基于深度学习的吉他谱识别  

Guitar Tablature Recognition Based on Deep Learning

作　　者：陈超艺 陈新度[1,2] 吴磊[1,2] CHEN Chao-yi;CHEN Xin-du;WU Lei(Guangdong Provincial Key Laboratory of Computer Integrated Manufacturing;State Key Laboratory of Precision Electronic Manufacturing Technology and Equipment,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学广东省计算机集成制造重点实验室 [2]广东工业大学省部共建精密电子制造技术与装备国家重点实验室,广东广州510006

出　　处：《软件导刊》2022年第1期141-145,共5页Software Guide

摘　　要：光学乐谱识别是音乐智能化发展的关键部分,在音乐教学、创作等领域有重要应用价值。针对现有吉他谱识别方法步骤繁琐、识别精度低等问题,提出一种基于深度学习的吉他谱识别方法。首先通过分析吉他谱的特点,将吉他谱水平分割为品格音符图像、减时线图像、休止符图像和增时线图像;然后将品格音符图像依次与减时线图像叠加,输入到第1个CRNN模型中进行识别,将减时线图像、休止符图像和增时线图像叠加,输入到第2个CRNN模型中进行识别;最后将识别出的各个符号全局关联,获取完整的乐谱语义。实验结果表明,基于深度学习的吉他谱识别方法可达到98.3%的品格音符识别准确率和99.1%的时值音符识别准确率,与传统的吉他谱识别方法相比,该方法具有更快的识别速度与更高的识别精度。Optical music recognition is a key part of the development of intelligent music,and has important value in the fields of mu⁃sic teaching and creation.Aiming at the cumbersome steps and low recognition accuracy of the existing guitar tablature recognition methods,a guitar tablature recognition method based on deep learning is proposed.First,by analyzing the characteristics of the guitar tablature,the guitar tablature is divided into fret note image,minus time line image,rest image and increased time line image;then,the fret note image is superimposed with the minus time line image and input to the firstCRNN model for recognition,and superimposes the reduced time line image,the rest image and the increased time line image,and inputs them to the second CRNN model for recogni⁃tion;finally,the recognized symbols are globally associated to obtain the complete musical score semantics.Through experimental analysis,guitar tablature recognition method based on deep learning can achieve 98.3%accuracy of character note recognition and 99.1%accuracy of time value note recognition,compared with the traditional guitar tablature recognition method,this method has fast⁃er recognition speed and higher recognition accuracy.

关 键 词：光学乐谱识别 吉他谱识别 深度学习 CRNN 图像识别 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]