Vision Transformer for Fast and Efficient Scene Text Recognition

Scene text recognition (STR) enables computers to read in natural scenes such as object labels, road signs and instructions. STR helps machines perform informed decisions what pick, which direction go, is the next step of action. In body work on STR, focus has always been accuracy. There little emphasis placed speed computational efficiency are equally important especially for energy-constrained mobile machines. this paper we propose ViTSTR, an with a simple single stage model architecture built compute parameter efficient vision transformer (ViT). On comparable strong baseline method TRBA accuracy 84.3%, our small ViTSTR achieves competitive 82.6% (84.2% data augmentation) at \(2.4\times \) up, using only 43.4% number parameters 42.2% FLOPS. The tiny version 80.3% (82.1% augmentation), \(2.5\times speed, requiring 10.9% 11.9% With augmentation, base outperforms 85.2% (83.7% without \(2.3\times but requires 73.2% more 61.5% terms trade-offs, nearly all configurations or near frontiers maximize accuracy, same time.