KPCA (kernel PCA) is derived from PCA. It can extract nonlinear feature components of samples. However, feature extraction for one sample requires that kernel functions between training samples and the sample be calculated in advance. So, the size of training sample set affects the efficiency of feature extraction. It is supposed that in feature space the eigenvectors may be linearly expressed by a part of training samples, called nodes. According to the supposition, an improved KPCA (IKPCA) algorithm is developed. IKPCA extracts feature components of one sample efficiently, only based on kernel functions between nodes and the sample. Experimental results show that IKPCA is very close to KPCA in performance, while with higher efficiency.