Análise de características para detecção de nudez em imagens

Abstract

The popularization of Internet access has lead institutions and parents to face serious problems on preventing employees, as well as children, to have access to inappropriate content, such as pornographic pages. This kind of content is available in different forms, including videos, sounds, text, and especially images, on the Web. Since most of this inappropriate content is provided as images, it is necessary to employ strategies which allow the analysis of image content in order to control access to inappropriate content. In this context, nudity detection in images plays an important role. Several approaches apply skin detection as a key step toward nudity detection. Skin detection is a difficult task due to the fact that it is necessary to use skin filters robust to shade variations caused by light. In addition, these methods employ a combination of features based on color, texture and shape, which may increase the complexity and time processing of detection algorithms. In despite of this drawback, feature analysis, or selection, is not carried out in most of the work available in the literature, The objective of this work is to investigate the features most frequently used in the literature for the description of nude images, as well as to select the most relevant subset of features taking into account classification accuracy. The feature analysis is carried out through three series of experiments focusing on investigating different scenarios of comparison. In the first series, we compare features extracted without applying skin filter, called global evidence in this work. In the second series, features extracted after skin filter are also compared. These features are called local evidence. Finally, in the third series of experiments, a zoning algorithm is used in order to allow us to analyze the impact of both local and global features in each area of the image. In all series of experiments, each feature is analyzed individually and all subsets of features are tested so as to determinate the best tradeoff between feature set and classification accuracy. In addition, an architecture called ANDImage (Architecture for Nude Detection in Image) is proposed. ANDImage allows that different modules may be used in order to provide the possibility of dealing with different scenarios of features comparison.