Kernel-based methods have been widely used in pattern recognition. But traditional kernel functions can only process 1D vectors, while image data are often 2D matrices. This paper presents a new kernel function based on RBF kernel function for image target recognition. This new kernel function can directly accept 2D image data as input data, and analyze the structural information of the targets in the images, which are often ignored by other kernel functions. This kernel function can also process target images which are obtained under different luminance conditions without any preprocessing. The experiments on UCI datasets and ALOI datasets show that, the classifier based on our kernel function can have higher classification accuracy. Because the features are not necessary in our method, the results also demonstrate a new framework of image target recognition.