基于核竞争学习算法的图像特征提取

利用核函数学习可有效解决图像特征线性不可分的特性,结合稀疏表示算法的优势,提出了一种新的图像特征提取方法。采用基于竞争学习规则的独立分量分析法对图像进行稀疏表示,该算法可提取数据的高维特征,且不需要优化高阶的非线性函数和进行稀疏密度估计,因而有较快的收敛速度。与仅使用基于竞争学习的独立分量分析法相比,在PolyU数据库上的实验结果表明,采用基于核函数学习和稀疏表示相结合的方法所提取的数据特征有利于提高特征分类精度。

Image Feature Extraction Using Kernel Winner-take-all Based on Independent Component Analysis Algorithm

Utilized the property that kernel function learning can solve efficiently the linearly inseparable problem of image features,and combined advantages of sparse representation algorithm,a novel image feature extraction method is proposed.Here the sparse representation of images is realized by the winner-take-all rule based independent component analysis method,which can extract image high-dimension features efficiently and has quicker convergent speed,because it is unnecessary to optimize high-order nonlinear function and estimate sparse density.Compared with the winner-take-all rule based independent component analysis method,the experiment results in PolyU database testify that image features extracted by the combination method of kernel function learning and sparse representation can favor to improve the precision of feature classification.