Comparison of Pornographic Image Classification based on Texture, Color, and Shape Features

Content filtering application on the internet is very important to protect children from various negative content such as pornography and violence. It is necessary to find a detection tool for pornographic images with high accuracy and fairly low computational time. The classification algorithm is responsible for distinguishing pornographic images and non-pornographic images. This research aims to find the optimum classifier among available algorithms. To achieve these objectives, we compare the existing classification algorithm namely: K Nearest Neighbors (KNN), Logistic Regression (LR), Linear Discriminant Analysis (LDA), Decision Tree (DT), Random Forest (RF), Multi Layers Perceptron (MLP), and Support Vector Machine (SVM). Classification algorithm needs features as the input. Therefore we need a feature extraction algorithm to do so. In this study, we evaluate some features which are Haralick texture (texture features), Color Histogram (color features), and Hu Moment (shape features). The test results show that the most suitable method for classification of pornographic images is the Random Forest (RF) with an accuracy of 91.04% and the computational time per image is 0.80 ms.