A New Multi-Modal Federated Learning Framework for Heterogeneous Change Detection

Federated learning can shorten training time in heterogeneous change detection through its decentralized processing mechanism. However, images in different modalities can exhibit non-independent and identically distributed characteristics. Therefore, we propose a new aggregation algorithm for federated learning to accelerate the training of heterogeneous change detection models. In the proposed method, each client corresponding to a specific modality is equipped with a backbone responsible for local data feature extraction, while the server is deployed over the whole heterogeneous change detection network. The client training parameters are gradually updated by first projecting the detection results and then aggregating them through averaging at the server. Our method focuses on shortening training time instead of privacy concern, which is not a typical federated learning algorithm. Experiments on representative datasets verify the efficiency of the proposed method, which requires substantially fewer training iterations to reach convergence and reduces communication significantly compared to current methods.