Abstract:In real open-world scenarios, object detection models inevitably encounter unknown categories of objects that are not annotated or seen during training. It is hard for traditional detection methods based on the closed-world assumption to deal with this problem effectively. To address challenges such as large scale variations of unknown objects, insufficient discriminative capability of candidate regions, and high false positive rates under complex scenarios, this study proposes an improved method for unknown object detection based on the Faster R-CNN detection framework. Firstly, the parallel dynamic multi-scale perception module u-PMS is introduced into the backbone network. By introducing an input-dependent dynamic gating mechanism to adaptively fuse multi-scale convolutional features, the model’s feature representation and perception for unknown objects of varying sizes are enhanced. Secondly, an objectness prediction head (OPH) independent of class labels is constructed. It utilizes CIoU-based geometric supervision to explicitly learn generic objectness features, effectively distinguishing real objects from background noise in the absence of semantic labels. Finally, a multi-signal post-processing filtering module (MSFM) is proposed to progressively filter candidate boxes by integrating multi-source prediction information, further suppressing redundant predictions and pseudo objects. Experimental results demonstrate that the proposed method outperforms baseline models on the COCO-OOD and COCO-Mix datasets, especially yielding superior results on the unknown object detection quality metric U-AP, which validates the effectiveness and robustness of the proposed method in complex open scenarios.