The German National Library of Science and Technology (TIB) researches and develops methods of automated content analysis and semantic web technologies to improve access to its library holdings and allow for advanced methods of information retrieval (e.g. semantic and cross-lingual search). Regarding scientific videos in the TIB AV-Portal spatio-temporal metadata is extracted by several algorithms analysing (1) superimposed text, (2) speech, and (3) visual content. In addition, the results are mapped against common authority files and knowledge bases via a process of automated Named Entity Linking and published as Linked Open Data to facilitate reuse and interlinking of information. Against this background the TIB constantly aims to improve its automated content analysis and Linked Open Data quality. Currently, extensive research in the fields of deep learning is conducted to significantly enhance methods of visual concept detection in the AV-Portal – both in terms of detection rates and coverage of subject-specific concepts. Our solution applies a state-of-the-art deep residual learning network based on the popular TensorFlow framework in order to predict and link visual concepts in audio-visual media. The resulting predictions are mapped against authority files and expressed as RDF-Triples. Therefore, in our presentation we would like to demonstrate how research in the field of machine learning can be combined with semantic web technologies and transferred to library services like the AV-Portal to improve functionality and provide added value for users. In addition we would like to address the question of data quality assessment and present scenarios of metadata reuse.