Knowledge distillation has made significant progress in anomaly detection. However, it is difficult for student network to distinguish the features between similar normal and abnormal samples due to similarities in structure or data processing with teacher network. To this end, this paper proposes a Cognitive Dissonance Reverse Distillation (CDRD) to optimize the training process of student network through the double-distillation learning mechanism. CDRD draws on the theory of cognitive dissonance in psychology. The student network will undergo a series of adjustment and optimization processes to reduce internal inconsistencies or contradictions and improve the model’s understanding of the differences between normal and abnormal features. Extensive experiments conducted on benchmark datasets demonstrate that the proposed model achieves detection accuracy of 99.59%/98.65%/95.70% on MVTec AD, 84.1% on MVTec Loco in pixel-level segmentation. Specifically, this proposed model outperforms state-of-the-art method RD++ by 0.15%/0.4%/0.71%, which underscores the potential for practical applications in anomaly detection task.