基于BiLSTM-Attention的钢板表面手写板号识别算法  

Handwritten board number recognition algorithm on steel plate surface based on BiLSTM-Attention

作　　者：徐萌 王雪飞 XU Meng;WANG Xue-fei(College of Electronic Information and Automation,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学电子信息与自动化学院,天津300300

出　　处：《中国冶金》2021年第10期86-93,共8页China Metallurgy

基　　金：国家重点研发计划资助项目(2018YFB1701601)。

摘　　要：国内钢铁企业生产厂的信息化物料跟踪大都依赖于钢板号。由于生产流程复杂,急需高准确率的板号在线识别技术。自然场景下机器喷号的识别技术较成熟,但复杂场景下的手写板号难以实现自动识别。针对复杂工作场景下钢板表面手写板号特点,提出一种以BiLSTM-Attention为主体结构的深度学习算法。首先结合复杂场景,对图像数据进行预处理,保证模型输入图片质量;然后利用残差神经网络(ResNet)提取图片特征、利用双向长短期记忆网络(BiLSTM)提取基于图像的序列特征;最后基于注意力机制捕获序列内的信息流,对每个字符的特征进行整合,形成文本特征向量以预测输出序列。经现场测试,实现钢板表面手写板号识别任务准确率达86.15%,结果表明算法可行有效,满足实际生产需求。The information-based material tracking of domestic steel production plants all relies on steel plate numbers.Due to the high complexity of production process,there is an urgent need for online plate number recognition with high accuracy.The recognition technology of machine spray numbers in natural scenes is relatively mature,but it is difficult to realize automatic recognition of handwriting board numbers in complex scenes.A deep learning algorithm with BiLSTM-Attention as the main structure based on the characteristics of handwritten board number on the surface of steel plate in complex work scenarios was proposed.First,combined with complex scenes,image data was preprocessed to ensure the quality of model input images,then residual neural network(ResNet)was used to extract image features,and Bi-directional Long Short-Term Memory(BiLSTM)was used to extract the sequence features of image.Finally,the attention mechanism(Attention)was based to capture the information flow in the sequence,and the feature of each character was integrated to form a text feature vector to predict the output sequence.After on-site testing,the result showed that the algorithm was feasible and effective,and the accuracy of recognition task of handwritten board number on the surface of steel plate reached 86.15%,which met the actual production needs.

关 键 词：双向长短期记忆网络 注意力机制 神经网络 手写钢板号 手写文本识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]