The Research Of Detecting Image Spam Based On K-labels Propagation Model

With the development of Internet technology, communication via email is becoming more andmore essential in people’s daily life. But the volume of spam has grown tremendously, and itbecame the most controversial byproduct of the Internet, especially in Image Spam. So it bringsabout the phenomenon that not only resources are wasted seriously, but also the network may isclogged and palsied, and the normal communication between users is affected. Therefore, it isnecessary to develop a high-efficiency detection technology of Image Spam.In connection to the detection technology of Image Spam, the background, purpose andmeaning are introduced, the definition, features, construction methods and detectable difficulty aredescribed, and some common detection methods are analyzed in this thesis. For the disadvantagesof existing methods, a method based on KLPM (K-labels propagation model) to detect Image Spamis proposed in this thesis. The main work and contribution of the thesis are:(1) The method based on Newton sparse representation is proposed to eliminate the noise of theimage: Newton Method is used to handle the coefficient after wavelet transform and make thecoefficient sparse. While this method is able to eliminate image noise, it also can keep the imagedetails as much as possible, so as to improve the precision of the detection method.(2) SURF algorithm is used to extract scale and rotation invariant features of the image. Thisalgorithm can keep certain invariance and distinguishable to the change of image, such as imagerotation, affine transform, scale zooming and so on.(3) The improved Means Clustering Algorithm is proposed to cluster image features, composefeature vector, and compute the similarity.And it guarantees the comparability between the images.(4) The LPM (Label Propagation Model) is proposed: first all images are labeled, then a completelyconnected graph (in this graph, each node is regarded as an image) with weight can be drawn, nextlabels are propagated on the graph, last all testing images are classified. After updated LPM, themethod based on KLPM is proposed: firstly, the image noise is eliminated with the method basedon Newton sparse representation; secondly, the SURF of all images are extracted and clustered;lastly, all testing images are classified via KLPM. According to the simulation experiment, themethod based on KLPM has ideal results.