Thesis (Master of Science/Engineering in Computer Science and Engineering)

Browse

Start date: 2020Subject: search.filters.subject.CNN

Search Results

Now showing 1 - 8 of 8
  • Thumbnail Image
    Item
    A deep learning framework for arsenic skin disease detection leveraging dual-teacher knowledge distillation with a depthwise-separable convolution and KAN-based lightweight student model
    (BRAC University, 2025-12) Mehedi, Md Humaion Kabir; Mridha, Muhammad Firoz
    Arsenic poisoning in groundwater poses a major public health issue in Bangladesh. Chronic arsenic poisoning often leads to arsenicosis, a chronic disease characterized by cutaneous effects, including melanosis, leukocomelanosis and keratosis. Timely treatment of these lesions on the skin is important since prompt diagnosis and timely action are taken within the medical profession. However, this is challenging in rural areas. In most areas, there are no trained dermatologists, diagnoses tend to be subjective and there may be inadequate healthcare facilities. The proposed study involves the development of an explainable and lightweight deep learning framework for the automatic identification of skin diseases caused by arsenic, which will be computationally efficient and interpretable on a clinical scale. This system uses the dual-teacher knowledge distillation (KD) approach, in which two high-capacity models, InceptionV3 and Xception+InceptionM, are used to distill discriminative and contextual information to a small student model called Inception-Residual- KANNet (IR-KANNet). ArsenicSkinImageBD dataset was trained and validated using a stratified data split and 5-fold cross-validation to ensure class balance and model stability. The final evaluation found accuracy 0.9692, precision 0.9610, recall 0.9867 and F1-score 0.9737 on both infected and not-infected cases using stratified data split and a factor of 78.39% reduction in parameters over the teacher 1 and 35.29% in teacher 2 networks. These results indicate the appropriateness of the model for use in low-resource healthcare settings by providing convenient and reliable AI-based screening for arsenicosis detection. Grad-CAM++ and LIME make the model explainable and the resulting transparent heatmaps are consistent with the clinical regions in terms of lesions. This study is relevant for developing interpretable, efficient and domain-flexible medical AI models. It also provides a basis for further study of explainable knowledge distillation and edge-deployable solutions for the diagnosis of other dermatological disease
  • Thumbnail Image
    Item
    Advanced video analytic system for posture and activity recognition: leveraging MediaPipe, CNN-LSTM, and ensemble learning for fall and unstable motion detection
    (BRAC University, 2024-10) Siraj, Farhan Md.; Zereen, Aniqua Nusrat
    Human posture detection and classification are vital in monitoring activities, especially in health and safety contexts, such as fall detection in elderly care. This thesis presents a comparative study of two machine learning approaches for real-time human posture calssification using real time video data, a traditional feature-based approach using a Voting Classifier, and a deep learning appraoch utilizing a Convolutional Neural Network-Long_Short-Term memory (CNN-LSTM) model. The feature-based method incorporates pose estimation using MediaPipe to extract human body landmarks, followed by classification using an ensemble of Rnadom Forest, Support Vector Machine (SVM), and K-Nearest Neighbors (KNN). On the other hand, the CNN-LSTM model captures both spatial and temporal dynamics of video sequences by extracting visual features through Convolutional Neural Network (CNN) and modeling temporal dependencies via LSTM. The models are evaluated on a dataset for four postures-Fall, SIt, Stand and Unstable-with promising result. This work demonstrate the effectiveness of combining pose-based features with voting classifiers and the power of deep learning in sequential data, offering a robust solution for real-time posture classification systems.
  • Thumbnail Image
    Item
    Feature-driven deep learning for Osteoporosis detection: leveraging explainable AI
    (BRAC University, 2024-10) Hosain, A.K.M. Salman; Zereen, Aniqua Nusrat
    Osteoporosis, a widespread bone disorder affecting over 200 million people globally, is associated with significant morbidity, particularly due to increased fracture risk. While traditionally considered a condition afflicting the elderly, younger individuals are also susceptible. In this study, we present a deep learning framework utilizing Convolutional Neural Networks (CNNs) to detect osteoporosis from knee X-ray images. Two models, MobileNet V3 and EfficientNet B0, were employed and fine-tuned on a dataset of 774 knee X-ray images. We further improved model performance by curating a new dataset, emphasizing critical features using explainable AI (XAI). Our results show that while both models achieved an accuracy of 0.77 on the original dataset, EfficientNet B0 consistently outperformed MobileNet V3 on the new dataset with an accuracy of 0.9189 and an F1 score of 0.9315, compared to MobileNet V3’s accuracy of 0.8243 and F1 score of 0.8169. These findings demonstrate the effectiveness of CNNs, particularly EfficientNet B0, in accurately diagnosing osteoporosis from medical images, and underscore the importance of both model selection and feature-focused data preprocessing in improving diagnostic performance.
  • Thumbnail Image
    Item
    ResInvolution: an involution-ResNet fused global spatial relation leveraging model for histopathological image analysis under federated learning environment
    (BRAC University, 2024-05) Dipto, Shakib Mahmud; Alam, Md. Ashraful
    Accessing image data in the domain of medical image analysis is challenging owing to concerns regarding privacy. Federated Learning is the approach used to get rid of this challenge. With millions of learning parameters, Residual Network (ResNet) is one of the most advanced architectures for classifying medical images. Because of its resource-hungry nature, using this ResNet architecture in the Federated learning framework has an impact on the entire system. This research introduces a novel architecture called Residual Involution (ResInvolution), specifically developed for analyzing histopathological images within a federated learning environment. The architecture utilizes a cutting-edge model, the Involution-ResNet Fused Global Spatial Relation Leveraging model, to enhance the analysis process. This model is impressively lightweight, boasting less than 190,000 parameters. Its efficiency and ease of deployment make it ideal for medical image analysis tasks. By incorporating involution operations into the ResNet framework, it becomes possible to adjust the spatial weighting of features dynamically. The proposed model enables a comprehensive analysis of intricate structures that exceed the capabilities of traditional convolutional networks. This model has been deployed within a federated learning environment, where privacy is prioritized. Also utilize decentralized data sources, thereby eliminating the necessity of centralizing sensitive medical images. This approach ensures strict adherence to medical data privacy regulations while simultaneously leveraging collective insights from multiple institutions. The model has undergone rigorous testing on three distinct datasets: GasHisSDB, NTC-CRC-HE- 100K, AND LC25000. In Federated Learning scenarios, the model achieves accuracies of 91%, 95%, and 99% on these datasets, respectively. However, in the context of federated learning, the accuracies exhibited are 91%, 93%, and 97%, respectively. The model’s effectiveness is evaluated through various performance metrics, including the confusion Matrix, Accuracy, Precision, Recall, F1-Score, Receiver operating Characteristic (ROC) curve, and Area under the ROC Curve (AUC) Score. The results highlight the model’s ability to adapt to various challenges, such as limited data and irregular data distribution, commonly encountered in federated learning environments. ResInvolution sets a revolutionary benchmark in medical image analysis, enhancing the ability to interpret intricate medical images and paving the way for future advancements in scalable, privacy-preserving deep learning technologies.
  • Thumbnail Image
    Item
    Into the heart of Bangla speech: advancing speech sentiment recognition with semi-supervised multimodal machine learning model leveraging an iterative SHAP-based feature selection
    (BRAC University, 2024-06) Shruti, Abanti Chakraborty; Alam, Md. Golam Robiul
    Automatic sentiment recognition from speech data is crucial for various applications. As AI has grown in popularity, the application of the importance of speech sentiment analysis is increasing along with the amount of speech in every industry. Bengali is the seventh most spoken language in the world, yet research on voice sentiment analysis in this language is lacking. This thesis investigates novel techniques to enhance speech sentiment recognition in underresourced languages like Bengali. We explore the efficacy of both unimodal (speech only) and multimodal (speech, Image, and text) approaches for different fusion techniques. This research proposed a semi-supervised Random Forest model, which achieved consistent and robust performance across different modality combinations. This model demonstrated high accuracy with fewer features, showcasing the efficiency and effectiveness of SHAP-based semi-supervised learning in handling unlabeled data. Additionally, eight different feature extraction techniques have been employed to extract acoustic features and VGG19 and Bangla Word2Vec are used to extract image and text features. Moreover, this study has experimented with different modality-based methods such as LSTM, CNN, and BanglaBERT. We have used BanglaSER, SUBESCO, and KBES datasets for our experiments. Among the various models tested, early fusion techniques proved the most effective, achieving an accuracy of up to 83% when combining speech and text modalities with LSTM classifiers and the proposed semi-supervised model acquired the highest 77% accuracy for audio, text, and image modals. In contrast, late fusion techniques showed reduced performance, though including speech and image modalities improved accuracy to 62%. Detailed performance comparisons for unimodal systems indicate that traditional Random Forest models perform well with fully labeled datasets, but our semi-supervised model works comparatively well with only 20% labeled data. Moreover, our proposed semi-supervised AdaBoost model, using only 20 features and SHAP-based feature importance, outperformed the traditional model trained with 50 features. Remarkably, the proposed Random Forest model trained with 20% labeled and 80% unlabeled data achieved over 70% accuracy across different feature selection methods, with the weighted feature selection technique achieving the highest accuracy of 72%. We believe this thesis will contribute significantly to Bangla speech sentiment recognition by providing a robust, efficient, and interpretable framework.
  • Thumbnail Image
    Item
    Machine fault diagnosis using a modified transferable CNN
    (BRAC University, 2023-07) Shammi, Sanjana Khan; Chakraborty, Amitabha
    Detecting prior bearing faults is an essential task of machine health monitoring because bearings are the crucial parts of rotating machines. The performance of traditional intelligent fault diagnosis methods depends on feature extraction of fault signals, which requires signal processing techniques, expert knowledge, and human labor. Deep learning algorithms have recently been applied for industrial machine health monitoring with their advanced features. With the capacity to automatically learn complex features of input data, deep learning architectures have great potential to overcome the drawbacks of traditional intelligent fault diagnosis. This paper proposes a rolling bearing fault diagnosis method based on Convolutional Neural Network and Leaky ReLU to solve the above problems. Firstly, the Continuous Wavelet Transform converts one-dimensional original vibration signals into two-dimensional time-frequency images. Secondly, the obtained time-frequency images are used to train the constructed model. Finally, the diagnosis of the fault location and severity is completed. The method is verified on the MFPT, MIMII data set, and vehicle engine. The results demonstrate that the suggested approach achieves higher diagnostic accuracy which is 95.49% on average and 2% greater than other advanced techniques. We have also incorporated XAI in the input images to make the network more transparent.
  • Thumbnail Image
    Item
    Deep Learning based Medicinal Plants Leaf Recognition
    (BRAC University, 2022-09) Mahalanabish, Tonusri; Rabiul Alam, Md. Golam
    Plants assume a significant part in Earth’s nature by giving food, cover and keeping a solid environment.These plants contain some significant therapeutic qualities. Due to having fewer negative side effects and being more affordable than contemporary medicine, medicinal plants are receiving interest in the pharmaceutical business.In this work, I tried to classify the plant’s images through classical methods and Deep neural network.30 medicinal plants leaves are represented by 1835 images in the proposed dataset.First, I applied CNN to classify the images and got 65.66% ac curacy.Then I applied SVM with Normal features, GrayScale features, HOG fea tures and combined features extraction and got 72.28% accuracy for Normal fea tures,73.91% accuracy for GrayScale features, 79.34% accuracy for HOG features and 80.0% for Combined feature extraction.Next I applied the VGG-19 pre-trained model and got 96.74% accuracy.At last, I applied a GradCam explainable AI method to interpret the results generated from VGG19.From all these experiments, I got the best accuracy for the VGG19 pretrained model.That’s why I used Grad Cam on the VGG19 results for getting the explanation for the predictions.
  • Thumbnail Image
    Item
    A hybrid deep learning model and explainable AI-based Bengali hate speech multi-label classification and interpretation
    (BRAC University, 2022-09) Shakil, Mahmudul Hasan; Alam, Md.Golam Robiul
    Data innovation has moved quickly in recent years, and various unfavorable alter ations have been made to the network medium. Social media platforms like Face book, Twitter, and Instagram are becoming more and more popular because they allow users to express their opinions through messages, photographs, and notes. In particular, in Bangladesh and other locations where the Bengali language is spoken. In any case, it has regrettably turned into a space with toxic remarks, cyberbully ing, and unidentified hazards. Numerous studies have been conducted in this area, but none have produced accurate results. Some effective pre-trained transformer models have been introduced. To identify Bengali malicious and non-malicious text at an early stage using simple Natural Language Processing (NLP). This study sug gests a Convolutional Neural Network with Bi-Directional Long Short-Term Memory (CNN-BiLSTM) hybrid strategy. This model can also classify any Bengali text data into six levels. Additionally, the transformed dataset is subjected to several conven tional Machine Learning methods using an estimator, and Explainable AI interprets these techniques (XAI). In the last stage, Stacking Classifier which is superior to any prior activity is used to ensemble all classifiers and the estimator.