一种基于N元ECOC的大类别K-shot满文识别方法  

A K-shot Manchu Recognition Method for Large Category Based on N-ary ECOC

作　　者：郑蕊蕊 辛守宇 周瑜 刘文鹏 党佳伟 贺建军 ZHENG Ruirui;XIN Shouyu;ZHOU Yu;LIU Wenpeng;DANG Jiawei;HE Jianjun(College of Information and Communication Engineering, Dalian Minzu University, Dalian 116600, China)

机构地区：[1]大连民族大学信息与通信工程学院,辽宁大连116600

出　　处：《郑州大学学报（理学版）》2021年第4期53-60,共8页Journal of Zhengzhou University:Natural Science Edition

基　　金：国家自然科学基金项目(61702081,61972068);辽宁省自然科学基金项目(2020-MS-134,2020-MZLH-29,20180550625)。

摘　　要：由于训练数据获取困难,满文识别被视为典型的K-shot学习问题。但在实际应用场景中,满文识别需要面对的类别数量是极大的,传统的K-shot学习算法并不适用。构建了一种面向大类别识别问题的K-shot学习算法,基本策略是利用N元纠错输出编码(error correcting output coding,ECOC)技术将原本的大类别分类问题分解为一系列较小类别的分类问题再进行处理。算法包括编码和解码两个阶段:在编码阶段,利用N元ECOC编码矩阵将大类别支持集分解为一系列小类别的子支持集,并根据子支持集生成多个K-shot学习基分类器;解码阶段利用上述基分类器对测试样本分类再合并为一个预测编码,然后将预测编码对照编码矩阵纠错,进而确定最终分类类别。实验结果表明,在500类满文数据集上获得了87.8%的识别准确率。Manchu recognition was regarded as a K-shot learning problem because it was difficult to obtain training data.However in actual application scene,the number of categories that Manchu recognition needed to face was extremely large,the traditional K-shot learning algorithm was not suitable.A K-shot learning algorithm for large category recognition was constructed in this paper.The basic strategy was that the original large-category classification problem was decomposed into a series of smaller-category classification ones by N ECOC(error correcting output coding,ECOC)technology before further processing.The algorithm was composed of two stages:encoding and decoding.In the encoding stage,the large-category support set was decomposed into a series of smaller-category sub support sets by the N ECOC coding matrix,and then multiple K-shot learning base classifiers were generated according to the sub support sets.In the decoding stage,the test sample was classified by the aforementioned base classifiers to form a prediction code.Finally,the prediction code was compared with the coding matrix for error correction to determine the final class.The proposed algorithm achieved recognition accuracy of 87.8%on a 500-category Manchu data set.

关 键 词：满文识别 小样本学习 K-shot学习 纠错输出编码(ECOC) 深度卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]