基于轮廓分层描述的目标识别算法研究

Hierarchical Representation Method for Object Recognition

有效的目标轮廓分段是描述目标局部特征的关键环节.针对现有轮廓描述算法存在轮廓分段不合理的问题,本文基于认知心理学,提出了分层描述的轮廓描述算法.算法思想是首先根据角点特征将整个轮廓划分成一些轮廓分段,接着对轮廓分段的分布特点提出价值尺度,然后将多级轮廓分段按照价值尺度原则合并得到有限个能够完整描述目标轮廓的特征分段,最后将特征分段综合考虑长度尺度应用到Shape Context相似度检测模型中进行目标识别.通过对MPEG-7图像数据库中的图像进行实验分析表明该算法能够完整描述目标图像的形状特征,提高了目标识别率和形状检索率,并对部分遮挡的目标也具有良好的鲁棒性.基本满足目标识别识别和形状检索对准确率、稳定性、抗遮挡能力等方面的要求.

Efficient object contour segment is a critical problem to describe the local features of objects. In order to solve the improper contour segments obtained by the existing recognition methods, a hierarchical description algorithm of contour description is proposed. Firstly, the whole contour is divided into several contour segments by the corners on the contour. Then the valuation scale is put forward via the distribution of contour segments. Thirdly, combine these contour segments into several contour feature segments according to the valuation scale. Finally, the similarity of different contour feature segments, in combination with their lengths,is jointly used to get the best recognized results. The experimental results of MPEG-7 database indicate that this algorithm has great advantage over recently published algorithms, especially for the objects with partial occlusion. Hence, this novel algorithm satisfies the requirements of accuracy, robust and anti-occlusion in object recognition and shape retrieval.