An Adaptive Porn Video Detection Based on Consecutive Frames Using Deep Learning


          

刊名：Revue d'Intelligence Artificielle
作者：Mohammad Reza Mazinani（Faculty of Electrical and Computer Engineering, Malek Ashtar University of Technology）
Kourosh Dadashtabar Ahmadi（Faculty of Electrical and Computer Engineering, Malek Ashtar University of Technology）
刊号：737F0004
ISSN：0992-499X
出版年：2021
年卷期：2021, vol.35, no.4
页码：281-290
总页数：10
分类号：TP3
关键词：Porn frame detection；Adult video recognition；Real-time video processing；Convolutional neural network；Adaptive classification；Computer vision
参考中译：
语种：eng
文摘：Many videos uploaded to online video platforms contain adult content that violates these platforms' policies and should be removed immediately. To recognize obscene videos, we developed a model that can process video frames in real-time while also adapting to time budget or hardware processing capacity. Thus, a deep convolutional neural network with multiple outputs was used. A decision-maker module was then designed to decide which neural network outputs to process and which label to assign to each frame. Using the reinforcement learning method, the decision-maker module is trained based on the results of previous frames as well as the results of neural network outputs while keeping the time budget in mind. Experiments showed that sacrificing a small amount of accuracy can increase speed by up to 4.7 times over the base model. We conclude that using a content correlation between consecutive frames not only reduces processing time by eliminating unnecessary frame processing but also improves the accuracy of the frame classification. It was also discovered that while using more of the previous frames, increases processing speed, the error in classifying the frame increases when the scene is changed.



©2016机械工业出版社（机械工业信息研究院） 京ICP备05055788号-35