Research Projects

"The internet's vast multimedia content necessitates automated technologies for organizing, analyzing, and recognizing it. However, the lack of labeled images in new visual domains poses a significant challenge. Domain Adaptation (DA) models have been developed to address this issue by reducing the distribution gap between the target and source domains. However, existing models struggle with preserving discriminative information, maximizing variance, minimizing distributions, aligning subspaces, maintaining sample similarity, and minimizing distances between within-class and between-class matrices. This project proposes a novel unsupervised domain adaptation framework for image classification, comprising six models: Unsupervised Domain Adaptation (UDA), Few-shot UDA (FUDA), Zero-shot UDA (ZUDA), Unsupervised Deep DA (UDDA), Few-shot UDDA (FUDDA), and Zero-shot UDDA (ZUDDA). UDA and UDDA integrate all the objectives necessary for minimizing the distribution gap into a unified framework. They are extended to FUDA and FUDDA models to address limited labeled data issues in the source domain. The proposed framework can accurately detect and classify objects or abnormalities under different lighting conditions, reducing the need for extensive labeled data and making it cost-effective. It can improve the performance and robustness of machine learning models, enabling them to generalize well to new and unseen data and improve accuracy across various applications, including medical, surveillance, and object recognition."