Date of Award
2025
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Science
Committee Chair
Tathagata Mukherjee
Committee Member
Chaity Banerjee
Committee Member
Sudhir Aggarwal
Research Advisor
Tathagata Mukherjee
Subject(s)
Image processing--Digital techniques, Pattern perception, Deep learning (Machine learning)
Abstract
Processing and analyzing multi-source image data is vital across a wide range of fields, including artificial intelligence (AI), deep learning, computer vision, geolocation, autonomous navigation, remote sensing, and forensic investigations. Although conventional image matching methods such as Scale-Invariant Feature Transform (SIFT) and Speeded-Up Robust Features (SURF) have proven effective in controlled setting, they often face challenges when applied to diverse datasets that exhibit substantial variations in viewpoint, scale, lighting conditions, and modality. Recent developments in deep learning, including convolutional neural networks (CNNs), transformers, generative adversarial networks (GANs), and self-supervised learning, have led to the creation of more efficient and precise methods for image matching and retrieval. This thesis proposes a novel AI-driven deep-learning-based framework designed to enhance the robustness and scalability of image matching. The framework is tailored for indoor scene-based image matching to support image localization, addressing challenges such as lighting variations, occlusions, lack of sensor calibration and sensor discrepancies. It employs a channel-aware autoencoder that captures meaningful features by leveraging spectral differences in RGB channels, thereby improving the model’s ability to distinguish between visually similar indoor environments. By incorporating latent feature representations, the framework enhances scene classification and retrieval accuracy, potentially contributing to advancements in deep learning-based indoor positioning systems, robotics, and forensic analysis. The experimental results highlight that the proposed AI-driven frameworks outperform traditional image matching methods by a significant margin. Together, these contributions advance the field of deep learning for image-based search and localization, providing scalable and intelligent solutions that can be applied across various real-world contexts.
Recommended Citation
Bhatia, Vidhi M., "A deep learning multi-channel framework for indoor image retrieval" (2025). Theses. 755.
https://louis.uah.edu/uah-theses/755
