基于Adaboost权值更新以及K-L距离的特征选择算法

对于颜色、纹理变化较大的目标, 边界片段是一种较为稳定的特征. 手工分割样本再提取边界片段的传统做法由于工作量巨大而限制了样本的数目, 不能满足统计学习中大样本训练的要求. 但是如果对训练样本采用自动分割的方法就不可避免地引入很多背景中的噪声. 在这种情况下特征的选择就显得尤为关键. 本文提出一种基于Adaboost权值更新以及K-L距离的特征选择算法, 在Adaboost的每一轮训练中动态地选择所有备选边界片段的一个子集作为Adaboost训练的特征集. 选择算法以边界片段在正面样本与负面样本中分布的鉴别信息为依据, 有效地减少了背景中边界片段的干扰. 实验证明该算法是有效的.

Feature Selection Based on Weight Updating and K-L Distance

Edge-fragment feature is very stable in detecting objects with large variances in color, texture, and shape. Traditional methods that extract edge-fragments from a few manually segmented samples cannot meet the requirement of statistical learning in case of large number of training samples. However, if the training samples are automatically segmented, it is inevitable that huge amount of edge-fragments from background of the training samples will appear in the feature set. In that case, the feature-selection algorithm is very critical to the detection task. In this paper, a feature-selection algorithm based on weight updating scheme of Adaboost and K-L distance is proposed. In each round of Adaboost learning, a subset of all the edge-fragments is selected as the feature set for training Adaboost weak classifier. Because the proposed feature-selection algorithm takes into account the edge-fragments' discrimination information between positive samples and negative samples, it can effectively reduce the number of edge-fragments from background in the final classifier. Experimental results show that the proposed algorithm is effective.