P300 Speller中基于权值重采样的ABSVM字符识别方法研究  

Study of ABSVM Character Recognition Method Based on Weighted Resampling in P300 Speller

作　　者：綦宏志[1,2] 孙长城[1] 许敏鹏[1] 明东[1] 万柏坤[1] 刘志朋[2] 殷涛[2] 

机构地区：[1]天津大学精密仪器与光电子工程学院,天津300072 [2]中国医学科学院生物医学工程研究所,天津300192

出　　处：《电子学报》2011年第11期2534-2539,共6页Acta Electronica Sinica

基　　金：国家自然科学基金(No.30970875;No.90920015;No.60501005);国家自然科学基金委-英国爱丁堡皇家学会联合研究项目(No.30910494);"十一五"国家高技术研究发展863计划(No.2007AA04Z236);天津市科技支撑计划重点项目生物医学工程专项(No.07ZCKFSF01300);国际科技合作专项(No.08ZCGHHZ00300)

摘　　要：P300 Speller是脑-机接口中重要的信息交互方式,由于其诱发的脑电特征信噪比较低与训练样本量庞大等问题,常规的线性识别算法和支持向量机等非线性识别算法难以获得理想的识别效率.本文引入了一种基于权值样本重采样过程的Adaptive Boosting SVM(ABSVM)方法,在大样本集上利用AdaBoost重采样方法建立一系列小样本子集,在其上训练支持向量机并将其集成后进行识别.对6位受试者P300 Speller字符辨识实验的脑电特征识别结果发现,该方法能够显著提高字符识别效率,在合并使用5次重复刺激特征的情况下字符识别准确率达到97.5%.使用国际脑机接口竞赛数据库数据进一步验证,在合并使用5次重复刺激特征的情况下该方法识别正确率较竞赛报告的最优方法提高7.35%,最大信息传输速率的提高达到48.9%.研究结果表明,ABSVM方法能够有效提高P300 Speller的识别效率和信息传输速率,值得进一步研究和发展.P300 Speller is an important information transforming method in Brain-Computer Interface.However,because the signal-to-noise ratio is comparably low and the size of training sample set is too large,the recognition efficiency in P300 Speller is not ideal by using popular linear recognition method or non linear methods such as Support Vector Machine.This paper introduces a novel method based on Adaptive Boosting SVM using weighted resampling process.We use AdaBoost resampling method to construct a series of little size training sample sub sets on the large integer set and then train the SVMs on these sub sets,finally run the recognition by combining these SVM＇s output.Using 6 subjects＇ EEG features from P300 Speller character identifying experiment we find that this method improves character identifying accuracy significantly.We achieve a character identifying accuracy of 97.5% on combining 5 repetitive stimulus features.A further evaluation using international BCI Competition dataset has proved that this method achieves a 7.35% enhancement in character identifying accuracy and a 48.9% enhancement in information transforming velocity on combing 5 repetitive stimulus features.This study demonstrated that ABSVM has the ability of improve recognition accuracy and information transforming velocity in P300 Speller and it is worthy of a further study and development.

关 键 词：自适应增强支持向量机 事件相关电位 脑-机接口 权值重采样 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]