一种快速的全局场景分类算法  

A fast algorithm for holistic scene classification

作　　者：付毅[1] 田畅[1] 吴泽民[1,2] 曾明勇[1] 胡银记 

机构地区：[1]解放军理工大学通信工程学院,江苏南京210007 [2]光电控制技术重点实验室,河南洛阳417009

出　　处：《红外与激光工程》2013年第S01期242-248,共7页Infrared and Laser Engineering

基　　金：光电控制技术重点实验室和航空科学基金联合资助项目(20125186005)

摘　　要：传统的空间金字塔匹配算法是一种成功的场景分类方法,先利用SIFT特征和K均值聚类的思想,构建"视觉字典",再形成包括图像空间信息的"空间金字塔"。传统算法构建"视觉字典"时采用了128维SIFT描述子,存在特征描述符计算复杂、聚类时间长等问题。针对这些问题,提出一种基于改进SIFT的空间金字塔匹配方法,采用圆形局部区域和梯度、灰度特征将特征描述符压缩到只有18维。结合特征匹配和SVM训练器,通过与原空间金字塔匹配算法进行场景分类性能对比,结果表明:该方法在特征提取和聚类速度方面具有显著提升,同时在分类准确度上也略有提高。Traditional spatial pyramid matching algorithm is successful for scene classification, based on SIFT descriptor and K - means cluster algorithm, by constructing ＇ spatial Pyramid＇ with ＇ visual words＇. But it has met the problems that the calculation of SIFT descriptor was complex and clustering time is long. Aiming these, this paper presents a new spatial pyramid matching algorithm based on improved SIFT, using a circular window with 12 gradient values and gray - scale difference to construct the feature descriptor with 18 dimension. With fea- tures matching and SVM tool, we simulate the performance of the new algorithm compared with the traditional one, and the experiments show that the new one is more accurate and take less time in feature description and cluster.

关 键 词：空间金字塔匹配 SIFT 特征描述 场景分类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]