Recent Trends in Data-Efficient Scene Text Recognition: A Survey on Zero-Shot and Few-Shot Learning

Scene Text Recognition (STR) is widely applied in computer vision for extracting and interpreting textual information from natural scene images and in this survey, a comprehensive analysis of Zero-Shot Learning (ZSL) and Few-Shot Learning (FSL) approaches for STR is presented. The limitations of traditional Optical Character Recognition (OCR) systems, which rely heavily on large annotated datasets, are addressed through data-efficient learning paradigms. It is demonstrated that ZSL is utilized to recognize unseen text categories without requiring prior labelled examples, while FSL is employed to enable learning from a limited number of annotated samples. Various methodologies are examined, including semantic embedding techniques, transfer learning strategies, meta-learning frameworks, and generative models such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), through which robust feature representations are learned. In ZSL, relationships between seen and unseen classes are established using semantic representations, whereas in FSL, rapid adaptation is achieved through optimization-based and metric-based approaches such as Model-Agnostic Meta-Learning (MAML), Prototypical Networks, and Siamese Networks. Transformer-based architectures and vision-language models are also incorporated to enhance performance under limited data conditions. In this paper a comparative analysis between ZSL and FSL is conducted based on training data requirements, generalization ability, feature extraction mechanisms, and application domains. Also discuss benchmark datasets such as SynthText, ICDAR, SVT, and Synth90k are utilized for evaluation, and performance is measured using metrics including Word Recognition Accuracy, Character Recognition Accuracy, Edit Distance, Unseen Class Accuracy, and Harmonic Mean. Key challenges are identified, including semantic-visual domain gaps, high intra-class variability, domain shifts, data scarcity, and overfitting, while additional issues such as noisy backgrounds, distorted text, and low-resolution images are also considered. It is concluded that ZSL and FSL are capable of improving scalability, adaptability, and efficiency in STR systems by reducing dependency on large labelled datasets, and future research directions are suggested in areas such as self-supervised learning, domain adaptation, multimodal embeddings, and the development of specialized evaluation benchmarks.