基于稀疏编码和多核学习的图像分类算法

Image Classification Using Multiple Kernel Learning and Sparse Coding

提出了一种基于稀疏编码和多核学习的图像分类算法.首先从图像中提取Dense-SIFT(Dense Scale-Invariant Feature Transform)和Dense-SURF(Dense Speeded Up Robust Feature)2种特征,使用稀疏编码对特征点进行处理,得到一系列高维向量,然后对这些高维向量应用max-pooling算法,将图像表示成单个向量.最后,使用改进的多核学习方法对这些向量进行分类,对于不同的特征,使用不同核的组合以达到最好的分类效果.实验结果表明,该算法作为词袋(BoW)模型的改进,能够提高分类准确率.

An image classification algorithm based on sparse coding and multiple kernel learning （MKI.） was proposed. First, D-SIFT （Dense Scale-Invariant Feature Transform） and D-SURF （Dense Speeded Up Robust Feature） are extracted from images. Then, sparse coding method is adopted to represent an im- age as a vector and max pooling method is also utilized for both features. Finally, an improved MKL is used to classify those vectors. Appropriate kernel combinations are selected for each feature and the final result is the fusion of both. The experiments demonstrate that the algorithm remarkably improves the clas- sification accuracy.