Inappropriate scene detection in a video stream

Abstract

In this paper, we attempted to propose a model to detect inappropriate scenes, such as nudity, weapons, danger, drugs, gore, etc. in any video stream. The detection part is divided into different steps. The very first step of the proposed model is to convert the video stream into its frames. After fragmentation is completed, the image detection algorithms are used on the individual extracted frames to detect our required inappropriate scenes. For the detection part, we decided to fuse three different algorithms to detect the required inappropriate scenes. The nude detection part is handled using two algorithms. The first algorithm would find human figures on the fragmented frames and if found, the human figures would be cropped out and a separate image file would be formed containing the cropped-out part only. Next, the second algorithm would use this cropped image to find the presence of nudity in the original uncropped image. To detect dangerous objects like knives, guns, swords and detect gore, bloody scenes in the fragmented frames, the object detection algorithm is used. For object detection, we used CNNs Object Detection [1,2] algorithm to detect objects and scenes, and for nudity detection, we have used nudepy library from python which is based on detecting skin-colored pixels and identify nudity based on pixel count and its region [18]. After successful detection via the two algorithms, the model is going to give an output to show the percentage of how much of violent scenes and nudity is present in the video sequence. Furthermore, the model will also be able to determine and classify the video as pornography if the percentage nudity detected is over our base scale of what percentage of nudity in a video may be considered as a pornography.