【摘要】 针对基于知识蒸馏的工业图像异常检测中，预训练网络在迁移到工业图像领域时因领域差异和数据分布的不同而导致获取的特征存在偏差问题，提出了一种特征自适应师生模型。该模型利用特征自适应器来调整教师网络提取的预训练特征，实现跨域特征转换和减少领域偏差。为了避免特征自适应器过度调整预训练特征，导致与原始预训练特征差异过大而降低泛化性能，提出监督损失来约束调整后的特征。此外，进一步为了提高模型对异常特征的表征和判别能力，设计了一个新的学生网络和提出了一种对抗损失来拉远教师和学生网络中异常特征之间的差异，拉近两者正常特征之间的差异。在多个工业数据集MVTec AD、BTAD、VisA和MVTec 3D AD上验证了该模型的有效性。更多还原

【Abstract】 In the industrial image anomaly detection based on knowledge distillation, the features obtained by the pre-trained network are biased due to domain differences and data distribution when transferred to the industrial image field. To solve this problem, a feature-adaptive teacher-student model was proposed. The pre-trained features extracted by the teacher network were adjusted utilizing a feature adapter to achieve cross-domain feature transformation and reduce domain bias. To prevent the feature adapter from excessively adjusting the pre-trained features, which could lead to significant deviation from the original features and degrade generalization performance, a supervised loss was proposed to constrain the adjusted features. Furthermore, to enhance the model’s ability to represent and discriminate anomalous features, a new student network was designed, and an adversarial loss was introduced to increase the discrepancy between anomalous features in the teacher and student networks while decreasing the discrepancy between their normal features. Experimental results on multiple industrial datasets(MVTec AD, BTAD, VisA, and MVTec 3D AD) validate the effectiveness of the proposed model.更多还原