An Exceedingly Fast Model for Low Resolution Handwritten Digit String Recognition

In this paper, we present a simplified target detector in character level for low resolution handwritten digit string recognition (HDSR). By utilizing the bounding boxes, some difficulties such as connecting, overlapping and arbitrary length can be well solved. The designed model takes advantage of residual network and anchor boxes of digital shape to recognize the images effectively and accurately. Experiment results show that the accuracy on unknown length strings can be up to 95.6%, meanwhile it has an extremely fast speed, which is up to 257 FPS on 32*160 px resolution images when using a Tesla P100 GPU.