School of Electrical and Computer Engineering

Permanent URI for this college

http://etd.aau.edu.et/handle/123456789/47

Browse

Now showing 1 - 20 of 724

(7,3) Maximum - Length Binary Cycle Code Applied to Single Channel Digital Communication System for Error Correction
(Addis Ababa University, 1996-06) snegash, Yohanne; Alemu, Ketema(phD)
This paper presents the design and hardware implementation of the (7,3) maximum-length binary cyclic code applied to a single channel communication system. In hardware implementation of the system a PC with a data acquisition board with time <- sharing for interfacing the analog signals. A 12-bit digital output of the PC is divided into blocks of 3-bits for processing by the channel encoder The implementation of the system is carried out using shift registers and logic gates. A sinusoidal input waveform is applied to the system input and a circuit designed with a combination of D-type flip-flops and logic gates is used to introduce the effects of a single-random-error and a double-adjacent-burst error to observe the performance of the system. The code resulted in good performance in correcting a single-random-error and a double-adjacent-burst-error.
A Video Coding Scheme Based on Bit Depth Enhancement With CNN
(Addis Ababa University, 2023-06) Daniel Getachew; Bisrat Derebssa (PhD)
Raw or uncompressed videos take a lot of resources in terms of storage and bandwidth. Video compression algorithms are used to reduce the size of a video and many of them have been proposed over the years. People also proposed video coding schemes which works on top of existing video compression algorithms by applying down sampling prior to encoding and restoring them to their original form after decoding for further bitrate reduction. Down sampling can be done in spatial resolution or bit depth. This paper presents a new video coding scheme that is based on bit depth down sampling before encoding and use CNN to restore it at the decoder. However unlike previous approaches the proposed approach exploits the temporal correlation which exists between consecutive frames of a video sequence by dividing the frames into key frames and non-key frames and only apply bit depth down sampling to the non-key frames. These non-key frames will be reconstructed using a CNN that takes the key frames and non-key frames as input at the decoder. Experimental results showed that the proposed bit depth enhancement CNN model improved the quality of the restored non-key frames by an average of 1.6dB PSNR than the previous approach before integrated to the video coding scheme. When integrated in the video coding scheme the proposed approach achieved better coding gain with an average of -18.7454% in Bjøntegaard Delta measurements.
Accelaration of Preprocessors of the Snort Network Intrusion Detection System Using General Purpose Graphics Processing Unit
(Addis Ababa University, 2015-04) Yihunie, Simegnew; Assamnew, Fitsum
Advances in networking technologies enable interactions and communications at high speeds and large data volumes. But, securing data and the infrastructure has become a big issue. Intrusion Detection Systems such as Snort play an important role to secure the network. Intrusion detection systems are used to monitor networks for unauthorized access. Snort has a packet decoder, pre-processor, detection engine and an alerting system. The detection engine is the most compute intensive part followed by the pre-processor. Previous work has shown how general purpose graphics processing units(GP-GPU) can be used to accellerate the detection engine. This work focused on the pre-processors of Snort, speci cally, the stream5 pre-processor as pro ling revealed it to be the most time consuming of the pre-processors. The analysis shows that the individual implementation of stream5 using Compute Uni ed Device Architecture(CUDA) achieved up to ve times speed up over the baseline. Also, an over all 15.5 percent speed up on the Defense Advanced Research Projects Agency(DARPA) intrusion detection system dataset was observed when integrated in Snort. Key words: Intrusion Detection System, Snort, Graphics Processing Unit, CUDA, Parallelization, Porting, Preprocessor.
Acceleration and Energy Reduction of Object Detection on Mobile Graphics Processing Unit
(Addis Ababa University, 2019-06) Fitsum, Assamnew; Jonathan, Rose (Prof.); Dereje, Hailemariam (PhD)
The evolution of high performance computing in today’s smartphones is enabling their use in compute-intensive applications. As the compute requirement increases, the energy required to do the computation cannot increase in proportion because the cost of providing that energy available and cooling would become prohibitive. An alternative, potentially power-reducing approach is to use graphics processing units or special accelerator cores. Today’s smartphones are equipped with systemon-chip (SoC) devices that house many cores such as graphics processing units, digital signal processors, and special multimedia encoder/decoder hardware along side multi-core central processing units. Their inclusion enables applications that require greater computational power such as real-time computer vision. In this work, we study the capability of the recently introduced general-purpose graphics processing unit (GPU) in a smartphone SoC to enable energy-efﬁcient object detection. This will include understanding the architecture of the recent GPUs that will be used (the Adreno 320 and Adreno 420 from Qualcomm), the implementation and optimization of the object detection algorithm used in the Open Computer Vision library (OpenCV) using these GPUs and measuring the energy consumption of this implementation. We implemented the Viola-Jones based object detection on the GPU in an Android tablet. The implementation is 35% faster on average than the same algorithm running on the CPU on the same device. The implementation also reduces the average energy consumption by 68% compared to the CPU on the same device. An application that utilized the object detector on the mobile GPU to detect Ringworm skin disease was developed. A classiﬁer was trained for this application and it has an accuracy of 75%.
Acceleration of Convolutional Neural Network Training using Field Programmable Gate Arrays
(Addis Ababa University, 2022-01) Guta, Tesema; Fitsum, Assamnew (PhD)
Convolutional neural networks (CNN) training often necessitates a considerable amount of computational resources. In recent years, several studies have proposed CNN inference and training accelerators, which the FPGAs have previously demonstrated good performance and energy efficiency. To speed processing, the CNN requires additional computational resources such as memory bandwidth, a FPGA plantform resource usage, time, and power consumption. As well as training the CNN needs large datasets and computational power, and they are constrained by the requirement for improved hardware acceleration to support scalability beyond existing data and model sizes. In this study, we propose a procedure for energy efficient CNN training in collaboration with an FPGA-based accelerator. We employed optimizations such as quantization, which is a common model compression technique, to speed up the CNN training process. Additionally, a gradient accumulation buffer is used to ensure maximum operating efficiency while maintaining gradient descent of the learning algorithm. Subsequently, to validate our design, we implemented the AlexNet and VGG16 models on an FPGA board and a laptop CPU and GPU. Consequently, our designs achieve 203.75 GOPS on Terasic DE1-SoC with the AlexNet model and 196.50 GOPS with the VGG16 model on Terasic DE-SoC. This, as far as we know, outperforms existing FPGA-based accelerators. Compared to the CPU and GPU, our design is 22.613X and 3.709X more energy efficient respectively.
Active Power Flow Control in Ethiopian High Voltage Transmission Networks Using Phase Shifting Transformer to Enhance Utilization of Transmission Lines
(Addis Ababa University, 2018-06) Yemane, Esayas; Fekadu, Shewarega (PhD)
The electricity supply industry of Ethiopia is undergoing a major transformation that requires a redefined approach to increase the utilization of existing transmission line assets. Overloading of transmission lines in a power system sometimes result stability issues, which may lead to unwanted tripping or failure of equipments. The cause could be uneven loading of interconnectors or parallel transmission lines in meshed networks due to different impedances caused by the tower geometry, conductor sizing, number of sub-conductors and line length. Under these conditions, to ensure economical and reliable operation of the grid, active power flow through the lines should be controlled within their capability limits. In view of above, the power flow needs to be controlled in order to enhance utilization of high voltage transmission lines and secure the power system. Thus Control of power in AC network requires special technology to be implemented on case to case basis. Operating efficiency of electric transmission system can be improved by using appropriate Flexible Alternating Current Transmission System (FACTS) devices. Phase shifting transformer is one of the FACTS families, which can be used for power control in ac network. This thesis presents a study on active power flow control within Ethiopian network for optimum utilization of transmission lines using phase shifting transformer (PST). The study is performed first by reviewing literatures on the use of phase shifting transformers how to redirect active power flow in transmission networks throughout the world. To demonstrate the active power flow control in the network, a 400/400 kV phase shifting transformer having a size of 685 MVA with a phase shifting angle range of -200 to+200 and the high voltage transmission networks was modeled using PSSE software(Power System Simulation for Engineers) for the peak load of 2040 MW in the year 2017. From the power flow studies/solution, various overloaded and under loaded transmission lines are identified. By varying the phase angle of the phase shifting transformer, several simulations are conducted to investigate the impact of PST on the active power flow distribution. In this study, it has been demonstrated that the active power flow patterns which originally flow via the low impedance and lower voltage system is fully controlled and restructured using phase shifting transformer. By varying the phase shifting transformer angle, the active power flow in the transmission lines can be redirected towards the alternate high voltage path. As the Phase shifting transformer angle increased from -20° to +20°, the loading of Wolayta - Gibe II and Sebeta IIGibe II 400kV transmission lines vary from 4% to 35% and 11% to 42% respectively. Similarly, Gelan - Wolayta400kV transmission line load increases from 15% to 43% as the Phase shifting transformer angle decreases from +20° to -20° Conventional ways of solving the network bottlenecks based on reinforcement and building new transmission lines cannot be taken as sufficient and fast due to the problems of acquiring new corridors and environmental limitations. Installation of Phase Shifting Transformer in the transmission network is a better solution for controlling the active power flow and effective utilization of existing high voltage transmission network assets.
Adaptive Antenna Array Algorithms and Their Impact on Code Division Multiple Access Systems (CDMA).
(Addis Ababa University, 2004-03) Hadgu, Dereje; Abdo, Mohammed (PhD)
In mobile communications there is a need to increase the channel capacity. The increasing demand for mobile communication services without a corresponding increase in RF spectrum allocation (channel capacity) motivates the need for new techniques to improve spectrum utilization. The CDMA and adaptive antenna array are two approaches that shows real promise for increasing spectrum efficiency. This research focuses on the application of adaptive arrays to the Code Division Multiple Access (CDMA) cellular systems. The adaptive antenna has an intelligent control unit, so the antenna can follow the user, direct the radiation pattern towards the desired user, adapt to varying channel conditions and minimize the interference. Therefore there can be several users in the same channel in the same cell. The driving force of this intelligent control unit are special kinds of algorithms and we are going to investigate the performance of these different adaptive array algorithms in the CDMA systems. Four each blind adaptive array algorithms are developed, and their performance under different test situations (e. g. AWGN (Additive White Gaussian Noise) channel, and multipath environment) is studied. A MATLAB test bed is created to show their performance on these two test situations and an optimum one can be selected.
Adaptive Control of Multi-layer Switched Reluctance Motor
(Addis Ababa University, 2018-12) Alem, Gebreziher; Mengesha, Mamo (PhD)
The multi-layer switched reluctance motor (MSRM) are receiving significant attention from industries because of its simple structure, inexpensive manufacturability and reliability. In addition multi-layer switched reluctance motor receiving renewed attention as a viable candidate for various adjustable speed and high torque applications such as in the automotive, traction and aerospace industries. Simple power electronic drive circuit and fault tolerance of converter are specific advantages of multi-layer switched reluctance motor drives, but excessive torque ripple has limited its use to special applications. It is well known that controlling the current adequately can minimize the torque ripple because current is directly proportional to torque. The magnetization characteristics of the SRM is highly non-linear making the flux linkage and torque as the non-linear functions of both the current and rotor position. Establishing this high precision nonlinear mapping between current and rotor position is used to control the motor accurately for the analysis and control of any switched reluctance motor system. The generating or motoring mode of operation of the motor depends greatly on the value of rising or falling torque and hence it needs to be controlling more accurately the torque ripples for the practical applications. This thesis investigates the use of fuzzy logic controller and a hybrid intelligent system which is adaptive neuro fuzzy inference system (ANFIS) to reduce the torque ripples of multi-layer switched reluctance motors. Matlab simulink models of multi-layer switched reluctance motors with fuzzy logic controller and adaptive neuro fuzzy inference system (ANFIS) are developed to carry out simulation studies under loaded conditions. A comparison results shows that with fuzzy logic controller, the torque ripple is reduced by twenty two percent (22%) as compared to that without any controller. It is further observed that the adaptive neuro fuzzy inference system (ANFIS) controller reduces the torque ripples by twenty six percent (26%) as compared to that without any controller. This clearly shows that the torque ripple is reduced by using fuzzy logic controller as well the adaptive neuro fuzzy inference system (ANFIS). Moreover, performance of the adaptive neuro fuzzy inference system is better because it includes learning mechanism to adapt itself to new dynamic conditions. Key words: Multi-layer switched reluctance motor, fuzzy logic controller, adaptive neuro fuzzy inference system, torque ripples
Adaptive modulation based cooperative MIMO in fading channel for future wireless technology
(Addis Ababa University, 2017-03) Ahmed, Niema; Ridwan, Murad (PhD)
With the rapid growth of multimedia services, future generations of cellular communications require higher data rates and a more reliable transmission links while keeping satisfactory quality of service. The data rate and reliability of wireless communication links can be improved by employing multiple antennas at both ends, thereby creating Multiple-Input Multiple-Output (MIMO) channels. However, the use of multiple antennas in mobile terminals may not be very practical. Certainly there is limited space and other implementation issues which make this a challenging problem. Therefore, to harness the diversity gains order by MIMO transmitter diversity techniques, while maintaining a minimal number of antennas on each handset, cooperative diversity techniques have been proposed. The main drawback of cooperative diversity is the throughput loss due to the extra resources needed for relaying. Therefore, cooperative MIMO together with adaptive modulation is used to meet the demands for high data rate and transmission reliability. This thesis presents performance analysis of a cooperative MIMO schemes with adaptive modulation for different detection techniques in Long Term Evolution network. In this scheme, each link uses MIMO Vertical Bell-Labs Layered Space Time architecture over Rayleigh flat fading channels and the cooperation strategy uses amplify and forward protocol with one relay node. For cooperative and non-cooperative MIMO, the SNR criterions to switch from one modulation order to the next for attaining maximum spectral efficiency subject (SE) at a target bit-error rate are determined. The simulation results shown that the cooperative MIMO system with adaptive modulation not only compensate for the throughput loss but also achieve considerable throughput gain compared with fixed modulation at comparable complexity. The switching criterion of optimal schemes for adaptive modulation of cooperative hybrid network with minimum mean square error (MMSE) detection, as it has a lower complexity compared to maximum likelihood (ML) detection, is also investigated. As an example, in the downlink scenario adaptive modulation based cooperative and non-cooperative MIMO network have shown optimal SE performance for and for respectively, while satisfying target BER constraint, . Key words: Adaptive Modulation, Cooperative Diversity, MIMO, LTE, SNR, Spectral Efficiency
Adaptive Radial Basis Function Neural Network Based Hierarchical Sliding Mode Controller for 2-Dimensional Double Pendulum Overhead Crane
(Addis Ababa University, 2024-01) Wosene Yirga; Dereje Shiferaw (PhD)
Several control methods for an overhead crane modeled as a double pendulum with constant cable length have been published in various studies. Most of the proposed control methods were open-loop and linear control methods or nonlinear control methods that fully depended on the system model.However, the dynamic of an overhead crane is a complex nonlinear function of uncertain or unknown parameters, which reduces the performance of such control methods. In this thesis, an adaptive radial basis function neural network-based hierarchical sliding mode controller (ARBFNNHSMC) is designed to control a 2-dimensional overhead modeled as a double pendulum system with variable cable length using the Lagrange equation of motion. To reduce the chattering effect of the sliding mode controller as well as increase its robustness, ARBFNN is designed to estimate unknown or uncertain nonlinear functions in the system. The overall control law, which contains only some parts of the crane model, is designed, and the adaptation law is derived from the Lyapunov stability condition to update the weight of the network based on observed errors. The proposed control strategy and derived model are verified using MATLAB/Simulink software.For the same controller parameters,500% changes in model parameters are taken, and trolley displacement settling time and rising time for HSMC are 12.3 seconds and 6.95 seconds, respectively. On the other hand, the maximum hook’s and payload’s swing angles are around 1.34 deg and 1.9 deg for HSMC, and it is around 1.04 deg and 1.64 deg for ARBFNN-HSMC. The residual hook’s and payload’s swing angles are 0.0137 deg and -0.0319 deg, respectively, in the case of HSMC and -0.0011 deg and -0.0022 deg for ARBFNN-HSMC. This numerical result shows that ARBFNN-HSMC has better performance than HSMC for large parameter variations. In addition, the controller output of ARBFNN-HSMC is smoother than that of HSMC, as evidenced by the result.
Adaptive Super Twisting Sliding Mode Controller Design of Quadcopter for Wheat Disease Detection
(Addis Ababa University, 2023-11) Nardos Belay; Lebsework Negash (PhD)
Brown wheat rust is a fungal disease that can cause huge destruction in wheat production and quality. Collecting accurate large scale crop data and detecting these diseases based on certain standards through visual inspection is labor intensive, time consuming, and prone to human error. This paper focuses on the design of adaptive super twisting sliding mode controller of a quadcopter for detection of brown wheat rust disease. First, the dynamics of the system was understood then the Newton-quaternion approach was used to model the dynamic system and verified in simulink. Then, the adaptive super twisting sliding mode controller was developed for attitude and position trajectory tracking of a quadrotor. Controller design involves tuning the parameters of the supertwising sliding mode controller using adaptation laws. Comparison of conventional sliding mode controller with the adaptive super twisting sliding mode controller was analyzed. The effectiveness of the proposed control scheme has been verified by developing simulation results for quadcopter in MATLAB/SIMULINK software. The results show high tracking accuracy, chattering reduction, and disturbance rejection capability of the proposed controller. For the task of brown wheat rust detection, transfer learning technique was applied using the state of the art ResNet152v2 model to perform feature extraction for the convolutional neural network architecture. The trained model achieved an accuracy level of 93.28% in the training phase and 92% in the test set.
Addressing User Cold Start Problem in Amharic YouTube Advertisement Recommendation Using BERT
(Addis Ababa University, 2024-06) Firehiwot Kebede; Fitsum Assamnew (PhD)
With the rapid growth of the internet and smart mobile devices, online advertising has become widely accepted across various social media platforms. These platforms employ recommendation systems to personalize advertisements for individual users. However, a significant challenge for these systems is the user cold-start problem, where recommending items to new users is difficult due to the lack of historical preference of the user in a content-based recommendation system. To address this issue we propose an Amharic YouTube advertisement recommendation system for unsigned YouTube users where there is no user information like past preference or personal information. The proposed system uses content-based filtering techniques and leverages Sentence Bidirectional Encoder Representations from Transformers (SBERT) to establish sentence semantic similarity between YouTube video titles, descriptions, and advertisement titles. For this research, 4500 data were collected and preprocessed from YouTube via YouTube API, and 500 advertisement titles from advertising and promotional companies. Random samples from these datasets were annotated for evaluation purposes. Our proposed approach achieved a 70% accuracy in recommending semantically related Amharic Advertisements (Ads) to corresponding YouTube videos with respect to the annotated data. At a 95% confidence interval, our system demonstrated an accuracy of 58% to 76% in recommending Ads which are relevant to new users who have no prior interaction history on the platform with the Ads. This approach significantly enhances privacy by reducing the need for extensive data sharing.
Adoption of Electric Vehicle and Efficiency Improvement of Storage System in Addis Ababa
(Addis Ababa University, 2021-10) Benyam, Girma; Mengesha, Mamo (PhD)
This thesis contributes to the problem description of the impact of geographical landscape and road dynamics in the process of adoption of electric vehicles in the case of Addis Ababa/Ethiopia. The impact of Addis Ababa road dynamics and topographic distribution of the city have been investigated and it was compared with various international drive cycles like EUDC, NYDC, and WLTP. By considering the worst possible scenario for modelling electric vehicle dynamics and energy storage system; this thesis provides an alternative engineering solution using battery and ultra capacitors. The investigation was done concerning electric vehicle regenerative braking energy gain possibilities and the magnitude of energy consumption from an energy storage system. From the simulation result, it was found that the magnitude of the maximum acceleration-deceleration was about 1.57m/s² and 2.81m/s² respectively and the frequency of stop time was around 31 with in the city. In addition to that, because of rugged nature of the topography; the tractive and regenerative energy consumption of the city was about 6637KJ and -281KJ respectively. It was observed that a larger magnitude of acceleration-deceleration rate and regeneration braking energy has been recorded compared to any other international city. Due to this nature of the city, 8 % state of charge battery variation has been found within 400 seconds of simulation time. From the finding, due to the nature of the city; battery Energy storage system is less efficient when compared to hybrid energy storage system hence electric vehicle implemented in the city of Addis Ababa/Ethiopia need to be redesigned. This thesis recommends fuzzy logic control based battery and ultra capacitor hybrid energy storage system which consider topographic distribution and road dynamics of the city. Adopting electric vehicles without considering the above issue may lead to performing under the manufacturer's specification. Specifically, degradation of batteries and reduced use of the range of travel per single charge.
AI-Based Mobile Robot for Agricultural Application using Sliding Mode Controller
(Addis Ababa University, 2023-05) Zewdu Jemema; Dereje Shiferaw (PhD)
Recent advancements in agricultural robotic systems have greatly enhanced their functionality, usability, and integration into various tasks, particularly in the field of agriculture. The primary goal of designing agricultural robots is to enhance efficiency, save time, and decrease production costs by incorporating controllers, sensors, actuators, and communication systems. These robots have versatile applications and are widely embraced in the agricultural sector in industrialized countries. Extensive research has been dedicated to developing mobile robot platforms tailored for agricultural tasks, including plant health monitoring, pesticide spraying, fruit picking, and harvesting, with the aim of supporting farmers in developing nations like Ethiopia, where approximately 67% of the population is involved in agriculture. The thesis specifically targets fruit harvesting and focuses on the challenging task of modeling, designing, and simulating a mobile manipulator with advanced capabilities for agricultural businesses, making it one of the most difficult undertakings in this field. The study encompasses the presentation of the mobile manipulator’s 3D design, kinematics, and dynamics. In addition, AI techniques are employed to analyze fruit images, facilitating the accurate detection and determination of fruits. Based on the results, the effectiveness of the training technique has been assessed using an RMSE value of 0.19 and a loss value of 3.6e-02. An SMC utilizes the input generated from the image to govern the mobile manipulator’s position. The system’s stability and robustness have been assessed by considering uncertainties and variations in mass. When comparing the performance of a designed controller with a PID controller in the presence of uncertainty and parameter variation, it was found that SMC outperformed. According to the evaluation using the ITAE, SMC proves to be more effective, demonstrating a 75% improvement compared to the PID controller. Overall, this research contributes to the development of a robust and intelligent mobile manipulator for fruit harvesting in the agricultural sector, with potential applications to support farmers in countries like Ethiopia.
Alarm Prediction for Fault Management using Deep Learning Approach: The Case of Ethio Telecom
(Addis Ababa University, 2021-09) Betelhem, Berhanu; Rosa, Tsegaye (PhD)
Telecommunication networks play a critical role in our society. They make it possible to share a massive amount of data across the globe. While networks are complex systems in terms of size and technological diversity, a failure results in numerous alarms from the series of devices. This makes monitoring and maintenance activities challenging due to the growing complexity of alarm management systems and the need for highly educated experts to deploy. Alarms are generated in vast quantities every day by today's large and complicated telecom networks. The alarm sequence offers significant information on the network's activity, but most of it is fragmented and hidden in the massive amount of data. Alarm regularities can be utilized in fault management systems, for example, to filter redundant alarms, locate network problems, and even anticipate catastrophic faults. In the presence of flooding alarms, alarms that are inadequately configured and maintained, and a large number of nuisance alarms, operators are expected to make vital judgments. If the incoming alarms can be correctly predicted before they occur, the operators may be able to address and possibly avoid anomalous behaviors by taking corrective actions in a timely manner. This paper presents an alarm prediction method based on data mining to generate patterns from historical alarm data, and use such patterns to train three deep learning approaches, namely long short-term memory (LSTM), bidirectional LSTM (Bi-LSTM) and gated recurrent unit (GRU). The prediction performance of the three deep learning approaches has been compared. Domain trained word embedding and pretrained word embedding (Word2vec) are used to feed embeddings to neural networks. The relevance of applying different word embedding is to explore the effect of the data preparation on the model performance. The Frequent pattern growth (FP) algorithm implemented in Rapidminer studio has been used to mine five months' worth of alarm logs. Finally, the best performing model is selected based on the accuracy of the model. The models are tested with a sequence of alarms and Bi-LSTM with domain trained word embedding achieves 93% in predicting the target alarms. However, from the results, we can also say that all three deep learning approaches can be used for predicting telecom alarms.
Amharic Hateful Memes Detection on Social Media
(Addis Ababa University, 2024-02) Abebe Goshime; Yalemzewud Negash (PhD)
Hateful meme is defined as any expression that disparages an individual or a group on the basis of characteristics like race, ethnicity, gender, sexual orientation, country, religion, or other characteristics. It has grown to be a significant issue for all social media platforms. Ethiopia’s government has increasingly relied on the temporary closure of social media sites but such kind of activity couldn’t be permanent solution so design automatic system. These days, there are plenty of ways to communicate and make conversation in chat spaces and on social media such as , text, image, audio, text with image, and image with audio information. Memes are new and exponentially growing trend of data on social media, that blend words and images to convey ideas. The audience can become dubious if one of them is absent. Previous research on the identification of hate speech in Amharic has been primarily focused on textual content. We should design deep learning modal which automatically filter hateful memes in order to reduce hate content on social media. The basis of our model consists of two fundamental components. one is for textual features and the other is for visual features. For textual features, we need to extract text from memes using optical character recognition (OCR). The extracted text through the OCR system is pixel-wise, and the morphological complex nature of Amharic language will affect the performance of the system to extract incomplete or misspelled words. This could result in the limited detection of hateful memes. In order to work effectively with an OCR extracted text, we employed a word embedding method that can capture the syntactic and semantic meaning of a word. LSTM is used for learning long-distance dependency between word sequence in short texts. The visual data was encoded using an ImageNet-trained VGG-16 convolutional neural network. In the studies, the input for the Amharic hateful meme detection classifier combines textual and visual data. The maximum precision was 80.01 percent. When compared to state-of-the-art approaches using memes as a feature on CNN-LSTM, an average F-score improvement of 2.9% was attained.
Amharic Named Entity Recognition Using Neural Word Embedding as a Feature
(AAU, 2017-10) Dagimawi, Demissie; Surafel, Lemma (PhD)
In this paper, Amharic Named Entity Recognition problem is addressed by employing a semi-supervised learning approach based on neural networks. The proposed approach aims at automating manual feature design and avoiding dependency on other natural language processing tasks for classi cation features. In this work potential feature information represented as word vectors are generated using neural network from unlabeled Amharic text les. These generated features are used as features for Amharic Named entity classi cation. SVM, J48, random tree, IBk(Instance based learning with parameter k), attribute selected and OneR(one rule) classi ers are tested with word vector features. Additionally BLSTM(bi-directional long short term memory), LSTM(long short term memory) and MLP(multi layer perceptron) deep neural networks are also tested to investigate the impact of proposed approach. From the experiments the highest F-score achieved was 95.5% using the SVM classi er. Relative to state-of-the-art approaches (SVM and J48) an average F-score improvement of 3.95% was achieved. The results showed that, automatically learned word features can substitute manually designed features for Amharic named entity recognition. Also these features has given better performance while reducing the e ort in manual feature design.
Amharic Parts-of-Speech Tagger using Neural Word Embeddings as Features
(Addis Ababa University, 2019-01) Mequanent, Argaw; Surafel, Lemma (PhD)
The parts-of-speech (POS) tagging for Amharic language is not matured yet to be used as one important component in other natural language processing (NLP) applications. Previous studies done on Amharic POS tagger used hand-crafted features to develop tagging models. In Amharic language, prepositions and conjunctions usually are attached with the other parts-of-speech. This forces the tags to represent more than one basic information and also decrease the total number of instances in the training corpus. In addition, the manual design of features requires longer time, more labor and linguistic background. In this study, automatically generated neural word embeddings are used as features for the development of an Amharic POS tagger. Neural word embeddings are multi-dimensional vector representations of words. The vector representations capture syntactic and semantic information about words. Another additional aspect in this study is, prepositions and conjunctions attached with the other parts-of-speech are segmented using HornMorpho morphological analyzer. Stateof- the-art deep learning algorithms are also used to develop tagging models. Long Short-Term Memory (LSTM) recurrent neural networks and their bidirectional versions (Bi-LSTM RNNs) are used to develop tagging models from the possible deep learning algorithms. The maximum evaluation result observed is 93.67% F-measure obtained from the model developed by using Bi-LSTM recurrent neural network. From the results obtained, it can be observed that word embeddings generated by neural networks can replace manually designed features which is an important advantage. Segmenting prepositions and conjunctions attached with the other parts-of-speech also improved the accuracy of the POS tagger by more than 5%. The accuracy improvement of the POS tagger is obtained from the increased total number of instances and decreased number of tags due to segmentation.
Amharic Sign Language Recognition based on Amharic Alphabet Signs
(Addis Ababa University, 2018-03-16) Nigus, Kefyalew; Menore, Tekeba (Mr.)
Sign language is a natural language mostly used by hearing impaired persons to communicate with each other. At present day, sign language explainers are used to eliminate the language obstacles between people who are hearing impaired and non-impaired one. However, they are very limited in number. So, automatic sign language recognition system is better to narrow the communication gap between hearing impaired and normal people. This thesis work dealts with development of automatic Amharic sign language translator, translates Amharic alphabet signs into their corresponding text using digital image processing and machine learning approach. The input for the system is video frames of Amharic alphabet signs and the output of the system is Amharic alphabets. The proposed system has four major components: preprocessing, segmentation, feature extraction and classification. The preprocessing starts with the cropping and enhancement of frames. Segmentation was done to segment hand gestures. A total of thirty-four features are extracted from shape, motion and color of hand gestures to represent both the base and derived class of Amharic sign characters. Finally, classification models are built using Neural Network and Multi-Class Support Vector Machine. The performance of each models, Neural Network (NN) and Support Vector Machine (SVM) classifiers, are compared on the combination of shape, motion and color feature descriptors using ten-fold cross validation. The system is trained and tested using a dataset prepared for this purpose only for all base characters and some derived characters of Amharic. Consequently, the recognition system is capable of recognizing these Amharic alphabet signs with 57.82% and 74.06% by NN and SVM classifiers respectively. Therefore, the classification performance of Multi-Class SVM classifier was found to be better than NN classifier.
Amharic Speech Recognition System Using Joint Transformer and Connectionist Temporal Classification with External Language Model Integration
(Addis Ababa University, 2023-06) Alemayehu Yilma; Bisrat Derebssa (PhD)
Sequence-to-sequence (S2S) attention-based models are deep neural network models that have demonstrated some tremendously remarkable outcomes in automatic speech recognition (ASR) research. In these models, the cutting-edge Transformer architecture has been extensively employed to solve a variety of S2S transformation problems, such as machine translation and ASR. This architecture does not use sequential computation, which makes it different from recurrent neural networks (RNNs) and gives it the benefit of a rapid iteration rate during the training phase. However, according to the literature, the overall training speed (convergence) of Transformer is relatively slower than RNN-based ASR. Thus, to accelerate the convergence of the Transformer model, this research proposes joint Transformer and connectionist temporal classification (CTC) for Amharic speech recognition system. The research also investigates an appropriate recognition units: characters, subwords, and syllables for Amharic end-to-end speech recognition systems. In this study, the accuracy of character- and subword-based end-to-end speech recognition system is compared and contrasted for the target language. For the character-based model with character-level language model (LM), a best character error rate of 8.84% is reported, and for the subword-based model with subword-level LM, a best word error rate of 24.61% is reported. Furthermore, the syllable-based end-to-end model achieves a 7.05% phoneme error rate and a 13.3% syllable error rate without integrating any language models (LMs).

Browse

Browsing School of Electrical and Computer Engineering by Title

Results Per Page

Sort Options