In order to solve the problem that image interpolation based on adaptive steering kernel regression results in whole gray deviation and some empty pixels in interpolated images, an improved adaptive kernel function for image interpolation is proposed. The improved kernel function was constructed by merging the function of the geometry distance of pixels in neighborhood into adaptive steering kernel, which is produced according to the principle of self-similarity of image. Experimental results show that the proposed method for irregularly and regularly sampled image has a much better effect than interpolation methods based on classical kernel and adaptive steering kernel.