基于Tri-training的多特征融合图像检索

Multi-feature image retrieval based on Tri-training

为了有效地综合利用图像的多种底层特征进行图像检索,提出将Tri-training方法应用于图像检索过程,将图像的颜色、纹理和形状特征进行了有效的融合。分别提取图像的三维量化颜色直方图、方向可控金字塔二值图像投影和仿射不变区域来表示其颜色、纹理和形状特征,并将三种特征的匹配值作为Tri-training分类器的输入对分类器进行训练和测试。实验结果表明,该方法有效利用了图像的多种特征,达到了很好的检索效果。

To effectively use multi-features in images for image retrieval,this paper used Tri-training to fuse the color feature, texture feature and shape feature.First,it extracted the three-dimensional quantized color histogram,steerable pyramid binary image projection and affine invariant regions to represent the color feature,texture feature and shape feature of the image. Then,the matching results of these three features served as the input of the Tri-training classifier for its training and testing. Experiments show that by the fusion of these features,this method is very effective.