基于多方向空间词袋模型的物体识别  

Object recognition based on multi-direction spatial Bags of Words model

作　　者：齐梅[1] 胡敏[2] QI Mei;HU Min(Anhui Open University, Hefei 230022, China;HeFei University of Technology, Hefei 230022, China)

机构地区：[1]安徽广播电视大学,合肥230022 [2]合肥工业大学,合肥230022

出　　处：《计算机工程与应用》2017年第7期197-201,共5页Computer Engineering and Applications

基　　金：国家自然科学基金-广东联合基金重点项目(No.U1135003);安徽省科技计划项目(No.1206c0805039)

摘　　要：针对词袋模型完全忽略空间位置信息的问题,提出了一种多方向空间词袋模型的物体识别方法。该算法通过空间金字塔划分,形成图像的空间子区域特征表达;分别在水平、垂直和倾斜±45°上对图像局部特征向量进行投影,得到图像在多方向上的空间结构信息;采用样本视觉词典方法,既减少了不同物体类别样本带来的冗余影响,又降低了特征维数。在Caltech101和Caltech256物体库上进行了对比实验,实验结果验证了算法的有效性。To alleviate the problem that spatial position information is completely ignored in Bags of Words representation,an optimized method based on multi-direction spatial Bags of Words is proposed.Firstly,the representation of spatialsub region of an image is modeled by spatial pyramid.Then,projection method is applied to local features of imageblocks in horizontal,vertical and inclined±45°,spatial structure information is well expressed in multi-direction.Furthermore,it reduces redundant effects of different object and the dimension of features by means of samples visual codebook.Finally,the proposed method is evaluated on two object databases,the results of comparative experiment show thatthe proposed algorithm has inspiring performance.

关 键 词：物体识别 多方向投影 空间信息 样本词典 词袋模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]