2DPCA-SIFT:一种有效的局部特征描述方法

2DPCA-SIFT:An Efficient Local Feature Descriptor

PCA-SIFT(Principal component analysis–scale invariant feature transform)方法通过对归一化梯度向量进行PCA降维,在保留特征不变性的同时,有效地降低了特征矢量的维数,从而提高了局部特征的匹配速度.但PCA-SIFT中对本征向量空间的求解非常耗时,极大地限制了PCA-SIFT的灵活性与应用范围.本文提出采用2DPCA对梯度向量块进行降维的特征描述方法.该方法相比于PCA-SIFT,可以快速地求解本征空间.实验结果表明:2DPCA-SIFT在多种图像变换匹配和图像检索实验中可以实现与PCA-SIFT相当的性能,并且从计算效率上看,2DPCA-SIFT具有更好的扩展性.

Principal component analysis - scale invariant feature transform （PCA-SIFT） applies principal components analysis （PCA） to the normalized gradient vector. It effectively reduces the dimension of feature representation and improves the matching speed while maintaining the descriptor＇s invariance. However, PCA-SIFT needs an additional step of eigenspace computation which is time-consuming. This step greatly limits the flexibility and applications of PCA-SIFT. In this paper, we adopt the 2DPCA to reduce the descriptor＇s dimension and build the descriptors. Compared to the PCA-SIFT, this method can finish the eigenspace calculation in real time. The experiments show that the proposed method can get competitive performance when compared to PCA-SIFT in different image matching and image retrieval applications, and can be easier to be expanded for its good computational efficiency.