基于规范割的空间金字塔图像分类算法

对大型图像数据库进行图像分类是很困难的,空间金字塔算法针对这种问题提出,并能得到很好的分类精度,但有几点不足.针对这些不足,提出基于规范割的空间金字塔算法:使用规范割算法对特征词进行更准确的聚类;对每类训练图像计算子特征库,利用二次聚类生成总特征库,在特征字典中保留更多的稀疏类型图像特征词;用高斯模型量化未知特征生成特征直方图,并对直方图进行尺度重整,提高类间距.实验证明提出算法比原方法分类精度最多能提高4.6%.

SPM based on normalized cut for image classification

It is difficult to classify scene images with high accuracy when the dataset is relatively large. Spatial pyramid matching was proposed to deal with this problem, but there are some shortages. As an improvement, the algorithm based on normalized cut was proposed. Normalized cut was utilized instead of K-means for clustering. The size of codebook was regulated referring to quantity and size of the images, by calculating sub-codebook for every category and re-clustering the codes. Distance between categories was enlarged by quantifying unknown features with Gaussian model and rescaling the histogram features. Experiments prove that new approach can get higher precision than the original by 4.6% at most.