Referring expression generation from images via deep learning object extraction and fuzzy graphs

We propose an approach to obtain referring expressions for objects in images. Segmentation and categorization is performed via deep learning, providing the basis for a graph-based representation of the context that is later enriched with fuzzy properties. Referable objects and their corresponding expressions are obtained from the graph using extraction algorithms following a classical REG approach. The proposal is more flexible and scalable than end-to-end deep learning techniques.