Detection of sensitive images based on affinity propagation clustering

Abstract

ABSTRACT In order to detect sensitive images, a c ontent-based detection method using image retrieval technology is proposed in this paper. The method utilizes color correlogram and Gabor wavelet transformation to ex tract color and tex ture features, and then employs integrated similarity approach to calculate similarity and Dffinity propagation clustering algorithm to group the images in image database. In the end, the image is detected by the co ntent-based image retrieval. Experiment results show that the proposed method is better in detection accuracy ratio and retrieval efficiency than the traditional method and using K-means clustering for image database.Key words: Sensitive images, conte nt-based image ret rieval; color correlo gram; Gabor wavelet tran sformation; affinity propagation clustering; K-means clustering 1. INTRODUCTION With the rapid development of the Internet and mobile technology, more and more sensitive images, which have seriously affected teenagers, are spreading widely via Internet a nd multimedia message service (MMS), so the issue of detecting sensitive images becomes mo re important. Currently, the technology of content-based image retrieval (CBIR) is active. Therefore, CBIR for detecting sensitive imag es has become an interesting research topic. Using the color, texture, shape or other visual characteristics for feature extraction [2-4], CBIR calculates similarity between the interested image and each image in image datab ase to find the most likely images as the retrieval result. Although it can improve precision ratio, the method costs a lot of computational time when the database is large . Birds of a feather flock togetherŽ is quoted in clustering algorithm for classifying a data to a category. So we can utilize a clustering algorithm to classify imag e database into several categories. The retri eval time could be shortened greatly if the search scope is properly narrowed according to theory of the clustering. Some researchers have used K-means clustering algorithm as the first step cl ustering [5]. However, it is sensitive to noisy data. The algorithm largely depends on the selection choice of initial cluster centers. If not properly selected, the system can not yield a reasonable retrieval result [14]. To tackle this problem, the affinity propagation clustering al gorithm is used in this paper.