A new image representation based on distribution of local invariant features to be used in a discriminative approach to image categorization is presented. The representation which is called Probability Signature (PS) is combined with character of two distribution models Probability Density Function and standard signatures. The PS representation retains high discriminative power of PDF model, and is suited for measuring dissimilarity of images with Earth Mover's Distance (EMD), which allows for partial matches of compared distributions. It is evaluated on whole-image classification tasks from the scene and category image datasets. The comparative experiments show that the proposed algorithm has inspiring performance.