State space model and federated adaptive heterogeneous edge visual inspection method and system

Abstract

The application provides a state space model and a federal adaptive heterogeneous edge visual detection method and system, relates to the technical field of computer vision, and comprises the following steps: discretization reconstruction of a state space operator in a lightweight visual backbone network is performed through a zero-order hold method; streaming inference is performed on a standard video frame to obtain a classification probability distribution; a data distribution drift event is identified in combination with information entropy, a sliding window and a dynamic threshold; core parameters of the lightweight visual backbone network are frozen, and a trainable parameter subset is activated; local adaptive fine-tuning is performed on the trainable parameter subset to obtain a parameter increment; the parameter increment is stripped through memory comparison, and a preset privacy protection mechanism is used to disturb the parameter increment to obtain a disturbed parameter increment; cloud global model aggregation is performed to obtain global model parameters; the global model parameters are distributed to edge nodes, and a hot update mechanism is used to perform inference operator hot loading to realize adaptive heterogeneous edge visual detection.

Images