摘要： 基于类属超图模型给出简单图像和复杂图像目标的识别方法。通过提取简单图像的稳健尺度不变特征变换特征，得到其对应的属性图，采用RSOM聚类树的思想和K近邻方法快速实现对简单图像的目标识别。复杂图像存在较大的背景干扰和遮挡的影响，通过滑动窗方法在待识别图像中定位待识别目标区域，并将该区域从待识别图像中分出，然后采用与简单图像识别方法类似的方法完成目标识别，减少背景干扰和遮挡的影响。仿真实验表明，2种图像目标识别方法是有效的。

关键词: 图, 类属超图, 尺度不变特征变换, 目标识别

Abstract: This paper proposes object recognition methods for images with simple imaging conditions and challenging imaging conditions, which are based on Class Specific Hyper Graph(CSHG) model. In the process of recognition for images with simple imaging conditions, it extracts their robust Scale Invariant Feature Transform(SIFT) features and describe them using graphs. The objects in test images are recognized efficiently by using a RSOM clustering tree and K-nearest neighbor method. In the process of recognition for images with challenging imaging conditions, the approximate interest object regions in test image are located by sliding window method. The approximate object regions are expanded or shrunk iteratively and their corresponding graphs matche to graphs in CSHG model. The exact object regions are located by checking the number of matching features and segmented from test images. K-nearest neighbor graphs of the object regions are obtained in CSHG model and final recognition decision are made by using a majority voting strategy. Experimental results demonstrate that the methods are effective.

Key words: graph, class specific hyper graph, Scale Invariant Feature Transform(SIFT), object recognition

中图分类号: 

• TP391