A novel method for color image feature extraction is proposed. First of all, non-iterative bilateral projection based 2DPCA algorithm (NIB2DPCA) is employed to extracted feature information from three channels of a given color image respectively. Then the three extracted feature matrices are reconstructed to form a two-dimension middle matrix. After that, NIB2DPCA is employed again to extract features from the middle matrix to obtain the final features. Experimental results on CVL and FEI databases show that, contrast to existing similar methods, the recognition accuracy of our method increased by 7-16 percent.