Visual depth model knowledge recombination method for intelligent security and protection

A visual depth model knowledge recombination method for intelligent security protection comprises the following steps: firstly, collecting visual data of security protection edge measurement, defining a model task, collecting the data for labeling, and training a target task model; then, calculating a transferable score between the target model and different pre-trained models by utilizing an attribution graph, taking the normalized transferable score as a probability of model selection, and sampling in a replaceable manner to form a model set; on the basis of a federated learning normal form, a target model is distributed to different nodes, after learning and migration of the target model are completed, knowledge recombination is carried out at a center node in an averaging mode by utilizing the feature extraction capacity of a pre-training model representing distillation migration, and the recombined model is distributed to a selected model node again for further iteration. And finally, performing fine adj