Abstract
Is it possible to create an AI-based system, capable of detecting harmful video content? This is the research question in focus within the project “Machine Learning for Interactive Age Classification of Films”.
The volumes of video content published and accessible online are growing, on Youtube alone 720 000 hours of moving pictures are uploaded every day. Some video material, ie violence, is considered harmful to minors and in Sweden it is the Swedish Media Council (SMC) who is tasked by the government to work to protect children from this type of content. In order to be able to fulfill this task, the SMC identified the need to automate the process and together with researchers at the Linköping University we designed a project and managed to attract funding from Vinnova.
The first stage of the project, ending in August 2021, has delivered good results. A high quality dataset of close to 4000 fully annotated video clips for training the system has been created. Further, an efficient and high-performing technical solution for classification has been developed along with task aware evaluation metrics.
The technical solution uses state of the art machine learning techniques to integrate both video and audio modalities. Early results show that close to human expert performance can be achieved on parts of the created dataset.
We see great opportunities in continuing this work – in view of creating a more mature technology aimed at eventually, autonomously, detecting harmful video content. This technology would solve both a public sector challenge: supervising constantly growing content volumes on the internet and thereby securing minors online safety. Moreover, the technology has a strong market potential, ie if implemented in parental control apps and similar content filters, a growth market globally.
For the next stage, we are looking for funding and partners in order to develop the technology so as to bring it closer to market and possibly even broadening the potential application areas.
Bio of the speakers
Anette Novak, Director General of the Swedish Media Council, former CEO of ICT and design research institute RISE Interactive and head of the Swedish government’s Media Inquiry. Anette Novak has a long career within the media industry, both in executive and strategic roles. Her last operative position being editor-in-chief of regional media house Norran. Some examples of board assignments: Swedish Media Publisher’s Association, Swedish public service radio Sveriges Radio, Schibsteds Tinius Trust and Sweden’s representative to the World Editors Forum.
Michael Felsberg, Professor in Computer Vision at the Department of Electrical Engineering, Linköping University. His research focuses on machine learning for robot vision and machine perception. Michael has published more than 200 papers, gathering more than 18000 citations (H=44), and he is supervising 8 PhD students, most of them in the WASP graduate school. He is also vice-head of department and LiU representative in the WASP executive committee