摘要
本文构建了一种基于编码器-解码器结构的网络结构,用于处理手写数学表达式识别的问题。经过对不同卷积神经网络的测试,整个网络先用密集连接卷积网络提取图片特征,再引入一个门限循环单元来形成编码器-解码器结构来处理图片特征,并在其中使用带有Coverage的2D注意力机制。基于CROHME竞赛提供的手写数学表达式数据集,结果达到了17.1%的字错误率和36.5%的识别率,验证了该模型的有效性。
A model based on Encoder-Decoder structure is constructed to deal with handwritten mathematical expression recognition.After testing different convolution neural networks,the entire network first uses densely connected convolutional networks to extract image features,and then introduces two gated recurrent unit to form an encoder-decoder structure to process image features,and uses a 2D attenion with coverage mechanism in it.Based on handwritten mathematical expressions dataset provided by the CROHME competition.The result achieves 17.1%WER and 36.5%ExpRate,which verifies the validity of the proposed model.
作者
沈佳伟
周宇昂
赵天宇
周渊
周志豪
张娟
SHEN Jiawei;ZHOU Yuang;ZHAO Tianyu;ZHOU Yuan;ZHOU Zhihao;ZHANG Juan(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai,China,201620)
出处
《福建电脑》
2021年第7期59-61,共3页
Journal of Fujian Computer
