Abstract: Open-World Object Detection (OWOD) extends the object detection task to real and variable environments, requiring models to accurately identify known and unknown objects, and gradually learn new knowledge. In response to the low recall rate of unknown classes and the problem of false identification in existing open-world object detection methods, Gradient-discriminative and feature norm-driven open-world object detection method (GDFN-OWOD) is proposed. To address the issue of low recall rate for unknown classes, a Gradient-Discriminative Representation Module (GDRM) is proposed, which uses the gradient difference from backpropagation to accurately distinguish unknown classes from the background, thereby improving the recall rate of unknown classes. In addition, a Graph-Based Segmentation for Bounding Box Clustering (GSBC) algorithm is introduced, modeling the determination of object bounding boxes as a graph decomposition problem, reducing redundant bounding boxes, and thus reducing the computational load of the model. To tackle the problem of false identification of unknown classes, a FeatureNorm-Based Classifier (FN-BC) is employed, selecting the most optimal convolutional layer to recognize known and unknown classes for higher recognition accuracy. Experimental results on the OWOD dataset show that the recall rate of unknown classes in tasks T1, T2, and T3 has been increased by 1.1%, 2.1%, and 0.9%, respectively, and the absolute open-set error (A-OSE) has been reduced by 35.1%, 28.7%, and 12.2%, respectively. Compared with existing open-world object detection methods, it effectively alleviates the problems of low recall rate for unknown classes and false identification.

Key words: open-world object detection, backpropagated gradient, Graph segmentation algorithm, feature norm, convolutional neural network

摘要： 开放世界目标检测(Open-World Object Detection,OWOD)将目标检测任务拓展至真实多变的环境中，要求模型能够准确识别已知和未知对象，并逐步学习新知识。针对现有开放世界目标检测方法中未知类的召回率偏低和误识别的问题，提出了一种梯度区分与特征范数驱动的开放世界目标检测研究方法(Gradient-discriminative and feature norm-driven open-world object detection,GDFN-OWOD)。针对未知类召回率偏低的问题，提出梯度区分性表征模块(Gradient-Discriminative Representation Module,GDRM)，利用反向传播的梯度差异准确区分未知类别和背景，以提高未知类召回率；此外，引入基于图分割的框聚类算法(Graph-Based Segmentation for Bounding Box Clustering,GSBC)，将物体边界框的确定建模为图分解问题，减少冗余的边界框，进而减少模型的计算量；针对未知类误识别问题，采用基于特征范数的分类器(FeatureNorm-Based Classifier,FN-BC)，选择性能最优的卷积层识别已知和未知类别，以达到更高的识别准确率。在 OWOD 数据集上的实验结果表明，在T1、T2、T3任务中未知类召回率分别提升了1.1、2.1、0.9个百分点，绝对开集误差(A-OSE)分别降低了35.1%,28.7%,12.2%。与现有的开放世界目标检测方法相比，有效缓解了未知类的召回率偏低和误识别的问题。

关键词: 开放世界目标检测, 反向传播梯度, 图分割算法, 特征范数, 卷积神经网络