Sensitive media analysis through deep learning architectures

Abstract

With the growth of video data generated by many devices, such as cameras, smartphones, and CCTVs, allied with the Internet as a fast spreading venue, smart and continuous content filtering becomes paramount. In this context, classification of sensitive media content (e.g., pornography, violence) retains a considerable amount of attention because of its applications: it can be used for detecting, via surveillance cameras, inappropriate behavior; blocking undesired content from being uploaded to general purpose websites (e.g., social networks, online learning platforms, forums), or from being viewed in some places (e.g., schools, workplaces); preventing children from accessing adult content on personal computers, smartphones or smart TVs; and avoiding that improper content is distributed over phones (e.g., sexting). In this project, we aim at designing, developing and deploying solutions for detecting sensitive content in digital images/videos such as pornography, violence and adult content. We will advance the state of the art by investigating deep networks architectures, by proposing new efficient and effective deep learning approaches, and by incorporating mobile solutions for sensitive media analysis. Specifically, we will focus our efforts on: (1) designing and developing novel deep learning approaches for automatically extracting discriminative space-temporal characteristics from videos; and (2) devising new approaches for dealing with sensitive content that incorporate the benefit of robust mid-level representations such as Fisher Vectors with the powerful Deep Learning characterization pipeline. We highlight that our group is at the forefront of such research worldwide being responsible for groundbreaking results associated with digital forensics, machine learning, and computer vision. (AU)