Hi-MetaCap: Configuring Object Relational Transformer in Meta-Learning Environment for Image Captioning in Hindi

This article proposes a meta-learning-based, few-shot image captioning framework based on an ensemble of object-relational transformer models and a self-distillation strategy. Unlike traditional approaches, the proposed framework can train models using non-paired images and captions. In each iteration, multiple base models are trained on distinct data samples, forming an ensemble that generates pseudo-captions with confidence-based weighting. An efficient pseudo-feature generation method based on gradient descent enables learning from non-paired captions. These pseudo-features and pseudo-captions are then used to train the base models in subsequent iterations. The obtained results show that using just 1% of the paired image-caption training data significantly improves performance and generates meaningful captions. The proposed framework demonstrates the potential to reduce reliance on large, paired datasets while maintaining high-quality image captioning performance. 1