摘要
为了检测颜色、纹理、形状变化较大的目标,提出一种基于边缘片段特征的目标检测算法,自动提取训练样本的边缘片段作为模板,不需要任何手工分割和人工筛选过程。通过改进的统计区域合并分割算法以及自适应提升学习算法可以有效地减小由于自动提取边缘片段而引入的噪声的影响。每一个弱分类器都综合考虑了训练样本与边缘片段的形状匹配代价和空间位置匹配代价。由辨别能力较强的弱分类器组成一系列由简到繁的星状模型。实验结果表明:该检测算法对于颜色、纹理、形状变化的鲁棒性较强。在耳朵直立的宠物狗数据库上,检测率可达到90%。
A detection algorithm was developed to detect objects with large variances in color, texture, and shape. The algorithm does not need any manual selection or segmentation to automatically extract edge fragments from training samples to form a dictionary of templates. A modified statistical region merge segmentation algorithm and the Adaboost learning strategy are used to dramatically reduce the effect of noise caused by the automatic extraction of the edge fragments. The matching cost used in each weak classifier is based on the shape and distance similarities between the training samples and the edge fragments. Then the Cascaded Adaboost is used to select discriminative edge fragments to form a series of star models going from simpler models to more complex models to detect objects. The experimental results show that the method is very robust to large variances in color, texture, and shape. The system has a 90% detection rate on a database of up-right ear dogs.
出处
《清华大学学报(自然科学版)》
EI
CAS
CSCD
北大核心
2008年第10期1640-1642,共3页
Journal of Tsinghua University(Science and Technology)
基金
国家自然科学基金资助项目(60472028)
高等学校博士学科点专项科研基金项目(20040003015)
关键词
机器学习
目标检测
边缘片段特征
自适应提升算法
图像分割
machine learning
object detection
edge fragment features
Adaboost algorithm
image segmentation