High-Speed One-Shot Detection and Recognition of Low-Resolution Text Trained on Synthetic Data

In this study, we address the challenge of text detection and recognition in low-resolution images. Addressing the dual constraints of limited training data and the necessity for real-time processing, we adopt a twofold strategy. Firstly, we introduce a pipeline for the generation of synthetic datasets that requires minimal manual annotation and is specifically designed for the context of low-resolution real-life text images. Secondly, we employ a streamlined neural network architecture based on the U-Net model, which concurrently executes text detection and recognition across multiple contextual layers. Our method demonstrates superior performance, achieving real-time processing speeds exceeding 120 fps, and is accurate even in challenging conditions where text characters are as small as five pixels in width. Our findings suggest that both the synthetic dataset generation pipeline and the neural network model are highly adaptable and can be easily modified for a broad range of applications.