针对开放世界目标检测中未知类目标预测性能不佳的问题，提出了一种基于形状感知与类平衡优化的开放世界目标检测方法。未知类指在训练阶段未标注的类别，由于缺乏标签的指导，未知类目标的检测是一个具有挑战性的任务。构建了一种未知类增强探测器，作为未知类检测分支，在训练阶段只利用已知类标签进行监督，让探测器学习已知类目标特征的相似性，进而推广到未知类目标。为了提高探测器对未知类的敏感度，利用区域生成网络（region proposal network, RPN）模块区分前景和背景的特性，使用特定筛选方式，从RPN输出中选择“具有未知类潜力”的结果作为伪标签参与探测器训练过程。由于缺乏置信度得分，传统非极大值抑制（non-maximum suppression, NMS）方法难以应用于未知目标的后处理，因此设计了一种冗余未知类目标框抑制器，该抑制器由基于中心点的分组策略和基于形状感知冗余度得分矩阵构成。其中基于中心点的分组策略包含三种根据未知类中心点的分组方法，用于确定抑制范围。接着根据组内每一个预测框的冗余度得分构建冗余度得分矩阵，从而抑制高冗余预测结果。在开放世界目标检测数据集上的实验结果表明基于形状感知与类平衡优化的开放世界目标检测方法在保证未知类召回率的同时，具有较高的未知类预测精度。基于形状感知与类平衡优化的开放世界目标检测方法能有效应对开放世界的难题，避免产生大量的无用预测结果。

An open world object detection method based on shape perception and class balance optimization was proposed to address the issue of poor prediction performance of unknown class objects in open world object detection. Unknown classes referred to classes that were not labeled during the training phase. Due to the lack of guidance from labels, detecting unknown class objects was a challenging task. An unknown class enhanced detector has been constructed as an unknown class detection branch. During training, this detector was supervised using only known class labels, allowing it to learn the similarities in features of known class objects and generalize to unknown class objects. To improve the detector's sensitivity to unknown classes, the region proposal network (RPN) module's ability to distinguish between foreground and background was utilized. A specific filtering method was employed to select results with "unknown class potential" from the RPN output, which were then used as pseudo labels in the training process. Due to the absence of confidence scores, traditional non-maximum suppression (NMS) methods were difficult to apply for post-processing unknown objects. Therefore, a redundant unknown object suppression mechanism was designed, consisting of a center point-based grouping strategy and a redundancy score matrix based on shape perception. The center point-based grouping strategy included three methods based on the unknown class center points to determine the suppression range. Subsequently, a redundancy score matrix was constructed based on the redundancy scores of each prediction box within the group to suppress highly redundant predictions. Experimental results on open world object detection datasets demonstrated that the open world object detection based on shape perception and class balance optimization maintained high recall rates for unknown classes while achieving high prediction accuracy. This method effectively addressed the challenges of open world scenarios and avoided generating a large number of useless predictions.