计算机工程与应用 ›› 2018, Vol. 54 ›› Issue (3): 206-211.DOI: 10.3778/j.issn.1002-8331.1706-0267

• 图形图像处理 • 上一篇    下一篇

基于兴趣域检测的空间金字塔匹配图像分类

周华兵,朱国家,张彦铎,任世强   

  1. 武汉工程大学 湖北省智能机器人重点实验室,武汉 430205
  • 出版日期:2018-02-01 发布日期:2018-02-07

Image classification based on region of interest dection and spatial pyramid matching

ZHOU Huabing, ZHU Guojia, ZHANG Yanduo, REN Shiqiang   

  1. Hubei Provincial Key Laboratory of Intelligent Robot , Wuhan Institute of Technology, Wuhan 430205, China
  • Online:2018-02-01 Published:2018-02-07

摘要: 在图像分类过程中,一个图像区域内起着决定性作用的对象位置和大小都不确定,直接使用空间金字塔匹配算法的分类准确率并不高。为此,提出了一种基于兴趣域检测的空间金字塔匹配方法可以有效改善分类准确率。首先利用检测器得到的定位结果,证实了在图像分类里使用一种主流的目标检测算法去将图像的目标和背景分离,分别得到前景和背景的可行性,然后使用粗目标对齐方式匹配,为这两个区域分别构建基于空间金字塔匹配算法的空间特征直方图,最后结合检测器提供的兴趣域检测评分与支持向量机提供的评分为分类结果重评分。实验结果表明,所提方法比使用标准的空间金字塔匹配算法得到的平均准确度均值提升超过12%,同时在与三种主流算法的对比中,所提方法平均准确度均值最高,并且在超过一半的图像类别中获得了最高的平均准确度。

关键词: 图像分类, 空间金字塔匹配, 兴趣域, 特征直方图, 平均准确度均值

Abstract: In the process of image classification, the image region containing object which plays a decisive role is indefinite in both position and scale, and it can not get a high accuracy of image classification by using Spatial Pyramid Matching(SPM) directly. Therefore, a method for improving the performance of image classification based on Region of Interest (ROI) detection and Spatial Pyramid Matching(SPM) is proposed. This method first makes use of localization results of detector, and it verifies feasibility of using a state-of-the-art object detection algorithm to separate the goals of image and background for image classification. Then, a method which is called coarse object alignment matching is used to construct spatial histogram features for these two regions separately based on SPM. Finally, the scores provided by real detector and Support Vector Machine(SVM) are combined to rescore for the final result. Experimental results demonstrate that the mean average precision of proposed method has been promoted over 12 percent than the standard SPM. Compared to three state-of-the-art methods, the proposed method gets the highest mean average precision, and gets the highest average precision at more than half image categories.

Key words: image classification, spatial pyramid matching, region of interest, spatial histogram features, mean average precision