Few-shot vision recognition and generation for the open-world

• Main Author: Song, Nan
• Other Authors: Lin Guosheng
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Few-shot learning
• Online Access: https://hdl.handle.net/10356/181293
• Institution: Nanyang Technological University

Deep Neural Networks (DNNs) have achieved remarkable success across various computer vision tasks, but their reliance on extensive labeled datasets limits their applicability in data-scarce scenarios. Few-shot learning offers a promising solution by enabling models to learn from minimal data, yet traditional approaches assume a closed set of classes, which is impractical in open-world settings. This thesis addresses the challenges of few-shot learning in an open-world context by introducing three novel frameworks: Few-shot Open-set Recognition (FSOSR), Few-shot Class Incremental Learning (FSCIL), and Lifelong Few-shot Text-to-Image Diffusion. For FSOSR, we reserve space for unseen classes and leverage background features from seen classes as pseudo unseen classes to effectively learn decision boundaries. For FSCIL, we adopt a decoupled learning strategy that prevents knowledge forgetting by updating only classifiers during incremental sessions and introduce a Continually Evolved Classifier (CEC) using graph-based context propagation. In Lifelong Few-shot Text-to-Image Diffusion, we integrate data-free knowledge distillation and In-Context Generation (ICGen) to continuously generate high-quality images from limited examples while retaining prior knowledge. Extensive experiments on benchmark datasets demonstrate that these frameworks significantly improve adaptability and efficiency in dynamic environments, setting new state-of-the-art results. This thesis advances both theoretical and practical aspects of few-shot learning, enabling robust and scalable AI systems for real-world applications.