基于多阶段数据生成的自循环文本智能识别

Self-circulation Intelligent Text Recognition Based on Multi-stage Data Generation

在复杂多样场景下,极少存在同时对英文和中文都具有较优识别效果的大数据标注方法.因此文中提出针对复杂多样文本识别场景的数据生成和多阶段自循环训练算法.按照定义的生成数据参数随机生成文本数据,免去数据标注过程.在卷积循环神经网络的基础上,进行多阶段自循环训练,在循环过程中通过控制数据生成策略不断提升样本的识别精度.实验表明,文中算法在多个公开英文数据集及中文特定的复杂文本场景下都具有良好的识别性能.

There are few effective big data annotation methods for both English and Chinese recognition in complex and diverse scenarios.Therefore,multi-stage data generation self-circulation training algorithm(MSDG-OCR)for complex and diverse text recognition scenarios is proposed.Text data is generated randomly according to the defined generated data parameters,and the data annotation process is omitted.Grounded on convolutional recurrent neural network(CRNN)model,multi-stage self-circulation training is carried out,and the recognition accuracy of the samples is continuously improved by controlling the data generation strategy during the loop process.Experiments show that the proposed algorithm gains good recognition performance in multiple public English datasets and Chinese-specific complex text scenarios.