文献类型：期刊文献

中文题名：基于双路注意力机制的化学结构图像识别

英文题名：Chemical Structure Image Recognition Based on Dual Attention Mechanism

作者：季秀怡[1];李建华[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2020

卷号：46

期号：9

起止页码：213

中文期刊名：计算机工程

外文期刊名：Computer Engineering

收录：CSTPCD;;Scopus;北大核心:【北大核心2017】；CSCD:【CSCD_E2019_2020】；

基金：国家科技重大专项“现代中药组分资源库公共平台建设”(2018ZX09735002)。

语种：中文

中文关键词：化学结构图像识别;卷积神经网络;长短期记忆网络;双路注意力机制;序列生成

外文关键词：chemical structure image recognition;Convolutional Neural Network(CNN);Long Short-Term Memory(LSTM)network;dual attention mechanism;sequence generation

摘要：基于传统图像处理技术与流水线方式的化学结构图像识别方法通常依赖于人工设计的特征,导致识别准确率较低。针对该问题,提出一种基于空间注意力机制与通道注意力机制的化学结构图像识别方法。将化学结构识别视为序列生成任务,采用卷积神经网络(CNN)与长短期记忆(LSTM)网络相结合的深度神经网络模型实现化学结构图像到SMILES序列的转换。该深度神经网络模型由编码和解码两部分组成,编码部分使用CNN提取化学结构图像特征,解码部分融合双路注意力机制与LSTM网络生成SMILES序列。实验结果表明,该方法在Beam Size为3的情况下,识别准确率和BLEU-4值分别为81.63%和0.937,明显优于无注意力机制和单注意力机制的化学结构图像识别方法。
Most of the existing chemical structure image recognition methods based on traditional image processing techniques and pipeline methods usually rely on artificially designed features,resulting in a low recognition accuracy.To solve the problem,this paper proposes a chemical structure image recognition method based on spatial attention mechanism and channel attention mechanism.The method simplifies the recognition of chemical structure to a sequence generation task,and adopts a deep neural network model combining Convolutional Neural Network(CNN)and Long Short-Term Memory(LSTM)network to implement the transformation from chemical structure images to the SMILES sequence.The deep neural network model is composed of the encoder and the decoder.The encoder uses CNN to extract features of chemical structure images,and the decoder combines the two attention mechanisms with LSTM to generate SMILES sequences.Experimental results show that the proposed method improves the recognition accuracy to 81.63%and the BLEU-4 value to 0.937 under the condition that Beam Size equals 3,outperforming the chemical structure image recognition methods without attention mechanism or with a single attention mechanism.