Deep learning for image processing and restoration

• Main Author: Zhang, Heyi
• Other Authors: Wen Bihan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Computer and information science
• Online Access: https://hdl.handle.net/10356/177195
• Institution: Nanyang Technological University

In the realm of machine learning and data science, the issue of data imbalance significantly hampers the accuracy of attribute prediction models, leading to biased and unreliable outcomes. This project endeavors to address this pervasive challenge by implementing and evaluating various data imbalance mitigation techniques, aiming to enhance the robustness and predictive performance of attribute prediction algorithms. Through a comprehensive study involving a series of experiments on diverse datasets, we investigate the efficacy of several approaches, including oversampling the minority class, undersampling the majority class, and employing synthetic data generation techniques. Moreover, we explore the integration of advanced algorithms that inherently adjust to data imbalances, such as cost-sensitive learning models. Our findings reveal significant improvements in prediction accuracy and model fairness across various metrics, demonstrating the potential of these mitigation strategies in overcoming the adverse effects of data imbalance. This report not only highlights the critical importance of addressing data imbalance in predictive modeling but also provides a valuable reference for future research and applications in this area, suggesting a path forward for developing more equitable and effective machine learning systems.