A Hybrid Deep Learning Approach for Explicit Content Detection in Images on Social Media and Internet

Abstract

The exponential increase in the amount of explicit content has presented numerous obstacles to the current way of life. This is especially true in situations wherein children and minors have unrestricted access to the internet. This process of screening the image features of all the TV channels in Malaysia imposes a huge censorship cost on the service providers like Unifi TV because all films, both local and foreign, are required to obtain the suitability approval in Malaysia before they can be distributed or shown to the public. This paper proposes the use of a hybrid model of Deep Learning (DL) approaches, specifically CNN+SVM and CNN+XGBOOST, in order to further improve the process of explicit images recognition in visual contents. The goal of this paper is to use this issue to our advantage. Transfer learning was performed using the previously trained model in order to find a solution to a new binary classification problem including explicit and non-explicit images. The effectiveness of the model that has been developed is examined using a dataset that has recently been compiled and contains more than number of examples of explicit non-explicit images photographs. The CNN+XGBOOST approach was able to acquire the best performance in terms of accuracy of 99.81 % after tests were run on the dataset. This was in comparison to the 76.49% accuracy attained by the CNN+SVM model. For better evaluation we also present the comparison of the proposed systems with standard state of art mechanisms viz. NB, LDA, SVM, RF, KNN, DT, and LR.