We address the problem of filtering image spam, a kind of rapidly spread spam in which the text is embedded into images to defeat text-based spam filter. Particularly, we focus on image spam with Chinese text as “spam” which is a more challenging task. A popular way to detect image spam is by Optical character recognition (OCR) system, which detects and recognizes the embedded text, then followed by a text classifier that discriminate spam from ham. However, spammers start to obscure image text to prevent OCR system discovering the spam text. To compensate for the shortcomings of OCR system, a novel method which essentially is a keyword reconstruction algorithm based on Word activation force (WAF) model is proposed. It is effiective on discovering keywords, hence is benefit for the later classification stage and notably improve the performance of image spam filtering. The experimental results on a personal data set of spam images (publicly available) validate the effiectiveness of our approach that outperforms the original OCR system in practical usage with complex background in image spam.