Thesis (Bachelor of Science in Computer Science)

Browse

Subject: search.filters.subject.CNN

Search Results

Now showing 1 - 10 of 106
  • Thumbnail Image
    Item
    Image forgery detection comparison between MobileNetV2 and VGG16 convolutional neural networks
    (BRAC University, 2020-10) Nandy, Aritra; Hasan, Md. Mustakim; Sayad, Abu Bakar Md; Khan, Imteenan Akhter; Anindita, Amina Azad; Uddin, Jia
    As there are an immense scope of useful assets to alter images now, the requirement for confirming the authenticity of images is more necessary than any time in recent memory. While forgery techniques are progressively getting better that even human perception appears quite difficult to perceive these changes, regular algorithms, which attempt to identify altering patterns, frequently pre-define suppositions that restrict the extent of issue. In this manner, such strategies fail to detect forgery strategies in computer programs. Inside the following publication, we initiate structure which uses Machine Learning methods to distinguish forged photos. Consequently, the MobileNetV2 network in [40] is altered with the goal that it very well may be well equipped to the goal of image forgery identification. It is contended by the rest spatial measurements of initial layers, the system is probably going to learn prominent highlights in these layers, and afterward succeeding layers are to extract these prominent highlights and coming to a conclusion determining an image is tampered. Furthermore, by our e orts we additionally lead an extensive examination to demonstrate those contentions. Exploratory outcomes show that this architecture-modified system accomplishes an amazing accuracy of 93.15%, which outperforms VGG16 neural network on which the previously defined system depends with a margin of healthy amount up to 10.05%.
  • Thumbnail Image
    Item
    Aging face verification using deep learning
    (BRAC University, 2020-10) Bushra, Fairooz Nawar; Elma, Farhat Lamia; Khan, Ramisa Sadeque; Shahba, Shiana; Chakrabarty, Amitabha
    The era of technological security has grown more attention and interest than ever in the past few years. From wired video surveillance and passcodes, to wireless cameras and facial recognition. Over the years Deep learning has seen a high rate of improvement and its approaches for facial recognition and verification have been observed to have the most optimistic results. Our research focuses on the analysis of di erent Convolutional Neural Networks (CNNs) that have been developed in recent years. We carry out an extensive analysis of the differences in the performances of the VGG-19 architecture, the ResNet-50 architecture, the InceptionResNet v2 architecture and the Xception architecture while verifying images of the same or di erent identities with a large age gap on the two widely used datasets namely the MORPH-II dataset and the FG-NET dataset. Our results show that the VGG-19 model has an accuracy rate of 58.005%, InceptionResNet v2 has 44.26%, ResNet-50 has 35.26% and lastly, VGG-19 has 24.74%.
  • Thumbnail Image
    Item
    Detection of brain tumor using several convolutional neural network architectures
    (BRAC University, 2020-10) Salehin, Abrar; Ahmad, Md. Sizer; Islam, Moinul; Parvez, Mohammad Zavid; Reza, Md Tanzim
    The word "brain tumor" defines the unusual expansion of the cells in the brain. Among other tumors, brain tumors are possibly one of the most alarming and lifethreatening. So, detection of brain tumor in early stage is much needed because many individuals died as they were unaware of getting a tumor in the brain. For this purpose, di erent machine learning algorithms and image processing techniques are used for the early detection of brain tumor. The aim of this study is to detect brain tumor by observing di erent areas of brain and tumorous grow of brain tissues with the help of functional magnetic resonance imaging (fMRI) data. Our main goal is to determine whether the tumor is present in patient's brain or not. After data collection, we have pre-processed the data where di erent steps like image extraction, data segmentation were performed. We have used CNN architectures for the classi cation of brain tumor. For this purpose, di erent pre-trained CNN model VGG16, VGG19, Inception V3, ResNet50, DenseNet121 and Xception have implemented. Among those models we have identi ed 3 models (Inception V3, DenseNet121, VGG19) which gave higher accuracy compared to other models and selected them for further work. Rather than taking one model as most accurate we have used ensemble method in our study which produced better predictive solution in terms of brain tumor detection.
  • Thumbnail Image
    Item
    Sign language recognition using CNN and OpenPose and making 3D model of sign language
    (BRAC University, 2020-10) Hossain, MD. Jubel; Anik, MD. Asaduzzaman Sarker; Toma, Farzana Haque; Rahaman, Ananya; Alam, Golam Rabiul
    According to the World Health Organization (WHO), 466 million people over the world have impairing hearing misfortune (over 5 percent of the total populace), of whom 34 million are kids. There is an immense correspondence gap between ordinary individuals and hard of hearing and quiet people. Some guardians are unconscious of sending their hearing and talking hindered kid to class or some of them feel bashful. So for learning gesture based communication is hard for those youngsters and some incapacitated grown-up individuals are additionally don't know gesture based communication appropriately. This circumstance causes us to think of our thought. Our venture is making 3D model for all alphabets and numbers as well as some mostly used words and sentences of Bengali language and train the individuals as a virtual teacher. The model will help the individuals until he will do 100 percent correctly. Here for sign language recognition we use algorithm called Inception v3 which is an extended version of CNN (Convolutional Neural Network) and for activity recognition we use OpenPose. Tensor ow is used for coding as this is vastly used for machine learning and deep learning. We use video as our dataset and create 3d avater to teach user sign language. The 3d avatar is created using OpenPose unity plugin. Most of the researches based on sign language are hand gesture based. Unfortunately, sign language not only consist of hand gesture, It also includes face gesture, eye's gesture. And we are taking all of these things in consideration and we are trying to implement all of these. While detecting our Bangla numbers and letters in sign language we have got 89 percent accuracy.
  • Thumbnail Image
    Item
    AGHC-neurotransformer: an attention-guided hybrid CNN-transformer model for robust classification of neurodegenerative disorders from MRI scans
    (BRAC University, 2025-06) Saha, Saj; Akhter, Farzia; Tanny, Mitul Roy; Chowdhury, Nabil Hossain; Alam, Md. Ashraful; Kowsar, Sahib
    Neurodegenerative diseases like Alzheimer’s and Parkinson’s gradually damage the brain’s structure and function, leading over time to serious and lasting problems with memory, thinking, behavior, and movement. These diseases often manifest subtly in their early stages, making timely diagnosis challenging yet critical for slowing disease progression and improving patient outcomes. In our paper, we proposed an attention-guided hybrid deep learning framework for the early and accurate classification of AD, PD, and Healthy controls using axial PD-T2 weighted Magnetic Resonance Imaging (MRI) slices. Our model leverages a Convolutional Neural Network (CNN) backbone, specifically EfficientNetB0, to capture low- and mid-level spatial features with high computational efficiency.To help the model better understand the bigger picture and relationships across multiple MRI slices, we added a Transformer-based encoder after the CNN feature extractor. A Convolutional Block Attention Module (CBAM) is introduced between the CNN and Transformer components to refine feature maps by adaptively weighting spatial and channel dimensions. Furthermore, the model incorporates an attention-based slice pooling mechanism, allowing it to automatically prioritize and aggregate the most informative MRI slices from each subject without requiring manual slice annotations. The proposed hybrid architecture is trained and evaluated on a curated dataset of 2D axial brain MRI slices, which have been preprocessed and organized at the subject level. Through extensive experiments, we demonstrate that our model achieves competitive classification accuracy with a significantly lower parameter count compared to larger transfer learning baselines, such as ResNet-50 and VGG- 19. Our framework effectively combines local feature extraction, long-range dependency modeling, and adaptive attention mechanisms to address the complexities of MRI-based neurodegenerative disease diagnosis. This study states the potential of combining CNNs, Transformers, and attention modules for medical image analysis, offering a scalable, interpretable, and clinically relevant diagnostic tool for the early detection ofADand PD. Our experiments report a decent accuracy of 95% as opposed to a 93% accuracy reported by Yan’s (2025) model on our dataset, while using a significantly lower learnable parameter count. As a result, these systems could be integrated into real-life neurological screenings, thereby making it easier to identify diseases early and support early intervention and preventive strategies.
  • Thumbnail Image
    Item
    Cross-domain emotion recognition using SAM-based region extraction: a comparative study on FER and emotic datasets
    (BRAC University, 2025-06) Majumder, Niloy; Rosan, Tasnim Hasan; Musa, Shihab; Tabassum, Tasmia; Zaman, Samiha; Ahmed, Md Sabbir
    In the digital age, understanding facial expressions and recognizing emotions will be crucial for enhancing human-computer interactions. This research will help to develop a framework for better facial analysis and emotion recognition, integrating traditional computer vision with advanced deep learning models. To refine facial analysis and focus precisely on expressive regions, we will explore the application of advanced segmentation techniques, particularly the Segment Anything Model (SAM), to accurately delineate facial features before emotion classification. The detected facial regions, and potentially their segmented components, will be preprocessed with resizing, grayscale conversion, and normalization, ensuring consistency for emotion analysis. These preprocessed images will then be fed into a CNN model, which will be trained using the rich and contextually diverse EMOTIC dataset, subsequently mapped to predict one of seven fundamental emotions: Angry, Disgust, Fear, Happy, Sad, Surprise, and Neutral. Real-time emotion recognition will also be implemented using webcam video frames. Moreover, considering the significant role of facial expressions in communication, this research will contribute to mental health monitoring and improved human-computer interaction. By advancing the integration of OpenCV, advanced segmentation, and deep learning with a focus on comprehensive training data, we aim to create a robust and accurate system for facial analysis and emotion recognition. Instead, we want to create a model pipeline that is both objective and based in reality, capable of performing well in ordinary emotional settings. Rather than attempting to artificially balance the dataset, we accepted its inherent unevenness because emotions do not occur in equal amount in real life. We hoped to achieve more than just a high-scoring algorithm by allowing the model to learn from data that reflects how people actually express their feelings. We set out to create a model that understands emotions with the delicacy, depth, and diversity of human experience, detecting feelings not only in clean, controlled situations, but also in the chaotic, beautiful complexity of everyday life.
  • Thumbnail Image
    Item
    Transforming Bangladesh agriculture: AI for precision crop disease management
    (BRAC University, 2025-06) Hossain, Shahriar; Nahin, Al-Zaber; Hassan, Tasnuva; Haque, Zarif Ayman; Farin, Nusrat Jahan; Hossain, Muhammad Iqbal
    The agricultural sector encompasses a large chunk of the economy of Bangladesh as it has the necessary preconditions and factors to be suitable for agriculture. Agriculture is wholly at the whims of the environment and associated natural factors. Innovations from the time man has mastered the art of farming have allowed us to have in control some factors to ensure the desired output however there remains room for improvement and innovation especially in regards to disease detection. Currently even with a large agricultural sector, the methods for disease detection and risk management are lacking due to the inefficiencies in the system which can be very costly. To mitigate this technological innovations such as machine learning and image processing can be used to combat visible signs of disease and achieve early detection. In this paper we have explored the current options available and what can be done to make it suitable to our conditions, which ones are the best for our problem and finally we have proposed a solution we deem feasible. In our reviewed past works we have come across three models, namely Xception, VGG19 and ResNet50 which perform the best for our use cases, giving us the best results for leaf disease detection. These models have been implemented with a transfer learning approach to achieve the best results. Finally we have created a hybrid model approach combining Xception and a Vision Transformer to get the advantage of both a CNN and a Transformer to achieve the best result for our purpose.
  • Thumbnail Image
    Item
    An efficient deep learning approach to detect various diseases using chest X-ray images
    (BRAC University, 2025-02) Hassan, Sanzana Mahrukh; Khan, Md. Anik; Hossine, Md. Abid; Lamia, Mayesha Zaman; Sarkar, Pritom Kumar; Alam, Md. Ashraful
    We propose and demonstrate an efficient deep-learning approach to classify various diseases using chest x-ray images. The proposed system comprises several steps: image acquisition, preprocessing, and classification of various diseases. The datasets include X-ray images of various diseases such as pneumonia, COVID-19, lung opacity, and normal chest images. Raw X-ray images and the dataset from Kaggle is preprocessed using image resizing and augmentation. Finally, a network-based deep learning model is applied to classify the disease. Different CNN architectures: ResNet50, ResNet101, EfficientNet, DenseNet121, and AlexNet are investigated for the classification, and the best-performing architecture is used in the model, to design a custom-made model named X Net. By incorporating certain layers from both ResNet101 and DenseNet121. The ResNet101 and DenseNet121 models gave 94% and 92% accuracy, respectively, where the rest of the models gave lower accuracy than them. Our proposed model achieves a higher accuracy of upto 96.5%.
  • Thumbnail Image
    Item
    Enhancing lung diseases recognition through CNN-RNN methodologies
    (BRAC University, 2025-02) Ahsan, Md. Fardin; Orni, Ramisa Anan; Zahin, Israt Ara; Hossain, Adiba; Tabassum, Muntaha; Noor. Jannatun
    Diagnostics of respiratory disorders greatly benefit from medical imaging, especially X-ray imaging, which offers important information about the anatomical anomalies of the lungs. As we explore deeper into the field of lung illness recognition, it becomes clear that using multiscale Deep Convolutional Neural Network techniques has the potential to transform the detection of pneumonia and tuberculosis from Xray pictures. In this paper, we will classify images through a process that requires only chest-xray images. We have proposed a Deep Learning (DL) based algorithm for lung disease detection, which we termed as Convolutional Recurrent Network (CRNet). In our research, we classify chest X-ray images into four categories according to the publicly available dataset. Our proposed model can calculate the dependency and continuity properties of the intermediate layer output very precisely. At the same time, the features of these intermediate layers can be combined with the final fully-connected network for classification prediction, resulting in better classification accuracy. We have explored the potential of combining CNN and RNN with XAI to identify lung diseases from chest radiographs to improve diagnostic accuracy compared to traditional single-scale methods. Upon comparing our suggested model with the current models, we discovered that, with an accuracy of 93.51% on the full dataset, our suggested model achieved the best accuracy of all the architectures we compared. Moreover, our suggested model C-RNet was observed to accurately categorize and detect the regions of disease through approaches such as Grad-CAM.
  • Thumbnail Image
    Item
    Segmentation and classification of fish species using deep learning
    (BRAC University, 2025-02) Rahman, Faiaz Ibnee; Talukdar, Nahiyan Rahman; Akhand, Zaion Abrar; Reza, Md. Tanzim
    Fish Species Classification using deep learning surfaced as a formidable tool for automating and enhancing species identification in aquatic ecosystems. Moreover, leveraging Convolutional Neural Networks (CNNs), this approach presents us with an efficient and accurate way of identifying and categorizing fish under different species based on image data, providing us with a notable advancement for aquatic biodiversity surveillance and fisheries management. The switch towards deep learning addresses the constraints of traditional techniques such as manual labor and morphological analysis, which can be time consuming, require specialist knowledge and are inclined to human error. The primary objective of our study reviews recent advances in the field of deep learning, which focus on Convolutional Neural Networks and their application in classifying fish species as well as segmenting them through images of fishes overwater and maintaining a white background. We analyzed various approaches adopted in recent research, from CNN-based models like ResNet-50, Inception-v3, YOLOv11 etc. to innovative image preprocessing techniques, highlighting the evolution of methodologies from rudimentary to more sophisticated automated systems. We introduced a custom dataset of 7100 images which were captured during daylight as well as night time for better picture quality keeping in mind the reduction of computational intensity for real-time applications. In addition, through this research our aim is not only to refine the accuracy of fish species classification and segmentation but also to contribute significantly to the protection of marine ecosystems, aiding in the detection of rare species and assisting in sustainable fishing practices. To conclude, this study stands at the forefront of technological advancements in ecological conservation and offers in industrial use.