Weeds are a significant threat to the agricultural productivity and the environment. The increasing demand for sustainable agriculture has driven innovations in accurate weed control technologies aimed at reducing the reliance on herbicides. With the great success of deep learning in various vision tasks, many promising image-based weed detection algorithms have been developed. This paper reviews recent developments of deep learning techniques in the field of image-based weed detection. The review begins with an introduction to the fundamentals of deep learning related to weed detection. Next, recent progresses on deep weed detection are reviewed with the discussion of the research materials including public weed datasets. Finally, the challenges of developing practically deployable weed detection methods are summarized, together with the discussions of the opportunities for future research.We hope that this review will provide a timely survey of the field and attract more researchers to address this inter-disciplinary research problem.