Handwritten String Recognition Based on YOLOv4 and CRNN on Different Deep Learning Frameworks

Abstract

Character recognition is one of the most important contents of pattern recognition research. Among them, due to the instability and continuity of individual differences in handwritten strings, it is difficult to extract features in recognition. At present, the recognition of handwritten strings is based on single character segmentation. However, due to the difficulty of character segmentation of handwritten string, the accuracy of recognition based on segmentation method is relatively low. In this paper, we use the non segmentation method based on RNN for character recognition, combined with the time sequence characteristics to improve the recognition accuracy of continuous strings. Firstly, handwritten string detection methods are studied respectively, based on YOLOv4, YOLOv4-Tiny model and OpenCV, the detection methods are tested; then the handwritten string recognition methods are studied, the ‘CNN+RNN+CTC’ pattern is used to calculate loss function model, thus can identify undivided strings, and then through CTC decoding, string results with higher accuracy can be obtained; finally, after testing was carried out, the experimental result shows that the method proposed in this paper can solve the problem of stylus and variable length of handwritten strings, therefore improve the efficiency of converting handwritten manuscripts into electronic manuscripts.