DETECTION OF PORNOGRAPHY CONTENT IN VIDEO USING SUPERVISED MACHINE LEARNING METHOD WITH IMAGE BASED APPROACHES

• Main Author: Kurniawan, Budi
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39851
• Institution: Institut Teknologi Bandung

The growth of information and communication technology is rapidly increasing, and technology has changed how people's lives. One of the information and communication technology media that is often used every day is the internet. According to data presented by the Asosiasi Penyelenggara Jasa Internet Indonesia (APJII), in 2017 internet users in Indonesia have reached 54.68% of the total population of Indonesia. One concern that can arise from the use of the internet is the content of photos and videos that are commonly accessed. This concern is based on the activities that are often carried out by internet users to view photos or videos and according to the Kementrian Komunikasi dan Informatika (Kominfo), until the end of 2018, the most blocked negative content was pornographic content. One of the methods that can resolve the pornographic problem is image classification. This image classification method determines whether an image or frame contains pornographic content or not if there is pornographic content the image or the frame can be blurred. The definition of pornography is different for everyone, even a country has rules that are different from other countries related to pornography. In Indonesia, the definition of pornography is contained in Undang-Undang Nomor 44 Tahun 2008 tentang Pornografi. In the law pornographic content includes sexual intercourse, sexual violence, nudity, masturbation, genitals, and child pornography. This study focuses on nudity, masturbation, and genitals. The definition of nudity is also distinguished between men and women. Example: if in an image there is a woman wearing formal clothes, but showing her cleavage, then that image is a pornographic image. Convolutional neural network (CNN) is a method that is often used to carry out image classification processes. Many studies have implemented CNN to solve problems in the field of vision. However, CNN research on the pornography domain has not been done much. Therefore, this study uses the CNN method and is combined with other machine learning methods to provide the best accuracy. In this study, the CNN method was used in the feature extraction process. The CNN architecture implemented is VGG16, VGG19, ResNet50, InceptionResNetV2, and InceptionV3. The five architectures are architectures that perform goof performance in the world-class computer vision competition, namely ImageNet Large Scale Visual Recognition Challenge (ILSVRC). In the feature extraction process, the transfer learning process will also be carried out and the addition of one pooling layer to the last layer of the CNN architecture. Transfer learning process using pre-trained models from ImageNet. The use of transfer learning and the addition of one pooling layer aims to reduce the computational process, save resources, and the training process can run faster. The features produced by the CNN method are then used as input to be trained on three classification methods, namely multilayer perceptron (MLP), random forest, and support vector machine (SVM). After getting the CNN architecture and classifier method that produces the best accuracy, the fine-tuning process and the k-fold cross validation will be carried out against the hyperparameters of each classification method used to further improve its accuracy. The best model produced gives an accuracy of 97.06% in 1,600 images from the test dataset. The best results from the pornographic image classification method will be implemented in vide. If the frame is classified as a pornographic image, the frame will be blurred. In the implementation of video multithreading method is applied, this multithreading aims to be able to streamline the execution time between the image classification process and the process to display the frame to the user without buffering.