Video surveillance cameras are the backbone for containment and surveillance activities in international safeguards. Safeguards inspectors need to review the images of hundreds of surveillance cameras installed in nuclear facilities worldwide. Besides being very resource-intensive, there is a growing number of facilities that require (near) real-time analysis of safeguards data, including surveillance imagery, which can be difficult to achieve with the current workflows. Over the last decade, deep learning has revolutionized automated image analysis in a wide variety of applications, reaching from face recognition to autonomous driving. It has the potential to improve significantly the efficiency and effectiveness of surveillance image review in nuclear safeguards, but also poses a number of challenges, such as availability of labelled training data. The paper proposes a workflow for training deep neural networks for detecting application-specific objects of interest by combining pre-trained models, synthetic data and only a small number of real images. Results of initial experiments in a lab environment show the potential of the approach for nuclear safeguards applications.