Ioannis M. Stephanakis

Regularized image restoration in multiresolution spaces

A regularized image restoration model in multiresolution spaces is proposed. The image is transformed in the wavelet domain, and the restored image is found as the solution of a constrained mean- squared-error optimization. The image energies in the subbands of a wavelet decomposition of the image set the constraints of the optimiza- tion problem. A closed-form relationship of the image energy in the wavelet channels is derived. The restored image is obtained by solving iteratively the equation that sets the derivative of the error norm equal to zero. The Lagrange multiplier that corresponds to a subband is given as a function of the noise and the image energy in the subband. Results of numerical experiments on multiresolution image restoration using one and nine constraints, respectively, are presented and compared against image restoration results obtained from a conventional regularized im- age restoration algorithm that is based on generalized cross validation and uses the Laplacian operator as the smoothing filter. The experimen- tal results obtained by the proposed method indicate that the restored image is closer to the original if one solves the restoration problem using more than one smoothing filter.

A novel data preprocessing method for boosting neural network performance

Data preprocessing methods have been used in Machine Learning classificationproblems, transforming datasets into a proper form in order to boost the classification performance.In thispaper,a novel data preprocessing method is proposed and evaluatedin a difficult classification data set, in which various classifiers have average performance lower than 50%. The dataset is related to osteoporosis disease, which is a disease of bones that leads to an increased risk of fracture and it is characterized by low bone mineral density and micro-architectural deterioration of bone tissue. The dataset consists of 589 subjects whose diagnosis was based on laboratory and osteal bone densitometry examination. In all cases, thirty three diagnostic factors for osteoporosis risk prediction, were usedin order to categorize subjects into three classes (normal, osteopenia, and osteoporosis). The performance of the proposed multilayer perceptron classifierin various topologies and training parameters was evaluated using the well-known 10-fold cross validationmethodand the results are reportedanalytically.The results indicate that the generated features after preprocessing of the original dataset significantly improve the accuracy of the resulting classifiers.

Wavelet-based approach to adaptive Wiener filtering of images in the presence of colored noise

An adaptive Wiener filtering approach, based on image estimation in the wavelet domain, is proposed. Gradient-based estimation of the image is employed by minimizing an error functional that depends on estimates of the image and the power of the noise in each wavelet subband. The power of the noise is estimated from the variance of wavelet coefficients. The Wiener filter used in restoring the corrupted image is updated at each iteration step. Hard wavelet thresholding allows for appropriate selection of initial estimates of the restored image and the noise. The initial estimate of the restored image is then improved by updating these estimates iteratively. Conditions for convergence of the proposed procedure are derived. Experimental results for restoring images corrupted by colored noise are presented as well. Comparisons with conventional restoration techniques strongly favor the proposed method.

Color segmentation using self-organizing feature maps (SOFMs) defined upon color and spatial image space

A novel approach to color image segmentation is proposed and formulated in this paper. Conventional color segmentation methods apply SOFMs - among other techniques - as a first stage clustering in hierarchical or hybrid schemes in order to achieve color reduction and enhance robustness against noise. 2-D SOFMs defined upon 3-D color space are usually employed to render the distribution of colors of an image without taking into consideration the spatial correlation of color vectors throughout various regions of the image. Clustering color vectors pertaining to segments of an image is carried out in a consequent stage via unsupervised or supervised learning. A SOFM defined upon the 2-D image plane, which is viewed as a spatial input space, as well as the output 3-D color space is proposed. Two different initialization schemes are performed, i.e. uniform distribution of the weights in 2-D input space in an ordered fashion so that information regarding local correlation of the color vectors is preserved and jointly uniform distribution of the weights in both 3-D color space and 2-D input space. A second stage of Density-Based Clustering of the nodes of the SOM (utilizing an ad hoc modification of the DBSCAN algorithm) is employed in order to facilitate the segmentation of the color image.

Enhancement of medical images using a fuzzy model for segment dependent local equalization

A novel model for fuzzy equalization of medical images is proposed. The method requires a preprocessing step which accomplishes a fuzzy segmentation of the image into N segments with overlapping fuzzy boundaries. Histogram equalization is performed according to individual equalization functions derived from each segment of the image. A fuzzy membership function is associated with each segment. Enhancement results compare well against histogram equalization techniques that derive the equalization transform function from the global histogram of the image.

Optimal wavelet filter banks for regularized restoration of noisy images

Regularized image restoration methods efficiently handle the ill-posed problem of image restoration. Nevertheless, the issue of selecting the regularization parameter as well as the smoothing filter still constitutes an open research topic. A model of regularized image restoration is introduced and analyzed in this paper. The proposed model assumes that wavelet filter banks replace the smoothing filter of conventional regularized restoration. Filter factorizations for the optimal design of wavelet filter banks using the generalized-cross-validation (GCV) criterion are presented, and novel expressions of the influence matrix, which is used to calculate the GCV error, are derived. The error of the GCV method is expressed in terms of the modulation matrix of the filter bank and the modulation vector of the degradation filter. The expressions are given in general form for optimal wavelet filter bank design upon arbitrary sampling lattices. The numerical examples of image restoration using the proposed method that are presented indicate significant signal-to-noise ratio improvement, ΔSNR, compared to image restoration methods that employ the Laplacian as the smoothing filter.

A Particle Swarm Optimization (PSO) Model for Scheduling Nonlinear Multimedia Services in Multicommodity Fat-Tree Cloud Networks

Cloud computing delivers computing services over virtualized networks to many end-users. Virtualized networks are characterized by such attributes as on-demand self-service, broad network access, resource pooling, rapid and elastic resource provisioning and metered services at various qualities. Cloud networks provide data as well as multimedia and video services. They are classified into private cloud networks, public cloud networks and hybrid cloud networks. Linear video services include broadcasting and in-stream video that may be viewed in a video player whereas non-linear video services include a combination of in-stream video with on-demand services, which are originated from distributed servers in the network and deliver interactive and pay-per view content. Furthermore heterogeneous delivery networks that include fixed and mobile internet infrastructures require that adaptive video streaming should be carried out at network boundaries based on such protocols as HTTP Live Streaming (HLS). Distributed processing of nonlinear video services in cloud environments is addressed in the present work by defining Distributed Acyclic Graphs (DAG) models for multimedia processes executed by a set of non-locally confined virtual machines. A novel discrete multivalue Particle Swarm Optimization (PSO) algorithm is proposed in order to optimize task scheduling and workflow. Numerical simulations regarding such measures as Schedule-Length-Ratio (SLR) and Speedup are given for novel fat-tree cloud architectures.

Developing Innovative Live Video-to-Video Communications for Smarter European Cities

The LiveCity Project effort intends to create a city-based “Living Lab” and associated ecosystem to pilot live interactive high-definition video-to-video (v2v) on ultrafast wireless and wireline Internet infrastructure for the support of appropriate public service use cases among a number of city user communities initially in four major European cities (Dublin, Athens, Luxembourg (city) and Valladolid). The essential aim is to empower the citizens of a city to interact with each other in a more productive, efficient and socially useful way by using v2v over the Internet, as the latter can be considered to improve city administration, reduce fuel costs and carbon footprint, enhance education, improve city experiences for tourists/cultural consumers and save patients’ lives. LiveCity underpins technology which has the ability to massively scale while it integrates the necessary ingredients in an efficient low-cost manner and provides a proper testing ground for a mass market deployment to the cities in Europe.

A single layer linear feedforward neural network for signal estimation in the wavelet domain

A model of adaptive signal estimation in the presence of noise is proposed using the wavelet transform and single layer linear feedforward neural networks. The networks successively estimate the residual of the approximations (details) of the input at different scales from coarse scales (low frequencies) to fine scales by minimizing the error of each channel of the wavelet transform separately. Since learning takes place in subspaces of the signal space one needs to train the networks with fewer weights. The rate of convergence of the learning algorithm in several transform channels is investigated. It is shown that learning in the input space may diverge, whereas convergence is achieved in wavelet spaces. Experimental results are presented which compare the proposed approximation method with adaptive approximation using sigmodal functions and exponentially decaying kernels.

Anomaly Detection In Secure Cloud Environments Using a Self-Organizing Feature Map (SOFM) Model For Clustering Sets of R-Ordered Vector-Structured Features

Cloud computing delivers services over virtualized networks to many end-users. Cloud services are characterized by such attributes as on-demand self-service, broad network access, resource pooling, rapid and elastic resource provisioning and metered services of various qualities. Cloud networks provide data as well as multimedia and video services. Cloud computing for critical structure IT is a relative new area of potential applications. Cloud networks are classified into private cloud networks, public cloud networks and hybrid cloud networks. Anomaly detection systems are defined as a branch of intrusion detection systems that deal with identifying anomalous events with respect to normal system behavior. A novel application of a Self-Organizing-Feature Map (SOFM) of reduced/aggregate sets of ordered vector structured features that are used for detecting anomalies in the context of secure cloud environments is herein proposed. Multivalue inputs consist of reduced/aggregate ordered sets of vector and binary features. The nodes of the SOFM - after training - are indicative of local distributions of feature measurements during normal cloud operation. Anomalies are detected as outliers of the trained SOFM. Each structured vector consists of binary as well as histogram data. The aggregated Canberra distance is used to order histogram data whereas the Jaccard distance is used for multivalue binary data. The so-called Cross-Order Distance Matrix is defined for both cases. The distance depends upon the selection of a similarity/distance measure and a method for operating upon the elements of the Cross-Order Distance Matrix. Several methods of estimating the distance between two ordered sets of features are investigated in the course of this paper.