A large number of digital humanities projects focuses on text. This medial limitation may be attributed to the abundance of well-established quantitative methods applicable to text. Cultural Studies, however, analyse cultural expressions in a broad sense, including different non-textual media, physical artefacts, and performative actions. It is, to a certain extent, possible to transcribe these multi-medial phenomena in textual form; however, this transcription is difficult to automate and some information may be lost. Thus, quantitative approaches which directly access media-specific information are a desideratum for Cultural Studies. Visual media constitute a significant part of cultural production. In our paper, we propose Deep Watching as a way to analyze visual media (films, photographs, and video clips) using cutting-edge machine learning and computer vision algorithms. Unlike previous approaches, which were based on generic information such as frame differences (Howanitz 2015), color distribution (Burghardt/Wolff 2016) or used manual annotation altogether (Dunst/Hartel 2016), Deep Watching allows to automatically identify visual information (symbols, objects, persons, body language, visual configuration of the scene) in large image and video corpora. To a certain extent, Tilton and Arnold’s Distant-Viewing Toolkit uses a comparable approach (Tilton/Arnold 2018). However, by means of our customized training of state-of-the-art convolutional neural networks for object detection and face recognition we can, in comparison to this toolkit, automatically extract more information about individual frames and their contexts.