Project Report

Browse

Search Results

Now showing 1 - 10 of 333

Reasoning Over Context in Bangla: A Generative QA Approach to Factoid Understanding Using LLM
(Daffodil International University, 2025-01-14) Rana, Md. Masud
This paper presents the development of a Bangla Question Answering (QA) system using advanced transformer-based models to tackle the complexities of Bangla language processing. Specifically, it compares the performance of BanglaT5, a model fine-tuned for Bangla, with mT5, a multilingual variant of the T5 model. Both models were evaluated on a dataset of over 7,500 Bangla news articles, focusing on factoid-based question answering. The results show that BanglaT5 outperforms mT5 on key metrics such as ROUGE, BLEU, Character Error Rate (CER), and Word Error Rate (WER), showcasing its superior ability to handle Bangla’s unique linguistic features like morphology and syntax. BanglaT5 achieved a ROUGE-1 F1 score of 0.6979, Exact Match Accuracy of 0.49, and CER of 0.4054, demonstrating its ability to generate accurate, contextual answers. In contrast, mT5’s performance was much lower, with an Exact Match Accuracy of 0.0008 and WER of 0.9996. This comparison highlights the importance of fine-tuning models for specific languages like Bangla, emphasizing the limitations of multilingual models in tasks requiring deep linguistic understanding. The system developed in this research offers a scalable solution for Bangla QA, with potential applications in education, public services, and digital literacy, contributing to the growing field of Bangla NLP. Future work will focus on deploying the model in real time, expanding the dataset, and exploring multimodal capabilities to increase its use in real-world applications.
Identifying The Authenticity of Images Using Deep Learning Techniques
(Daffodil International University, 2025-01-14) Mondal, Sourav Kumar
The rapid development of generative artificial intelligence has produced a wave of hyper-realistic deepfakes, posing existential challenges for authenticity verification to occur in digital media. This study proposes a deep learning architecture for the binary classification of AI-generated and real images as a reaction to a growing need for credible detection techniques. We comparatively evaluate four various architectures: ResNetRS50, MobileNetV2, EfficientNetB0, and a specially designed CNN with integrated Gabor filters and attention mechanisms. All models were trained and evaluated on an equalized, high-quality dataset under the same experimental conditions to provide serious benchmarking. While MobileNetV2 and EfficientNetB0 achieved higher peak validation accuracies of 99.29% and 99.81% respectively, ResNetRS50 was the most powerful and most generalized model. Its robust convergence behavior, high interpretability, and resistance to overfitting— particularly under extended training durations and high-density data—make it the top choice even at a slightly lower peak accuracy of 97.24%. Extended testing using classification reports, confusion matrices, and performance curves supports this conclusion further. A web interface was also established to demonstrate real-time deployment capability, showing that the model is usable in practical applications. The proposed method not only elevates the state of AI image forensics but also serves as a basis for large-scale and trustworthy content verification systems in the face of rising synthetic media.
Fruit Quality Classification using Deep Learning and Explainable AI
(Daffodil International University, 2025-01-14) Annafi, Abdun Nafi; Muhtasim, Tasnuva
Fruit quality classification is a critical task in agriculture and food industries to ensure standardization and market value. This study evaluates the performance of VGG19, MobileNetV2, ResNet50, custom CNN and BiLSTM for fruit classification using deep learning. A dataset of 3,758 images across seven fruit classes was used for training and evaluation. Among the tested models, MobileNetV2 achieved the highest accuracy (99.48%), making it the most suitable for real-world applications due to its efficiency. LIME (Local Interpretable Model-Agnostic Explanations) was employed to interpret model predictions, verifying that fruit characteristics like color, shape, and texture were key factors in classification decisions. The study highlights dataset imbalance and lighting variations as primary challenges. Future improvements include dataset expansion, hyperparameter optimization, and real-time deployment of the best-performing model. This research provides insights into selecting optimal deep learning models for automated fruit classification, contributing to precision agriculture and quality assurance in food industries.
Bangla News Headline Generation
(Daffodil International University, 2025-01-14) Rifat, Mahmudul Hasan
This project presents a comparative study on Bangla news headline generation using two transformer-based models: the multilingual mT5 and the monolingual BT5-base. Aimed at addressing the scarcity of effective headline generation tools for low- resource languages like Bangla, the study evaluates both models on a curated dataset using standard performance metrics. While both models demonstrated stable training behavior, BT5-base exhibited faster convergence and lower validation loss, indicating more efficient learning. Evaluation results reveal a stark contrast in output quality: BT5-base achieved a ROUGE-1 F1 score of over 56% and a ROUGE- 2 score of 45.92%, significantly outperforming mT5, whose scores remained below 3% across all ROUGE metrics. Furthermore, BT5-base attained a 21.33% exact match rate and showed markedly lower Character Error Rate (CER) and Word Error Rate (WER), highlighting its superior ability to produce semantically and lexically aligned headlines. These results affirm the effectiveness of domain-specific pretraining, as the Bangla-focused BT5-base consistently delivered more fluent, accurate, and culturally appropriate headlines than the multilingual mT5 model. The findings underscore the value of monolingual transformer models for text generation in underrepresented languages and contribute a practical foundation for future advancements in Bangla NLP applications.
Multi-Model Deep Learning Approach for Monkeypox Detection: Evaluating Base and Enhanced CNN Architectures
(Daffodil International University, 2025-01-14) Yousuf, Mahmud Kabir
Global reemergence of monkeypox has underlined the urgency of rapid, specific and scalable diagnostic solutions. PurposeThis study investigates the possible use of deep learning techniques for automated classification of monkeypox from skin lesion images in a cohort of suspected and confirmed monkeypox patients. The study evaluates five known architecture of Convolutional Neural Networks (CNN)- VGG16, InceptionV3, MobileNet, Xception, and ResNet50 on a dataset of 998 images labelled with monkeypox, chickenpox, measles and normal. Implemented each model in a base and hybrid form, hybrid version improved by attention mechanisms and noise regularization, improves focus on lesions and encourages generalization. Evaluation Metrics for Performance Accuracy, Precision, Recall and F1- score Among all the tested models, the hybrid MobileNet and InceptionV3 models had the highest accuracies of 97% and 96%, respectively, and stable performance in classification. Hybrid versions of other models like Xception and VGG16 also performed very well. On the other hand, the ResNet50 hybrid model performed poorly, suggesting the difficulties adapting that architecture for this problem. The findings from this study corroborates that hybrid deep learning models are immensely beneficial for improving the accuracy and robustness of monkeypox classification from images. Thus, these results highlight the promise of AI-mediated diagnostic tools to address early detection and outbreak management, especially in resource- scarce settings. Research coming down the pipeline will identity suitable prospects for data expansion, transformer architectures, and integrating multimodal clinical data to ultimately yield a more reliable and clinically- applicable diagnostic.
Watermelon leaf disease detection and classification using Yolo architecture
(Daffodil International University, 2025-01-14) Hasan, Mir Saem
Watermelon is a fruit that people all over the world enjoy. However, eating watermelon is mostly in the summertime because the juicy and cooling effect to the body it has is highly appreciated. In countries like Bangladesh and other countries with long, hot summers, the demand for watermelon fruit is more or less similar. The yield quality is compromised due to several leaf diseases affecting watermelon production and causing a decrease in value of financial returns. Thus, very soon and accurate diagnosis of these diseases becomes very critical in order to minimize the losses and ensure sustainable agriculture. The deep learning-based approach of the study uses the YOLOv11 model for real-time detection and classification of watermelon leaf diseases. For that, a balanced dataset of healthy and diseased watermelon leaf images was collected and further added to for improvement of model performance. The specifications are that its detection speed and accuracy is very good while being lightweight in its design; because of that, the YOLOv8 architecture was selected. The model's robust precision under challenging field conditions in identifying multiple classes of diseases was attained with optimal learning parameters. Precision was achieved to be 93.3%; Recall 87.4%, mAP50, and mAP50-95 were 96.4% and 81.5% respectively. A web- based application was developed for realtime disease detection in uploaded leaf images to ensure easy reach and use for farmers and agricultural experts. The system presented in this integrated approach contributes promise toward smart agriculture through enhanced crop monitoring in favor of food security and economic stability.
Classification of Network Traffic Anomalies Using Deep Learning Techniques
(2025-01-14) Khan, Md. Rabby Hossain
Predicting network traffic is essential to optimize network resource management, congestion avoidance, anomaly detection, and QoS in general. This paper presents and investigates a deep learning technique for the precise forecast of future network traffic through historical data. A dataset of 1,000,000 records and 11 features was used (700,000 samples for training and 300,000 for testing). The raw traffic data was divided into a normal class and an anomaly class for the classification and anomaly detection task. Data Preprocessing The steps of missing value treatment, normalization, data cleaning, and reshaping the data for meeting the minimum input requirement of the deep learning model were performed. Three deep learning algorithms (multilayer perceptron (MLP), feedforward neural network (FNN), and autoencoder (AE)) were developed and evaluated. They were chosen for their ability to model complex, non-linear relations from the network traffic and in order to obtain representations that traditional statistical models have not been able to learn. The experimental results showed that the MLP and FNN models produced high accuracy rates of 0.99, which was indicative of high predictive ability. The Autoencoder, despite its inferior performance with an accuracy of 0.94, also performed well in unsupervised learning and anomaly detection. Performance measures such as precision, recall, F1-score, confusion matrices, and ROC/Precision-Recall curves demonstrated the robustness and generalization of the models. The comparative study demonstrated that deep learning models, such as MLP and FNN, were more efficient than the conventional statistical predicting anomalies models. These results demonstrate the efficiency and scalability of deep learning for real-time network traffic prediction and anomaly detection, providing an intelligent and proactive network management technique for advanced communication systems. We deployed our top-performing model online and are currently examining the results produced visually
A Deep Neural Network Approaches for Detection of Guava Leaf Disease
(Daffodil International University, 2025-01-12) Hasan, Mahedi
Guava production is seriously threatened by guava leaf diseases as anthracnose, rust, and leaf spot, which result in severe yield and quality reductions. Effective management and the reduction of financial losses depend on the early and precise detection of these illnesses. In this work, a deep learning-based system for automatically identifying and categorizing guava leaf diseases is developed. Images of both healthy and diseased guava leaves are analyzed using convolutional neural networks (CNNs), with preprocessing methods like scaling, normalization, and augmentation improving model performance. To maximize feature extraction and computational efficiency, transfer learning techniques are used, including architectures like VGG16, ResNet50, and MobileNet. Metrics like accuracy, precision, and recall are used to assess the system's efficacy, showing that it can accurately classify the conditions of guava leaves. With the help of this computerized technology, farmers can detect illnesses early and take prompt action, lowering their reliance on chemical treatments. This study demonstrates how artificial intelligence can be used practically to advance precision farming by promoting sustainable agricultural practices.
Individual Tree Identification and Classification Using Image Data Processing Method
(Daffodil International University, 2025-01-12) Ahmed, Shafin
Identifying trees based on their type, such as fruit-bearing, medicinal, or forest trees, plays a crucial role in environmental monitoring, agriculture, and forestry management. This project aims to develop a robust deep learning-based system to individually identify these tree types from images of their leaves and bark. Three state-of-the-art convolutional neural network (CNN) models were employed for this purpose: ResNet50, ResNet101, and InceptionV3. The dataset used for training and testing consisted of diverse images collected from various sources, covering different tree species under varying conditions. The images were preprocessed using techniques like resizing, normalization, and data augmentation to ensure the models could learn effectively and generalize well to new samples. Each model was fine- tuned using transfer learning, leveraging their pre-trained weights on ImageNet. Performance metrics such as accuracy and loss were evaluated during the training and testing phases to compare the models. Among the three models, ResNet101 demonstrated superior performance, achieving a test accuracy of 85%. ResNet50 and InceptionV3, while still effective, exhibited slightly lower accuracy. This result highlights the effectiveness of deeper architectures like ResNet101 in capturing intricate features and patterns in leaf and bark images for tree classification. The findings of this study provide a foundation for deploying automated tree identification systems in real-world applications, such as forest management and ecological research. Future work can focus on expanding the dataset and incorporating additional models to further improve accuracy and scalability.
Human security with supervised learning: Automatic hate speech detection with encoding method on violence features
(Daffodil International University, 2025-01-12) Tanjin, Subrina
This investigation goals to develop an intelligent system that identifies hate speech in audio recordings and replaces offending phrases with a beep sound while maintaining the speaker's natural voice quality. Feature extraction and noise reduction, especially Mel- frequency cepstral coefficients, are done through a dataset of over 3,000 voice samples of both hate and non-hate speeches, made possible by the Librosa package for effective audio processing. Various machine learning models, such as Random Forest, XGBoost, GBoost, KNN, and Logistic Regression, classify audio samples as hate or non-hate speech. It comes up to an incredible 85% detection accuracy. Wherever hate speech is detected, the deep learning capabilities ensure the system smoothly converts the objectionable words to a beep without influencing the overall tone and rhythm of speech. In days to come, real-time speech processing will also be developed whereby this system can mark and change speech during a live conversation. For the time being, the concentration remains on processing audio files. Furthermore, the integration of robust cybersecurity measures secures users' data in processing and storage with full compliance to privacy laws. Given its novelty in voice processing, this research incorporates a powerful method for moderating bad speech, with the opportunity to make digital communication platforms more inclusive, safe, and resistant to harmful material.

Project Report

Browse

Filters

Settings

Sort By

Results per page

Search Results