For compressive sampling and classify of images, an orthogonal transform basis based on graph presentation is proposed. Firstly, according to correlation of both edges and pixels, an improved graph presentation of an image is introduced. Then the orthogonal transform basis is constructed as the eigenvectors matrix after eigenvalue decomposition of the Laplacian of the graph, based on which the sparse representation of images is obtained. Finaiiy, the random projection of compressed features is used to train and test the classifier directly in the compressed domain. The proposed method has a similar classifying performance in the original domain with abundantly reduced training and testing time, as well as computational/store cost.