基于自适应梯度最速下降的分类字典训练算法  

Discriminate Dictionary Training Algorithm Based on Adaptive Steepest Descent

作　　者：徐健[1,2] 常志国[3] 赵小强[1] 马祥[3] 

机构地区：[1]西安邮电学院通信与信息工程学院,西安710121 [2]西安交通大学电子与信息工程学院,西安710049 [3]长安大学信息工程学院,西安710064

出　　处：《计算机工程》2013年第5期225-229,共5页Computer Engineering

基　　金：国家自然科学基金资助项目(61101215);中央高校基本科研业务费专项基金资助项目(CHD2012JC012);陕西省社会发展科技攻关基金资助项目(2010K11-02-11);陕西省教育厅科学研究计划基金资助项目(11JK0994)

摘　　要：为提高稀疏表示字典用于图像分类时的正确率,提出一种基于自适应梯度最速下降的分类字典训练算法。该算法采用交替梯度下降法解决分类字典训练的双变量优化模型。为提高收敛速度,并保证稀疏表示残差和不同类别对应字典原子间的不相关性同时收敛,采用自适应步长,推导证明自适应步长的计算方法。通过固定字典,运用当前字典和训练样本计算出稀疏表示系数的下降方向和自适应步长,按照稀疏度约束将小系数置零,固定稀疏表示系数,利用稀疏表示系数和样本找到字典的下降方向和自适应步长。实验结果表明,该算法在手写字符识别中正确率能达到96.51%。To improve the correct rate of image classification using sparse representation dictionaries, a discriminate dictionary training algorithm based on adaptive steepest descent is proposed. This algorithm uses alternating gradient descent for bi-objective optimization model of discriminate dictionary training. To guarantee the convergence of sparse representation residual and the incoherence of atoms from dictionaries corresponding to different classes and increasing convergence rate, adaptive step size is adopted and the method of calculating the step size is proved. The steps of the algorithm are as following： Fix dictionaries and use current dictionaries and data samples to calculate the steepest descent directions and adaptive step sizes of the sparse representation coefficients. Small coefficients are set to zero according to sparseness Constraints. Fix sparse representation coefficients, and use sparse representation coefficients and data samples to calculate the steepest descent direction and adaptive step size of dictionaries. Experimental results of hand-writing recognition show that the correct rate is 96.51%.

关 键 词：图像分类 字典训练 稀疏表示 手写字符识别 双变量优化 自适应步长 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]