In this paper, a new method is proposed for discriminating spam images from non-spam images. This method extracts edge features of a binarized image by using higher-order local autocorrelation(HLAC), and then input those features to support vector machine (SVM) for classification. Our method has three unique characteristics. First, the method extracts edge features which can represent major edge properties of an image without limitations imposed by image edgespsila directions or distributions. Second, the method can tolerate effectively slight changes of color, texture, size, layout of an image, and characteristics of text embedded in it. Third, the method is fast because of no time cost of text location and recognition. Experimental results for the public personal dataset show that the proposed method can separate spam images from non-spam images with minimum recognition rates of 98%.