Many Internet users are potential victims of pornographic images, most of which are minor children. The ability to filter content for the content of pornographic images from multimedia sources has an important use. Most of the work related to the detection of pornography is based on the detection of human skin. The disadvantage of this approach is associated with obtaining a high level of negative results that are observed in sports and beach filming. To detect pornographic frames in still images, use the Bag of Visual Features (BoVF), which is a good method for recognizing naked people even at the frame level. According to the results of the experiment, the model is able to detect adult content with an accuracy of 75.08% in the validation process and 69.02% in the testing process. To achieve a positive result in the detection of pornographic content, the new ACORDE deep learning architecture is also proposed, which includes both convolutional neural networks and recurrent networks Long ShortTerm Memory (LSTM) for detecting adult content in video. A deep neural network is used to build a model that is able to automatically detect adult content.
With the rapid development of the Internet and social networks, there are a large number of pornographic images available to users. Most parents want to protect their underage children from accessing websites that contain pornographic multimedia. In addition, in educational institutions and the workplace, basic ethical norms and standards of behavior dictate that such images become inaccessible to the community. There are many different methods for solving this problem: creating a database of blacklisted websites. As the number of sites increases, it is not practical to manually blacklist all pornographic sites.
Automatic detection of adult (pornographic) content in images and videos is an important and complex task. The task of automatically identifying adult content is more difficult because of the degree of subjectivity and uncertainty. For example, even people find it difficult to correctly assess the degree of sensuality in scenes where people wear swimsuits or underwear. Obviously, the solution is a form of intelligent machine that can analyze text, audio, or visual signals.
Bag-of-Visual-Features (BoVF) approaches are used to detect naked bodies in still images, starting with the application of skin detectors and then applying geometric models of body posture to the detected areas of skin in order to search for a naked body area. BoVF approaches typically use local descriptors based on gray level values. However, color is fundamental information for detecting nudity. The HueSIFT-based bovf approach is applied to classifying still images between Nude and non-Nude images. Main problems associated with these approaches:
- accurate skin detectors themselves are far from trivial construction;
- the huge variability of body positions in such images makes it difficult to determine the overall geometric model.
Deep neural networks are very popular in the field of image recognition, especially in convolutional neural networks (CNN). The neural network layer is replaced by a set of image pixel parameters that can be trained through the neural network. The architecture of the ACORDE method (adult content Recognition with deep neural networks) uses a convolutional neural network (ConvNet) as a feature extractor and long-term short-term memory (LSTM) to perform final video classification. ACORDE extracts function vectors from the NPDI dataset, creating a sorted set of semantic descriptors. This set is used to feed the LSTM, which is responsible for analyzing video in an end-to-end format. The results show that ACORDE conveniently sets a new state for adult content detection in NPDI, reducing the number of false positives by half and the number of false negatives by a third. This model can be integrated into a mobile phone, tablet, or personal computer.