Abstract: Aiming at the problem that high dimension of SIFT feature description operator causes large calculating scale and high complexity in SIFT image feature matching algorithm, this paper proposes an improved SIFT image feature matching algorithm based on kernel projection. Instead of using smoothed weighted histograms to calculate gradient modulus and gradient orientations, this paper proposes an improved scheme based on kernel projection. It reduces the dimension of SIFT feature description and improves the efficiency of feature matching. Experimental result proves that the improved algorithm has higher matching accuracy and has less matching time, and it has higher instantaneity.

Key words: Scale Invariant Feature Transform（SIFT）, image matching, feature description, kernel projection

摘要： 针对传统SIFT图像特征匹配算法因其特征描述算子维度过高而造成的计算量大、实时性差的问题，提出一种基于内核投影的改进SIFT图像特征匹配算法。传统SIFT特征匹配算法采用平滑加权直方图计算特征点的梯度模值和梯度方向。采用内核投影算法对其进行改进，使生成的特征描述算子的维度降低，从而能够提高特征匹配效率。实验结果表明，改进后的SIFT算法具有较高的匹配精度，同时匹配时间有所减少，使实时性得到提高。

关键词: 尺度不变特征变换（SIFT）, 图像匹配, 特征描述算子, 内核投影