基于特征融合的核量化图像分类方法

摘要

Local feature descriptor and global feature descriptor are two significant descriptors of image, which play crucial role in image classification.Based on this, we propose an image classification method using kernel quantification for fusing global and local features.First, the respective pros and cons of global feature and local feature are analysed and used to extract the image features.Secondly, the features are mapped onto appropriate high dimension space through kernel approach to get the codebook and to do quantification, and the features fusion is carried out for better description on the image.Lastly, the SVM based on histogram intersection kernel is employed to classify the quantified features.It is proved through experiment the feasibility of the method proposed.%局部特征和全局特征是图像的两种重要的特征描述，在图像分类时起着至关重要的作用。据此提出一种通过融合全局与局部特征核量化图像分类方法。首先，分析全局特征及局部特征各自优缺点，并对图像进行特征提取；其次，通过核方法将特征映射到适当的高维空间中，来进行码书的获取与量化，并进行特征的融合以更好地对图像进行描述；最后，采用基于直方图交叉核的支持向量机对获取的量化特征进行分类。通过实验证明了所提出的方法的可行性。