摘要
针对传统图像文字识别技术采用模板匹配法和几何特征抽取法存在识别速度慢、准确率低的缺点,提出一种基于深度学习的图像文字识别技术,使用开源、灵活的Tensor Flow框架以及LeNet-5网络训练数据模型,并将训练好的模型应用于特定场景印刷体文字识别。实验结果表明,识别模型的top 1与top 5准确率分别达到了99.8%和99.9%。该技术不仅可快速有效地处理大量图片文件,而且能综合提高图像文字识别性能,节省大量时间。
In view of the shortcomings of slow recognition speed and low accuracy in the traditional image and character recognition technology based on template matching method and geometric feature extraction method,an image character recognition technology based on deep learning is proposed.Open source,flexible TensorFlow framework and LeNet-5 network are used to train data model,and the trained model is applied to the recognition of printed characters in specific scenes.The experimental results show that the accuracy of top 1 and top 5 of the recognition model reaches 99.8% and 99.9% respectively.This technology can not only process a large number of image files quickly and effectively,but also improve the performance of optical character recognition and save a lot of time.
作者
夏昌新
莫浩泓
王成鑫
王瑶
闫仕宇
XIA Chang-xin;MO Hao-hong;WANG Cheng-xin;WANG Yao;YAN Shi-yu(School of Computer Science,University of South China,Hengyang 421001,China)
出处
《软件导刊》
2020年第2期127-131,共5页
Software Guide
基金
衡阳市科技计划项目(2017KJ273)
南华大学大学生研究型学习和创新性实验计划项目(2018XJXZ117)。
