Zero sample target identification method driven by large language model

Abstract

The invention belongs to the technical field of target recognition, and particularly relates to a large language model-driven zero sample target recognition method, which comprises the following steps of: 1, constructing a target language intelligence knowledge base; step 2; constructing a component feature recognition expert model library; step 3, performing feature recognition on the target component to be recognized; step 4, performing man-machine interaction correction on the component characteristics; and step 5, target category combination reasoning identification is carried out. According to the method, a to-be-recognized target is abstracted into a plurality of stable and discriminative component feature combinations, the semantic reasoning ability of a large language model is introduced on the basis, and a zero sample recognition task is reconstructed into a three-stage task of component feature recognition, component feature correction and combined reasoning recognition. On the premise of completely lacking target category image samples, effective recognition of unknown categories can be completed only by depending on corresponding language intelligence information.

Images