针对场景图像种类增多、场景复杂度增加和场景内容增大的趋势,本文提出了一种基于多层次特征表示的场景图像分类算法.首先采用O bject Bank目标属性的高层特征表示方法,经分类器预测出该图像所属的场景主题;然后在同一场景主题内,采用基于底层特征的局部约束低秩编码方法提取图像特征;在低秩编码方法中加入局部约束正则化并采用F-范数替代核范数的优化方法,减少计算复杂度,实现对场景图像较为细致的理解.这种由高层特征和底层特征相结合的多层次特征表示方法,从对象特征的粗理解到底层细节特征的详细解析,充分利用了不同特征间层层递进和互补的关系,实验结果证明了本文算法的有效性.%With the increases in categories ,complexity and content of scene images ,a categorization algorithm based on multi-level features representation was proposed .First ,object attributes based on high-level feature representation were available .Using simple classifiers ,the topics of scene images were exported .Then in the same topic ,the low-level feature in the image was extracted by the way of fast locality-constrained low rank coding . Meanwhile , in order to reduce the computational complexity ,the method of adding local constraint regularization and replacing kernel norm with F-norm in the processing of low rank coding was adopted to achieve detailed understanding of scene images .Achieving scene classification from coarse understanding of object characteristics to detailed analysis of low-level feature , the method can make full use of the progressive and complementary relationship between different features .The experiment results show that better classification effect is obtained .
展开▼
