非独立同分布数据流下的持续学习语义分割方法OA

In order to alleviate the catastrophic forgetting phenomenon that occurs when the semantic segmentation model is in-crementally updated with knowledge,a continual learning semantic segmentation approach under non-independent homogeneously distributed data streams is proposed.Firstly,generative adversarial network generation and web crawling are used as data sources for the model,and old data are replayed during training to alleviate catastrophic forgetting.Secondly,to further optimise the knowl-edge recovery effect after replay,gating variables are introduced into the network,and a gating mechanism is built into the network to further improve model stability and plasticity.Experiments on the Pascal VOC 2012 dataset show that in the most complex incre-mental scenario 10-1,the mIoU of the initial class set and all classes improve by up to 2.5%and 2.2%compared to the baseline.