徐雪松 付瑜彬 于波

摘要：【目的】为解决图像分类模型面对传统闭集训练方式出现的模型缺乏开集泛化性的问题，提出了一种分离式的独立分类网络结构。【方法】每个类别都包含独立的线性特征层，特征层中设计的神经元节点能够在有限的数据样本下更准确地捕获类别特征。同时，在模型训练时，文中引入了一类无需标注的负样本，使得模型在构建决策边界时不仅依赖于已知类别的特征差异，在不增加额外标注样本的情况下，增加模型决策边界的开集泛化性。【结果】结果表明：独立分类网络开集识别（ICOR）模型结构和开集自适应训练策略均能有效改善传统模型开放集识别（OSR）性能；随着开放度的增加，能表现出更好的鲁棒性，能更有效地降低模型的OSR风险。【结论】提出的独立分类网络并融合开集自适应训练的算法比现有开集识别算法具有更优的开集识别性能。

关键词：深度学习；开集识别；图像分类；迁移学习

中图分类号：TP391；U495 文献标志码：A

文章编号：1005-0523（2024）02-0079-08

Research on Open Set Recognition Based on Independent Classification Network

Xu Xuesong， Fu Yubin， Yu Bo

（School of Electrical & Automation Engineering， East China Jiaotong University， Nanchang 330013， China）

Abstract： 【Purpose】In order to solve the problem of image classification models lacking open set generalization due to traditional closed set training methods when facing open set recognition problems， we propose a separate independent classification network structure. 【Method】 Each category contains an independent linear feature layer. The neural nodes designed in the feature layer can capture the category features more accurately under limited data samples. At the same time， a class of negative samples without labeling is introduced in the model training， so that the model not only relies on the feature difference of the known categories when constructing the decision boundary， but also increases the open set generalization of the model decision boundary without adding additional labeled samples. 【Result】The results show that both the ICOR model structure and the open-set adaptive training strategy can effectively improve the OSR performance of traditional models; with the increase of openness， it can demonstrate better robustness; can more effectively reduce the OSR risk of the model. 【Conclusion】The proposed independent classification network combined with open-set adaptive training algorithm has better open-set recognition performance than existing open-set recognition algorithms.

Key words： deep learning; open set recognition; image classification; transfer learning

Citation format： XU X S，FU Y B，YU B. Research on open set recognition based on independent classification network[J]. Journal of East China Jiaotong University， 2024， 41（2）： 79-86.

【研究意义】在实际应用中，由于无法准确预知样本的分布，预先设计好的分类器在面对未知类别样本时难以保持较高的分类准确性，这就是开放集识别问题（open set recognition，OSR）[1]。近年来许多学者针对这个问题提出不少方法，总的来说，分成生成式和判别式两种[2]。

【研究进展】判别式方法通过正则化收缩已知类在特征空间的决策边界，并在特征空间内给未知类分配区域，从而缓解模型的过度泛化，降低模型的OSR风险[3-5]。……