文字识别技术在电力系统、车辆驾驶等领域应用十分广泛.随着人工智能技术的兴起和万物互联(Internet of Everything,IoE)的发展,厂商对随时随地获取复杂场景文字的需求也越来越迫切.针对文字识别环境背景复杂、视角畸变、字迹浅显和中英文字符混杂形似等诸多问题,设计出具有文字区域提取与校正、图像增强、文本检测和文本识别的光学字符识别(Optical Character Recognition,OCR)算法框架.设计了基于双注意力机制和内容感知上采样的DBNet文本检测模块增强网络的特征提取选择能力,提高内容感知能力,设计了融入中心损失CRNN+CTC的文本识别模块增大字符之间的特征间距.实验结果表明,改进的文本检测网络在ICDAR2015数据集上准确率提升了 5.09%,召回率提高2.12%,F评分提高了 3.46%.在中英文文本识别数据集中,改进的文本识别网络对中英文字符识别准确率提高了 1.2%.
Optical character recognition technology is widely used in power system,vehicle driving and other fields.With the development of artificial intelligence and Internet of Everything(IoE)technology,the enterprises have an increasingly urgent need to obtain text for complex scenes anytime and anywhere.For many problems such as complex background,distorted visual angle,plain handwriting and mixed Chinese and English characters,an Optical Character Recognition(OCR)algorithm framework is proposed with text region extraction and correction,image enhancement,text detection and text recognition.A DBNet text detection module based on a dual-attention mechanism and content-aware upsampling is designed to enhance the feature extraction selection capability of the network and improve content-awareness capability,and a text recognition module incorporating center-loss CRNN+CTC is designed to increase the feature spacing between characters.The experimental results show that the improved text detection network has improved accuracy by 5.09%,recall by 2.12% and F-score by 3.46% on the ICDAR2015 dataset.The improved text recognition network improved the accuracy of Chinese and English characters recognition by 1.2%on the Chinese and English text recognition dataset.