摘要： 针对目前直接提取图片文本特征费时且分类准确率不高，以及使用图像属性特征过滤垃圾图片召回率低下等问题，提出一种快速有效的垃圾图片过滤方法。在使用4-gram切分Base64编码后的图片文本后，通过Binary特征将图片特征项表示为Binary向量，并训练支出向量机分类器来识别垃圾图片。实验结果表明，该方法不仅能够识别不同格式的垃圾图片，而且垃圾图片识别精确率、召回率和F1值分别可达99.85%、99.49%和99.67%。

关键词: 垃圾图片, Base64编码, 4-gram分词, 支持向量机

Abstract: Extracting embedded text from images to filter image spam is usually time-consuming and can not reach high classification accuracy. On the other hand, filtering image spam using image properties features has low recall rates problem. This paper proposes a simple but effective method to detect image spam. By tokenizing Base64-encoded image text into a series of 4-gram features and representing them as a binary vector, a trained Support Vector Machine(SVM) can distinguish spam images from legitimate ones very well. Experimental results show that the method achieves satisfactory performance in filtering image spam with different formats, with the precision, recall and F1 of 99.85%, 99.49% and 99.67% respectively.

Key words: image spam, Base64 encoding, 4-gram, Support Vector Machine(SVM)

中图分类号: 

• TP18