstract A minimum error neural network (MNN) model is presented and applied to a network of the appropriate architecture. The associated one-pass learning rule involves the estimation of input densities. This is accomplished by utilizing local Gaussian functions. A major distinction between this network and other Gaussian based estimators is in the selection of covariance matrices. In MNN, every single local function has its own covariance matrix. The Gram-Schmidt orthogonalization process is used to obtain these matrices. In comparison with the well known probabilistic neural network (PNN), the proposed network has shown improved performance.

Mouse chromosome classification by radial basis function network with fast orthogonal search

This paper provides the results of our study on automatic classification of mouse chromosomes. A radial basis function neural network was compared with a multi-layer perceptron and a probabilistic neural network. The networks were trained and tested with 3723 chromosomes presented to each network as 30-point banding profiles. The radial basis function classifier trained with the fast orthogonal search learning rule provided the best unconstrained classification error rate of 12.7% which was obtained with a training set of 2250 chromosomes.