Low Resolution Handwritten Digit String Recognition based on Object Detection Network

摘要

A novel object detection network is proposed in this paper for low resolution handwritten digit string recognition. It is composed of a convolutional neural network (CNN) and two independent output branches for classification and bounding box regression. The network is designed to effectively extract the features from low resolution images. Non-categorized non-maximum suppression (NMS) and mini-batch fine-tuning (MB-FT) are used to improve accuracy further. The experiments are conducted on a new dataset collected by a tablet and HDSRC 2014 benchmark datasets, and the high metrics are obtained. Furthermore, its prediction speed reaches 65 FPS achieving real-time recognition.