基于改进的LBG算法的SVM学习策略  

The SVM Learning Strategy Based on Improved LBG Algorithm

作　　者：李滔[1] 王俊普[1] 吴秀清[2] 张邵一[1] 

机构地区：[1]中国科技大学自动化系,合肥230027 [2]中国科技大学电子工程与信息科学系,合肥230027

出　　处：《复旦学报（自然科学版）》2004年第5期789-792,共4页Journal of Fudan University：Natural Science

基　　金：国家863项目资助(2002AA783055)

摘　　要：针对SVM方法在大样本情况下学习和分类速度慢的问题,提出了利用LBG算法对训练样本进行预处理,然后再使用传统的SVM算法进行训练的策略,并提出了一种改进的LBG算法.通过对仿真数据以及对实际的纹理图像的分类实验表明,这种预处理方法能在保持学习精度的同时减小训练样本以及决策函数中支持向量集的规模,从而提高学习和分类的速度.The strategy of using LBG algorithm to compress the example set is proposed to accelerate the speed of learning and classification in SVM applications. The original LBG algorithm is modified to get more stable codebooks. This strategy is used to classify artificial data and texture images from the Brodatz image set. The experimental results show that by this strategy both the scale of training data and the resulting support vector sets are effectively compressed, so the processes of learning and classifying are accelerated while keeping almost the same classification performance.

关 键 词：SVM算法 训练样本 决策函数 纹理图像 支持向量 大样本 数据 学习策略 真数 实际 

分 类 号：G633[文化科学—教育学] TP183[自动化与计算机技术—控制理论与控制工程]