Department of Computer Science

Permanent URI for this collectionhttps://repository.mu.edu.et/handle/123456789/147

Browse

Search Results

Now showing 1 - 4 of 4
  • No Thumbnail Available
    Item
    DEVELOP A Bi-DIRECTIONAL ENGLISH - NUER MACHINE TRANSLATION USING DEEP LEARNING APPROACH
    (Mekelle University, 2024-12-28) Lemlem Gebremedhin
    The advancement of deep learning has revolutionized natural language processing, with machine translation playing a pivotal role in bridging linguistic barriers. This research focuses on developing a bi-directional English-Nuer machine translation system using deep learning techniques. The primary challenge is the lack of linguistic resources for the Nuer language, hindering its technological representation and global accessibility. To address this, the study constructed a parallel corpus of 46,134 English-Nuer sentence pairs and employed models such as GRU, Bi-GRU, LSTM, LSTM with attention and transformer mechanisms. The findings revealed that the Transformer model achieved superior BLEU scores compared to the other architectures, scoring 0.2567 for Nuer-to-English and 0.2431 for English-to-Nuer translations. The results highlight the potential of the proposed deep learning-based machine translation for low-resource languages. As future work, the researcher highlights to explore integrating speech-to-text and textto-speech capabilities to enhance usability.
  • No Thumbnail Available
    Item
    Enhanced Solar Irradiation Forecasting Using LSTM
    (Mekelle University, 2025-03-12) Silas Gebretsadik Mebrahtom
    Accurate forecasting of solar irradiation is critical for improving grid stability, optimizing energy storage, and maximizing photovoltaic (PV) system efficiency. Traditional forecasting methods, including statistical and numerical weather prediction models, often struggle with the nonlinear and complex nature of solar irradiation data. This research work explores the application of deep learning techniques, specifically Long Short-Term Memory (LSTM) neural networks, to enhance day-ahead solar irradiation forecasting for solar energy applications. In this research, an LSTM-based model was developed and trained using historical irradiation data, covering the period from January 1984 to March 2025 in Mekelle. The methodology involved data collection, cleaning, and preprocessing, followed by model training and evaluation. Key preprocessing steps included removing anomalies and normalizing the data set using Min-Max scaling. The LSTM model demonstrated superior performance compared to traditional machine learning models, with a Mean Bias Error (MBE) of 0 kWh/m²/day, Mean Absolute Error (MAE) of 0.46 kWh/m²/day, and Root Mean Squared Error (RMSE) of 0.65 kWh/m²/day. The model demonstrated a 71% lower RMSE in winter compared to summer and achieved 56% higher accuracy on sunny days than on cloudy days. These improvements can be attributed to the more stable weather conditions of the winter season and the consistency of solar irradiation on sunny days in Mekelle. The model was evaluated using two different dataset sizes, 5556 and 15040 data points. With the larger dataset, performance improved by 15%, highlighting the critical role of data availability in enhancing model accuracy and reliability. The results suggest the potential of LSTM networks in providing reliable day-ahead solar irradiation forecasts, contributing to the broader adoption of renewable energy. This study recommends integrating the solar irradiation forecasting model with energy forecasting systems to optimize grid performance and storage utilization. Future research should explore extended datasets, additional meteorological parameters, and ensemble methods to enhance the adaptability and accuracy of LSTM-based forecasting models
  • No Thumbnail Available
    Item
    Enhanced Inception ResNet V2 Model for Grape Leaf Disease Detection and Classification
    (Mekelle University, 2025-01-25) Efrem Gebrewahd Gebreslassie
    Grapes are one of the most widely consumed and globally traded crops, benefiting both agricultural economy and healthy wise. However, their vulnerability to diseases can negatively affect the quality and quantity of the grapes being produced. The most common way of detecting and classifying these disease is through the use of human experts (manual inspection method), such as pathologists and botanists. This manual inspection method is prone to error, time consuming and inefficient for large scale farms. To tackle this problem several researchers have built an automated system that detects and classifies plant diseases in a more accurate and faster way. While these studies show a reasonable result but still struggle with several issues, like poor accuracy, increased computational complexity and strong overfitting. Thus, our model addressed these issues by introducing an enhanced version of Pretrained Inception ResNet V2 architecture, By integrating a lightweight Reduction Module C which is responsible for reducing the grid size of the input image from 8 X 8 to 3 X 3 while increasing the feature maps from 1536 to 1600. This module allows the model to capture more complex features while ignoring relevant information leading to improved feature extraction capability. The proposed model achieved a superior F1 score of 99.89% on the validation set with only 0.21 million trainable parameters, outperforming EfficientNet-B4 (99.81% F1 score, 0.46 million parameters), Xception (99.73% F1 score, 1.05 million parameters), and the baseline Inception ResNet V2 (99.87% F1 score, 0.79 million parameters) in both accuracy and computational efficiency. The results show that the suggested model presents a promising solution for accurate and efficient grape leaf disease detection and classification.
  • No Thumbnail Available
    Item
    Multimodal GPS Navigation System for Visually Impaired Users with Tigrinya Voice Commands
    (Mekelle University, 2025-02-25) Brkti Hadush Meche
    This thesis addresses the challenges faced by visually impaired individuals (VII), particularly in regions like Ethiopia, where traditional navigation aids such as white canes provide limited independence. Leveraging the widespread adoption of smartphones with GPS technology, this research introduces a mobile application that integrates Tigrinya voice commands, multimodal feedback (audio and haptic), and emergency support to enable autonomous navigation. By combining GPS-based turn-by-turn guidance with localized speech recognition, the system offers a culturally relevant and accessible solution tailored to low vision Tigrinya speakers. Rigorous development and user testing demonstrate its effectiveness, with initial results showing high accuracy in voice recognition and improved user mobility. The smartphone-based approach ensures broader accessibility compared to embedded systems, while the modular design allows for future scalability to other languages and regions. This work advances assistive technology by providing a practical, user-centric solution that enhances independence for VII