基于轮廓几何稀疏表示的刚性目标模型及其分级检测算法

刚性目标轮廓具有明显几何特性且不易受光照、纹理和颜色等因素影响.结合上述特性和图像稀疏表示原理,提出一种适用于刚性目标的分级检测算法.在基于部件模型(Part-based model, PBM)的框架下,采用匹配追踪算法将目标轮廓自适应地稀疏表示为几何部件的组合,根据部件与目标轮廓的匹配度,构建描述部件空间关系的有序链式结构.利用该链式结构的有序特性逐级缩小待检测范围,以匹配度为权值对各级部件显著图进行加权融合生成目标显著图. PASCAL图像库上的检测结果表明,该检测方法对具有显著轮廓特征的刚性目标有较好的检测结果,检测时耗较现有算法减少约60%~90%.

A Rigid Object Detection Model Based on Geometric Sparse Representation of Profile and Its Hierarchical Detection Algorithm

The profile of rigid objects has the geometrical characteristic and is insusceptible to illumination, texture or color. In this paper, a hierarchical detection algorithm for ridge objects based on geometric sparse representation of profile is presented. In the framework of part-based model(PBM), the object profile is automatically divided into geometrical parts by the sparse representation using the matching pursuit algorithm. To describe the spatial relationship of the geometrical parts, an ordered chain-like structure is constructed according to the order of the matching degree of the parts and the object profile. With the ordered chain-like structure, the detection range is gradually shrunk at each hierarchy. The final salient map of the object is the weighted summation of the parts'' salient maps, and the weights are defined as the matching degrees. The simulation on the PASCAL datasets shows that the proposed method outperforms the existing models in rigid objects detection, and saves 60% to 90% detection time compared to the state-of-art methods.