A Semantic Segmentation Approach to Recognize Assault Rifles in ISIS Propaganda Images

Between 2014 and 2018, the Islamic State in Iraq and Syria (ISIS) perfected the use of social media for its propaganda. To understand and counter these efforts by ISIS, it is critical to analyze their propaganda materials. During the past few years, a systematic effort has been made to catalog and annotate these materials which appear in the form of images, video and text. However due to the sheer volume of the material, it is an extremely onerous task to maintain. In this work, we present a deep learning solution to automatically identify and tag images for assault rifles. We present our experiments of a semantic segmentation approach to localization of assault rifles in a self-collected and maintained data set. Our goal is to consume minimal amount of data and cater to an analysis platform. The state of the art for object localization is the Convolutional Neural Network (CNN). A limitation of CNN is that it only handles images of fixed dimensions. One way to deal with this limitation is to re-size the input images, however this is not an ideal solution. A more flexible approach is to use a Fully Convolutional Network (FCN), which provides a robust solution for varied sizes of input images. We show that FCNs can achieve high performance in detecting and localizing objects in a real world setting, with non-curated data. We also show that by using a step wise training pipeline it is possible to learn a representation of the object using a bounding box annotation.