Compared with the dictionary learning and recognition method based on sparse constraints, the projective Dictionary Pair Learning (DPL) has faster learning speed and a higher recognition rate. In order to further improve the recognition ability of DPL, this study proposes an improved DPL algorithm K-DPL, which combines Kernel Principal Component Analysis (KPCA) and DPL method. In K-DPL, by using the kernel technique, the samples are mapped to high-dimensional space to solve the nonlinear problem, and then DPL is used to get more discriminant dictionary by training the dictionary. Experiments on ORL datasets show that recognition rate is increased by at least 1.5% and the recognition speed is increased by about 20 times compared to DPL at different training ratios by using K-DPL. On the extended YaleB and AR datasets, compared with DPL, the recognition rate is increased by 0.3% and 0.4% respectively, and the recognition speed is improved by using K-DPL as well. It indicates that K-DPL has good robustness to illumination and occlusion.