基于SVM稀疏表示的类特别字典学习算法  

Class-specific dictionary learning algorithm based on SVM sparse representation

作　　者：宋银涛 杨宝庆 刘计 赵宇 闫敬[2] SONG Yintao;YANG Baoqing;LIU Ji;ZHAO Yu;YAN Jing(School of Information Engineering,Yangzhou University,Yangzhou,Jiangsu 225009,China;School of Electrical Engineering,Yanshan University,Qinhuangdao,Hebei 066004,China)

机构地区：[1]扬州大学信息工程学院,江苏扬州225009 [2]燕山大学电气工程学院,河北秦皇岛066004

出　　处：《燕山大学学报》2024年第5期437-445,共9页Journal of Yanshan University

基　　金：国家自然科学基金资助项目(62205283)。

摘　　要：近年来,深度学习对大规模训练样本的依赖性成为一个突出问题。在面对小样本数据集时,字典学习算法被提出作为一种解决方案。为了进一步提升字典学习在图像分类领域的竞争优势,本文提出了一种基于支持向量机的类特别字典学习算法。该算法创新性地引入了类特别系数相异性约束项。该约束项将原本独立的重建项、稀疏项和判别项融合为一个统一的学习框架,以显著提升字典的判别能力。实验证明,该模型的分类性能优于其他先进的字典学习模型。此外,本文提出将深度学习预训练与字典学习算法相结合的方式,通过实验证明该方式可以显著提升字典学习算法在大规模训练样本中的分类性能。In recent years,the dependence on large-scale training samples in deep learning has become a prominent issue.Dictionary learning algorithms have been proposed as a solution for small sample datasets.To further enhance the competitive advantage of dictionary learning in image classification,a class-specific dictionary learning algorithm based on support vector machine is proposed in this paper.The coefficient disparity constraint is introduced innovatively.The constraint term fuses the originally independent reconstruction,sparse,and discriminative ters into a unified learning framework,significantly improving the discriminative ability of the dictionary.It has been demonstrated through experiments that the classification performance of this model outperforms other state-of-the-art dictionary learning models.Additionally,a method to combine deep leaming pre-training with dictionary learning algorithms is proposed,which has been experimentally demonstrated to significantly improve the classification performance of dictionary leaming algorithms in large-scale training samples.

关 键 词：字典学习 稀疏表示 支持向量机 系数相异性约束项 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]