Multi spatial resolution for image spam filtering

Abstract

Image spam filtering is a challenging task because spammers are constantly creating new tricks and making anti-spam filters ineffective. To overcome these problems, many new techniques have been developed. Most of these techniques use a basic bag-of-features representation where global approach is used to extract the feature. This global representation leads to limited descriptive power for the features due to neglecting the spatial information, which can create powerful cues for classification tasks. Spatial Pyramid Representation (SPR) is one of the most effective and widely used image processing techniques that embedding spatial information into a feature. Inspired by this technique, we propose Multi Spatial Resolution (MSR) approach, which transform the image to Base-64 encoding, divided the Base-64 encoding into a sequence of increasingly finer grids on different pyramid level. The n-gram technique is used to extract the features from each grid cell or partition. Frequency histogram for each partition is concatenated to form a single feature vector. The experiments were conducted on Dredze and SpamArchive data sets at four different resolutions using SVM classifier. The results show that MSR increased the classification performance compared to global approach.