混合采样与解耦协同的开放世界目标检测算法OA

Open world object detection(OWOD)is to balancing known category detection,unknown object recognition,and supporting incremental learning of new classes.To address issues low recall for unknown classes,insufficient feature decoupling,and poor incremental learning stability in existing methods,this paper proposes a hybrid sampling and decou-pling collaborative algorithm for OWOD(HSDC-OWOD).The method collaboratively optimizes through three modules:(1)mixed distribution proposal generation(MDPG)module integrates characteristics of Gaussian and long-tailed distribu-tions to generate region proposals with comprehensive coverage,enhancing the adaptability of object detection;(2)the multi-task self-attention decoupling module(MSDM)employs a dual branch self-attention mechanism to decouple fea-tures for category prediction and objectness prediction.It uses mutual information minimization to strengthen feature inde-pendence and maximum softmax probability verification to improve unknown category recognition robustness;(3)the multi-task self-calibration layer(MT-SCL)designs dual branch independent affine transformations to reconstruct task fea-ture spaces,mitigating catastrophic forgetting in incremental learning via a task routing mechanism.Experimental results show that on the M-OWODB dataset,algorithm increases the recall rate for unknown classes by 1.2-1.9 percentage points and improves the mAP by 1.1-1.5 percentage points compared with the best baseline.In the incremental learning task,the model achieves the mAP of 74.2%,which is 1.4-1.8 percentage points higher than existing methods.The results validate the algorithm's effectiveness in addressing low recall rate,insufficient feature decoupling,and unstable incremental learning.