基于特征共生矩阵的图像分类方法研究

Image classification based on a feature co-occurrence matrix

针对词袋算法(BOW)忽略局部特征空间关系的弱点,本文提出了基于特征共生矩阵的图像表达方法。该方法利用局部特征的空间共生统计代替直方图统计,充分考虑了局部特征的空间关系,增强了对图像的表达能力。实验利用标准的景物15数据库,在灰度、尺度不变特征变换(SIFT)和局部二进制模式(LBP)3个特征空间,比较了本方法、词袋法以及空间金字塔方法(SPM)的图像分类性能,结果表明本方法比词袋法的图像分类性能分别高出21.2%、6.4%、4.67%,在灰度及LBP空间,本文方法比空间金字塔法分别高出17.07%、3.87%。

In order to solve the problem whereby the bag-of-words model disregards all information about the spatial layout of features,we have used a co-occurrence matrix to study the spatial layout of features,and then represented an image with a feature co-occurrence matrix. This approach uses a co-occurrence matrix in place of a histogram.We evaluated the proposed method using a database of 15 scenes. The results show that the performance of our method was 21. 2%,6. 4%,and 4. 67%,higher than the bag-of-words image classification method in gray space,scale invariant feature transform( SIFT) space and local binary pattern( LBP) space,respectively. In gray space and LBP space,our method gave a higher performance than the spatial pyramid matching classification method by17. 07% and 3. 87%,respectively.