基于Conv-Res-LSTM的空中手写数字识别  

Aerial handwritten digit recognition based on Conv-Res-LSTM

作　　者：刘伎昭 王昭发 姜辉 LIU Jizhao;WANG Zhaofa;JIANG Hui(School of Computer Science,Zhongyuan University of Technology,Zhengzhou 450007,China;Henan Development and Innovation Laboratory of Industrial Internet Security Big Data,Zhongyuan University of Technology,Zhengzhou 450007,China)

机构地区：[1]中原工学院计算机学院,河南郑州450007 [2]中原工学院河南省工业互联网安全大数据发展创新实验室,河南郑州450007

出　　处：《传感器与微系统》2025年第4期83-87,共5页Transducer and Microsystem Technologies

基　　金：河南省科技攻关计划项目(232102210134);中原工学院“学科骨干教师支持计划”项目(GG202417)。

摘　　要：针对现有基于声波感知的手写数字识别方案中,一般直接将多普勒频移对应的时频图作为图片输入卷积神经网络(CNN)进行特征提取,对时序特征提取不足导致识别准确率较低的问题,提出了一种结合CNN和长短时记忆(LSTM)网络的卷积残差(Conv-Res)-LSTM模型。首先,通过Conv-Res网络对时频图的深度局部特征进行提取,然后,将特征矩阵按列展为时间序列,送入LSTM中提取时序依赖关系。在智能手表上实现了一个系统原型,实验结果表明:手写数字的平均识别精度达到了94.8%,与现有的3种基于经典CNN的方法相比平均提高了3.79%,且平均识别响应时间为167 ms。Aiming at problem that in existing acoustic wave-aware handwritten digit recognition schemes,the time-frequency map corresponding to the Doppler frequency shift is generally directly fed into convolutional neural network(CNN)as a picture for feature extraction,and the insufficient extraction of temporal features leads to low recognition accuracy,a convolutional residual long short-term memory(Conv-Res-LSTM)model combining CNN and LSTM is proposed.Firstly,the deep local features of the time-frequency map are extracted by a Conv-Res network.Then,the feature matrix is expanded into a time series by column and fed into the LSTM to extract the temporal dependencies.A system prototype is implemented on a smartwatch and experimental results show that the average recognition precision of handwritten digits reach 94.8%,which is average improved 3.79%,compared to existing three classical CNN-based methods,and the average recognition response time is 167 ms.

关 键 词：手写数字识别 多普勒频移 时频图 长短时记忆网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]