Filtering Chinese image spam using Pseudo-OCR

Abstract

For image spam filtering, the Optical character recognition(OCR) based methods often achieve a better performance due to the more complex structure of recognizing corresponding text. However, applying traditional OCR techniques usually introduced shortcomings like the expensive computational cost, vulnerability to image noises and artificial interferences, especially for Chinese image spam filtering. So, by optimizing recognition procedure of traditional OCR, we propose the idea of pseudo-OCR more suitable for Chinese image spam filtering. During which discriminating the potential image spam character features from ham ones is sufficient, instead of recognizing them. What's more, a novel Chinese key-point based character feature specific for pseudo-OCR is also devised and extracted using a carefully designed algorithm, which outperforms classic corner detection methods in finding such key-points. Experiment results show that our proposed system usually has a better performance than traditional OCR based method while maintaining a low false positive rate.