Since different kinds of heterogeneous features (such as color, shape and texture) in image shave different intrinsic discriminative power for image understanding, this paper proposes a multiple kernel learning with group sparsity (MKLGS) to select groups of discriminative features for image annotation to effectively utilize those heterogeneous visual features. Given each image label, the MKLGS method embeds the nonlinearity image data with discriminative features into a Hilbert space, and then utilizes the kernel function in the Hilbert space and group LASSO to select groups of discriminative features. Finally, a classification model can be trained for image annotation. In comparison to other image annotation algorithms, experiments show that the proposed MKLGS for imageannotation achieves a better performance.