Case Study of Few-Shot Learning in Text Recognition Models

Abstract

Optical text recognition models are widely applied in document processing systems. However, a high-quality text recognition model usually requires large number of samples, extensive amount of time and computation resources. In this paper, we propose a few-shot learning framework for unsegmented text recognition, which comprises of a conventional encoder-decoder recognition module, as well as a generative module for convolutional feature generation. In the meta-training stage, a base model for general text recognition and feature vector generation is trained with large synthesized text image dataset. In the meta-testing stage, the base model is adjusted with a small number of authentic samples. With the complementation of synthesized feature vectors, the base model is adapted to the target dataset distribution. The proposed framework only requires a few authentic samples. It is both data- and time- efficient in adapting existing models to new target datasets. Experimental results on authentic datasets used in industrial applications show that the proposed meta-testing approach outperforms conventional transfer learning by up to 5.84%.