SoC Theses
Permanent URI for this collection
Browse
Recent Submissions
Item Empowering spatial knowledge acquisition: Navigation and mapping solutions for people with visual impairments(Wichita State University, 2024-12) Abraham, Ajay; Namboodiri, VinodBlind and visually impaired (BVI) individuals face significant challenges when navigating unfamiliar environments, including difficulties with spatial orientation, and accessing and interpreting wayfinding cues like signage or information displays, often requiring them to rely on sighted guides for navigation. Sighted users, too, can struggle to navigate complex indoor environments, particularly when faced with poorly marked signs or maps, unfamiliar signage conventions, or disorienting spatial layouts. There is also a notable lack of wayfinding systems that allow users to navigate seamlessly between indoor and outdoor spaces. This often results in users—both BVI and sighted—taking unnecessarily complex routes instead of the most efficient paths to their destinations. To address these challenges, this work introduces an inclusive wayfinding system, accessible via a smartphone, designed to assist both BVI users and individuals without visual impairments in navigating the shortest route to their destinations. The system uses an indoor module that leverages Bluetooth Low Energy (BLE) beacons and an outdoor module powered by GPS, with transition beacons ensuring seamless navigation between indoor and outdoor spaces. Many existing indoor navigation systems focused on turn-by-turn guidance fail to deliver meaningful spatial information that enables BVI users to develop a mental map of the space. Indoor maps designed for BVI users often fail to provide a complete understanding of the layout, and users may not have access to these maps before visiting the space. To empower BVI users in independently exploring and cognitively mapping unfamiliar indoor spaces, this work introduces two strategies for spatial knowledge acquisition: pre-visit exploration, where users familiarize themselves with a space using a mobile application featuring grid-based route and layout maps, and in-situ exploration, which uses the beacon-based indoor navigation system to help users understand the space during navigation. By integrating these exploration techniques into wayfinding systems and map applications, smartphones are transformed into powerful cognitive mapping tools that enhance spatial awareness, mobility, autonomy, and confidence for BVI users in navigating complex indoor environments.Item Enhancing generalizability in building damage detection: Domain adaptation and augmentation approaches for post-disaster assessment(Wichita State University, 2024-12) Parupati, Bharath Chandra Reddy; Kshirsagar, ShrutiThe increasing frequency of natural disasters necessitates rapid and reliable methods for assessing building damage to aid timely disaster response. This thesis investigates the potential for deep learning models to generalize across diverse geographical and environmental contexts for building damage detection, a critical requirement for models deployed in real-world scenarios. The primary objective is to evaluate biases within these models, specifically those that may limit performance in out-of-domain datasets, and to propose methodologies to enhance model robustness. This study leverages two datasets, xBD and Ida-BD, utilizing both in-domain and out-of-domain data to analyze model biases. We introduce a novel Fusion Augmentation technique designed to enhance the model’s ability to capture building edges, thereby improving classification of damage levels, especially in regions with dense vegetation. A series of supervised and unsupervised domain adaptation techniques, including CORAL, were applied to improve model generalizability across varied disaster scenarios without requiring target labels. Grad-CAM visualization techniques further support explainability by offering insights into the areas of focus in model predictions. Results demonstrate that combining Fusion Augmentation with domain adaptation significantly improves model accuracy, especially in damage classes, and reduces location specific biases. This research contributes a practical framework for developing reliable and generalizable building damage detection models, which can serve as valuable tools in post-disaster assessment, ultimately supporting faster and more effective humanitarian responses.Item A machine learning approach for integrating phonocardiogram and electrocardiogram data for heart sound detection(Wichita State University, 2024-12) Mains, Thu; Kshirsagar, Shruti; Sawan, EdwinHeart sound detection (HSD) is crucial for diagnosing cardiovascular diseases and monitoring cardiac health. While the traditional diagnostic methods often rely on either phonocardiogram (PCG) or electrocardiogram (ECG) data and often cause several performance degradations. In this work, we propose to combine the augmentation methodology with ECG and PCG fusion. Experiments are conducted with physioNet dataset used in CINC2016 Challenges. Experimental results show the proposed method outperforming benchmark systems by providing complementary information, hence improving performance with modality fusion.Item Brain tumor segmentation using deep learning techniques on multi-institution al MRI datasets(Wichita State University, 2024-12) Karji, Fatemeh; Kshirsagar, ShrutiBrain tumor segmentation uses medical pictures, usually MRI scans, to locate and characterize tumor areas. Separating tumor tissue from healthy brain structures allows for a more detail analysis of tumor size, shape, and location. Accurate segmentation is vital for diagnosing brain tumors, as it helps clinicians differentiate between benign and malignant tumors, understand their growth patterns, and assess their impact on surrounding brain regions. It also plays an essential role in treatment planning, allowing precise targeting of the tumor during surgery, radiation, or other therapies while minimizing damage to healthy brain tissue. This thesis tackles the crucial topic of accurate and effective brain tumor segmentation in MRI data for clinical diagnosis and therapy planning. This study offers a deep learning-based brain tumor segmentation method to overcome the drawbacks of manual segmentation, which is time-consuming, error-prone, and requires specialized expertise. The experimental pipeline includes multi-modal MRI scan pre-processing, data augmentation to overcome data scarcity, and rigorous training using the BraTS2020 and BraTS2024 datasets. In this thesis, we aim to improve the generalizability of deep learning model for accurate detection of brain tumour segmentation. This cross-data validation technique ensures model robustness and generalizability. The study also compares 2D and 3D U-Net segmentation accuracy using Dice Similarity Coefficient measures. This comparison highlights the pros and cons of each model type in representing brain tumor morphology's complexity. Additionally, this study addresses complex tumor forms, MRI modalities, post-treatment effects, and data imbalance to obtain high segmentation accuracy, showing that the proposed models could considerably impact clinical practice.Item Deep learning approaches for speech emotion recognition(Wichita State University, 2024-05) Srinivasan, Sriram; Kshirsagar, ShrutiThis thesis addresses the challenge of speech emotion recognition, focusing on contin- uous emotion estimation using deep learning techniques. Emotion detection plays a vital role in various domains, including healthcare, human-computer interaction, and affective com- puting. However, traditional approaches often struggle with accurately recognizing emotions across noise and reverberation, leading to limited diagnostic accuracy and applicability. To overcome these limitations, our study proposes a novel approach that integrates speech enhancement as a preprocessing step using advanced deep learning techniques. Our exper- imentation utilizes the AVEC 2018 challenge datasets, comprising audio/video recordings from diverse cultural backgrounds. The experimental pipeline involves several key components, including feature extrac- tion, model training, and data/speech enhancement techniques. We employ LSTM (Long Short-Term Memory) models for temporal dependency modeling and investigate the effec- tiveness of different hyperparameters, such as batch size, learning rate, and optimizer choice. We aim to evaluate the effectiveness of speech enhancement methods and explore the impact of various hyperparameters on emotion recognition performance. The results of our experi- ments demonstrate promising performance improvements when leveraging data/speech en- hancement techniques, such as single Spectral Enhancement (SSE) and Speech enhancement Generative adversarial network (SEGAN) show potential for capturing complex temporal relationships and contextual information, leading to enhanced emotion recognition capabilities. Overall, this research contributes to advancing the field of speech emotion recognition by providing insights into the effectiveness of different deep learning techniques and hyper- parameters. By improving emotion detection accuracy, our work lays the groundwork for future developments in healthcare monitoring technologies and human-computer interaction systems, ultimately enhancing patient outcomes and user experiences.