Radu DOGARU

University Politehnica of Bucharest, Applied Electronics and Information Engineering, Faculty Member

Followers

Following

Public Views

Interests

Uploads

Papers by Radu DOGARU

Pneumonia Detection on Chest X-Ray Images using Convolutional Neural Networks Designed for Resource Constrained Environments

2020 International Conference on e-Health and Bioengineering (EHB), 2020

The problem of detecting pneumonia from X-ray images is considered here in the context of artific... more The problem of detecting pneumonia from X-ray images is considered here in the context of artificial intelligence. Particularly, in this work the possibility to use light-weight convolutional deep learning models is investigated from the perspective of building efficient recognition tools operating on portable equipment (resource-constrained resources), handy available to the medical personnel. Several state of the art lightweight deep neural network architectures as well as more complex (for better performance) ones were compared not only from the point of view of recognition accuracy but also from the complexity perspective. MobileNetV3, ShuffleNetV2 and SlimNet models with Android implementations in TensorFlow Lite are proposed to give a good compromise between accuracy and portability in building intelligent medical instruments.

Optimized Super Fast Support Vector Classifiers Using Python and Acceleration of RBF Computations

2018 International Conference on Communications (COMM), 2018

An optimized implementation for a parameter-less radial-basis neural paradigm called super fast v... more An optimized implementation for a parameter-less radial-basis neural paradigm called super fast vector support classifier (SFSVC) is presented. Training resides only in a process of selecting centers and consequently it is a very fast process, linear in the number of hidden neurons. Several improvements make the resulting implementation a highly effective, high speed., machine learning engine convenient for various applications on a large variety of platforms. It is shown that among various programming environments., Python with optimized math library support represents the best choice for its implementation. An important (up to 40 times) speed-up for the testing time is achieved by implementing the distance calculations as matrix multiplications., thus accelerating RBF computations on platforms with optimized math and linear algebra libraries such as Intel's MKL. Another important aspect of SFSVC., making it suitable for various embedded platforms (e.g. micro controllers., FPGA., etc.) is the lack of any tunable parameter except a unique radius., leading to a very convenient structure. The overall test plus training speed is better than for support vector machines while SFSVC has the important advantage of accepting arbitrary RBF kernels.

Novel Applications of Complexity Inspired RDT Transform for Low Complexity Embedded Speech Recognition in Automotive Environments

2017 21st International Conference on Control Systems and Computer Science (CSCS), 2017

Embedded dictation, i.e. recognizing vocal commands in noisy environments, with good accuracy and... more Embedded dictation, i.e. recognizing vocal commands in noisy environments, with good accuracy and using low complexity implementations is a desirable task with many applications. Such applications include automotive infotainment solutions particularly when no connectivity is available, personal assistants including embedded dictation solutions for disabled people, and so on. This paper reports our novel results in applying a nonlinear transform (RD-transform) introduced in a previous work and inspired form complexity measurements of signals generated in cellular automaton. Such a transform is compact and has a low computational complexity yet it previously proved quite efficient in terms of accuracy for a standard task of recognizing user independent dictation of digits. Herein, we report results on employing RD-transform on a specially designed sound database containing commands for the non-critical automotive equipment in a realistic, noisy environment. In addition to specific nonlinear transforms, low complexity FSVC classifiers were employed proving that good accuracies can be achieved using a very convenient implementation solution.

Shape object classification using echoes inside a virtual environment application

2015 7th International Conference on Electronics, Computers and Artificial Intelligence (ECAI), 2015

This work investigates the geometric object-shape classification using the echoes generated by va... more This work investigates the geometric object-shape classification using the echoes generated by various kinds of obstacles in a cellular automata based virtual environment for ultra-sound propagation. The virtual environment is implemented as a JAVA platform [1] capable of emulate sound propagation in a controlled 2D environment. The echoes are preprocessed by a Feature Processor Vector Unit (FVPU) and then classified using 2 neural networks: Fast Support Vector Classifier (FSVC) and Extreme Learning Machine (ELM).

A low complexity solution for epilepsy detection using an improved version of the reaction-diffusion transform

2017 5th International Symposium on Electrical and Electronics Engineering (ISEEE), 2017

Recognition of epileptic seizures is an important issue and in certain circumstances it is desira... more Recognition of epileptic seizures is an important issue and in certain circumstances it is desirable to have portable equipment implementing the algorithm in order to better monitor the patients. This work considers a widely used EEG database from University of Bonn as reference for comparing our recognition method with other previously reported. In order to perform epileptic seizures we combine a low complexity nonlinear transform applied to the EEG time-series with a fast support vector classifier (FSVC), a low complexity classifier introduced in previous works. Since EEG signals used in epilepsy detection are complex signals produced in a large network of neurons it make sense to extract proper features using an optimized version of a nonlinear transform (previously introduced as the reaction-diffusion transform — RDT) having its roots in our previous work in the field of cellular nonlinear networks. Results reported here confirm that very good performance (98.67% accuracy) can be obtained, while using a very low complexity solution which is easy to integrate in a portable device. Since RDT provided already very good results in recognizing speech commands, there is a good evidence to consider it as an useful preprocessing method for a wider range of signals or time-series to be recognized.

2003] “Simplicial RTD-based cellular nonlinear networks

Recently, a novel structure called the simplicial cellular neural network (CNN) has been introduc... more Recently, a novel structure called the simplicial cellular neural network (CNN) has been introduced [1], which permits to implement any Boolean/Gray-level function of any number of variables. This paper is devoted to explore novel circuit architectures for the implementation of the simplicial CNN based on resonant tunneling diodes. The final objective is to implement a fully programmable CNN in a hardware platform based on nanoelectronic devices.

Download

RD-CNN: A Compact and Efficient Convolutional Neural Net for Sound Classification

2020 International Symposium on Electronics and Telecommunications (ISETC), 2020

Classification of signals (sounds, biomedical, etc.) occurs in various circumstances, often requi... more Classification of signals (sounds, biomedical, etc.) occurs in various circumstances, often requiring low-complexity implementations for various resources-constrained platforms such as mobile or embedded systems within the IoT framework. Herein we describe preliminary results for a novel recognition system where a compact and fast transform, the reaction-diffusion transform (RDT), is used to generate spectral images. Such images are then processed into a novel type of compact convolution neural network (called NL-CNN) where nonlinear convolution is emulated. The ESC-50 environmental sound database with 50 sound categories was considered for parameter tuning and performance evaluation. Despite its low complexity, a reasonable well accuracy is demonstrated (up to 76.5%) close to the human accuracy reported on the same dataset and better in both accuracy and complexity when compared to similar approaches reported in the literature and based on mel-cepstrum. These results provide a basis for a novel and effective method for time-series recognition in general.

A Python Framework for Fast Modelling and Simulation of Cellular Nonlinear Networks and other Finite-difference Time-domain Systems

2021 23rd International Conference on Control Systems and Computer Science (CSCS)

This paper introduces and evaluates a freely available cellular nonlinear network simulator optim... more This paper introduces and evaluates a freely available cellular nonlinear network simulator optimized for the effective use of GPUs, to achieve fast modelling and simulations. Its relevance is demonstrated for several applications in nonlinear complex dynamical systems, such as slow-growth phenomena as well as for various image processing applications such as edge detection. The simulator is designed as a Jupyter notebook written in Python and functionally tested and optimized to run on the freely available cloud platform Google Collaboratory. Although the simulator, in its actual form, is designed to model the FitzHugh Nagumo Reaction-Diffusion cellular nonlinear network, it can be easily adapted for any other type of finite-difference time-domain model. Four implementation versions are considered, namely using the PyCUDA, NUMBA respectively CUPY libraries (all three supporting GPU computations) as well as a NUMPY-based implementation to be used when GPU is not available. The specificities and performances for each of the four implementations are analyzed concluding that the PyCUDA implementation ensures a very good performance being capable to run up to 14000 Mega cells per seconds (each cell referring to the basic nonlinear dynamic system composing the cellular nonlinear network).

Download

NL-CNN: A Resources-Constrained Deep Learning Model based on Nonlinear Convolution

2021 12th International Symposium on Advanced Topics in Electrical Engineering (ATEE), 2021

A novel convolution neural network model, abbreviated NL-CNN is proposed, where nonlinear convolu... more A novel convolution neural network model, abbreviated NL-CNN is proposed, where nonlinear convolution is emulated in a cascade of convolution + nonlinearity layers. The code for its implementation and some trained models are made publicly available. Performance evaluation for several widely known datasets is provided, showing several relevant features: i) for small / medium input image sizes the proposed network gives very good testing accuracy, given a low implementation complexity and model size; ii) compares favorably with other widely known resources-constrained models, for instance in comparison to MobileNetv2 provides better accuracy with several times less training times and up to ten times less parameters (memory occupied by the model); iii) has a relevant set of hyper-parameters which can be easily and rapidly tuned due to the fast training specific to it. All these features make NL-CNN suitable for IoT, smart sensing, biomedical portable instrumentation and other applications where artificial intelligence must be deployed in energy-constrained environments.

Download

Choosing the Most Efficient Classifier for a Fast Logo Recognition System

In this paper we address the challenging problem of logo recognition. As logos appear in various ... more In this paper we address the challenging problem of logo recognition. As logos appear in various scales, colors and with illuminations coming from source lights of various intensities, positions or colors, their recognition is a tedious task. The system is based on a Bag-of-Words framework with Scale Invariant Feature Transform (SIFT) features that are matched using a Random Sample Consensus (RANSAC) homography detection. The logo recognition system is completed by a classifier and we aim to choose the one providing the best compromise between time and quality performance from a set of five hardware-oriented implementations of classifiers in the context of FlickrLogos-32 database.

Download

Disease detection on medical images using light-weight Convolutional Neural Networks for resource constrained platforms

2020 International Symposium on Electronics and Telecommunications (ISETC), 2020

This paper presents a thorough investigation to optimize lightweight convolutional neural network... more This paper presents a thorough investigation to optimize lightweight convolutional neural networks for diseases detection from medical images. Particularly, we are interested in building an efficient recognition convolutional deep learning architecture running on portable, energy efficient, resources constrained platforms, as aids to medical personnel. Several state of the art lightweight deep neural network architectures as well as more complex (for better performance) ones were compared not only from the point of view of recognition accuracy but also from the complexity perspective. An original solution with Android and Raspberry-Pi implementations is proposed to give a good trade-off between complexity, detection accuracy, latency and portability in building intelligent medical instruments. Experiments on the Chest X-ray dataset indicate that our proposed solutions achieves state of the art accuracies with 91.22% less trainable parameters and with 3.62 times less memory usage storage, when compared with other state of the art light weight models.

Download

Optimization of GPU and CPU acceleration for neural networks layers implemented in python

2017 5th International Symposium on Electrical and Electronics Engineering (ISEEE), 2017

Many neural architectures including RBF, SVM, FSVC classifiers, or deep-learning solutions requir... more Many neural architectures including RBF, SVM, FSVC classifiers, or deep-learning solutions require the efficient implementation of neurons layers, each of them having a given number of m neurons, a specific set of parameters and operating on a training or test set of N feature vectors having each a dimension n. Herein we investigate how to allocate the computation on GPU kernels and how to better optimize the problem parameters (neural structure and training set size) as well as the GPU parameters in order to maximize the acceleration (relative to a CPU implementation). It is shown that by maximizing the load (number of threads on each computational GPU core) and by a proper allocation of the GPU global memory, very large speedups (100-250 times) with respect to the CPU implementation can be achieved while using the convenient NUMBA Python package supporting CUDA programming of GPU. Consequently, it is shown that given a problem to be posed to a neural network a convenient decomposition of the network can be done in order to allocate optimally the parts of the computation to the GPU in order to maximize efficiency. Also, for CPU implementations it was found that Intel's MKL library (called from NUMPY package) can offer efficient implementation of neural layers, comparable to what is achieved using GPU.

Download

n+of+two+Cryptographically FPGA Implementation and Evaluation of two Cryptographically Secure Hybrid Cellular Automata DRAFT

Ioana Dogaru and R. Dogaru, "FPGA implementation and evaluation of two cryptographically sec... more Ioana Dogaru and R. Dogaru, "FPGA implementation and evaluation of two cryptographically secure hybrid cellular automata", in Proceedings of the 2014 10th International Conference on Communications (COMM), final version of the paper is available on IEEE Xplore Abstract -Two solutions for designing cryptographically secure pseudo-random number generators are considered from the perspective of FPGA implementation. Both are based on hybrid cellular automata (HCA) and are designed to maximize the efficiency/throughput ratio. While the first PRNG employs a simple HCA, the second solution builds a chain of HCA devices in order to improve cryptographic properties. Both solutions were implemented on FPGA-based platforms and their performances are evaluated in terms of complexity, speed and power. Pseudorandom sequences generated by the hardware implementations were extracted, saved and further analyzed using the standardized NIST statistical tests. All 188 tests were passed.

Download

Predicting Emergence from Cell’s Structure

Systematic Design for Emergence in Cellular Nonlinear Networks, 2008

A Convenient Method for Learning to Classify Biomedical and Other Complex Signals based on Neuronal Boolean Complexity

2019 11th International Symposium on Advanced Topics in Electrical Engineering (ATEE), 2019

This paper proposes a novel method for fast and compact recognition of biomedical and other compl... more This paper proposes a novel method for fast and compact recognition of biomedical and other complex signals. It is essentially based on decomposing the signal sequences to be recognized into binary strings. Then each binary string is divided into smaller strings corresponding to Boolean functions. Based on complexities for neural implementation of Boolean functions a feature vector is computed by counting the occurrences of functions with certain complexities. Using enough labeled signal sequences, a classifier can be easily trained and then used for the automatic recognition of the signals. To decrease costs, a hardware oriented Extreme Learning Machine is used. Herein, in addition to describing the method, we consider the widely used Bonn dataset for epileptic seizure detection to demonstrate that our method offers good accuracies (more than 97%) with the advantage of very low implementation costs. Given its convenience and simplicity, the proposed method can be easily integrated ...

Comparative Study of Extreme Learning Machine using Various Computing Platforms

2019 6th International Symposium on Electrical and Electronics Engineering (ISEEE), 2019

This paper focuses on testing the performance of an optimized extreme learning machine (ELM) algo... more This paper focuses on testing the performance of an optimized extreme learning machine (ELM) algorithm using various resources-constrained computational platforms, to be used in autonomous intelligent vehicles. The purpose of this comparative study is to find the best suited embedded computational platform for future use in an experimental autonomous mobile robot, where computing power, responsiveness and power consumption are key to further development. The aforementioned algorithm was used as a benchmarking tool in this performance review, but also as an important starting point for processing large amounts of different data types from the robots' sensorial units. For now, the conducted experiments are limited to pre-trained datasets, to have a precise output of the very same data type.

A super fast vector support classifier using novelty detection and no synaptic tuning

2016 International Conference on Communications (COMM), 2016

A novel classifier architecture is introduced and its performances are evaluated against state of... more A novel classifier architecture is introduced and its performances are evaluated against state of the art shallow classifiers. Its main advantage consists in a very fast learning ensured by a novelty detection algorithm, selecting a list of prototypes among the training samples, used as centers in a radial basis functions neurons layer. Only the radius of the basis functions is optimized to improve generalization in conjunction with an overlapping parameter and there is no need for synaptic tuning. Compared to state of the art models such as SVM (Support Vector Machine) or ELM (Extreme Learning Machine) our SFSVC (Super Fast Vector Support Classifier) it offers equal performance while having a more compact and fast algorithm. Thus SFSVC is well suited for embedded processing of big collections of data such as data from satellite remote sensing units, automotive sensors etc. with a good potential of being directly integrated into sensing platforms.

Download

Logo localization and recognition in natural images using homographic class graphs

Machine Vision and Applications, 2015

We propose a method for localization and classification of brand logos in natural images. The sys... more We propose a method for localization and classification of brand logos in natural images. The system has to overcome multiple challenges such as perspective deformations, warping, variations of the shape and colors, occlusions, background variations. To deal with perspective variation, we rely on homography matching between the SIFT keypoints of logo instances of the same class. To address the changes in color, we construct a weighted graph of logo interconnections that is further analyzed to extract potentially multiple instances of the class. The main instance is build by grouping the keypoints of the graph connected logos onto the central image. The secondary instance is needed for color inverted logos and is obtained by inverting the orientation of the main instance. The constructed logo recognition system is tested on two databases (FlickrLogos-32 and BelgaLogos), outperforming state of the art with more than 10% accuracy.

Download

Fast Support Vector Classifier for automated content-based search in video surveillance

2015 International Symposium on Signals, Circuits and Systems (ISSCS), 2015

In this article we present and test a specialized classifier, i.e., Fast Support Vector Classifie... more In this article we present and test a specialized classifier, i.e., Fast Support Vector Classifier (FSVC), which is employed for multiple-instance human retrieval in video surveillance. Thanks to its low complexity and high performance in terms of computation and speed, FSVC is adapted to ease the generalization of the feature space using only a limited number of samples in the training process. To validate the performance, FSVC is evaluated on two standard video surveillance datasets. It obtains superior or similar results in terms of precision and recall compared to the close related state-of-the-art Support Vector Machines approaches. I.

Download

Physical modeling of RTD-based CNN cells

Proceedings of the 2000 6th Ieee International Workshop on Cellular Neural Networks and Their Applications (Cnna 2000), 2000

Martin HZnggi, Radu Dogaru, and Leon 0. Chua Nonlinear Electronics Laboratory University of Calif... more

Radu DOGARU

Uploads

Papers by Radu DOGARU

Log In