Zero-shot Hand-held Objects Recognition Based on Global Feature Relationships: Zero-shot Hand-held Objects Recognition Based on Global Feature Relationships.

In recent years, there have been frequent terrorist attacks both at home and abroad. Accurate detection of a wide variety of handheld objects is a problem that needs an urgent solution. To better exploit the effects of handheld movements on handheld object recognition, this paper proposes a zero-shot hand-held object recognition based on global feature relations that comprise two modules: an image processing module and a semantic relations module. The image processing module learns the visual classifier of the input image to obtain the classification weights. The semantic relations module propagates structural knowledge through semantic learning through graph convolution operations on the knowledge graph and obtains the classification weights for all categories. Results indicate that this method significantly improves the recognition rate compared to traditional handheld object recognition methods.