基于原型对齐的持续图像分割OA

Continual image segmentation suffers from catastrophic forgetting of previously learned classes,insufficient representation of newly introduced classes,and dynamic background semantic shifts in continual image segmentation.To address these issues,this paper proposes PAIS(Prototype-Aligned Incremental Segmentation),a prototype-aligned incremental segmentation method for continual image segmentation.From the perspective of object-level representation stability,PAIS jointly models the initialization,evolution,and historical semantic compensation of decoder object queries during continual learning.Specifically,prototype-aware initialization(PAI)selects semantically salient positions from the current feature map to initialize object queries,thereby improving the modeling capability for new classes.Temporal consistency learning(TCL)constrains the consistency of key semantic positions and their response distributions across adjacent learning stages,which alleviates objectness degradation caused by feature drift.In addition,category-aware memory(CAM)maintains compact class-level memory representations to enhance old-class retention without relying on large-scale image replay.Experiments on continual panoptic segmentation and continual semantic segmentation on the ADE20K dataset demonstrate that the proposed method consistently outperforms representative methods such as ECLIPSE,BalConpas,and CoMFormer under multiple task settings,data splits,and input sequences.Under the setting covering 150 classes,the proposed method achieves a clear improvement in PQ while reducing storage requirements by approximately 10 times.These results show that PAIS provides a favorable balance among new-class adaptation,old-class retention,and memory efficiency.