Few-shot symbol classification via self-supervised learning and nearest neighbor

The recognition of symbols within document images is one the most relevant steps involved in Document Analysis field. While current state-of-the-art methods based on Deep Learning are capable adequately performing this task, they generally require a vast amount data that has to be manually labeled. In paper, we propose self-supervised learning-based method addresses task by training neural-based feature extractor with set unlabeled documents and performs considering just few reference samples. Experiments different corpora comprising music, text, symbol report proposal tackling high accuracy rates up 95% few-shot settings. Moreover, results show presented strategy outperforms base supervised learning approaches trained same that, some cases, even fail converge. This approach, hence, stands as lightweight alternative deal classification annotated data.