FusionDiff: a dual-path diffusion-based framework for few-shot authenticity analysis of ceramic microstructures

Abstract The authenticity of ceramic components is closely tied to their microscopic structures, making automatic and accurate identification essential for quality control. However, this task is often constrained by the scarcity of labeled samples. This study investigates the potential of large-scale pretrained diffusion models as feature extractors, leveraging the rich visual priors embedded in their generative processes to provide a robust semantic foundation for small-sample learning. To address the limitations of the original U-Net in global representation modeling and the weak local-detail sensitivity of DeiT, we propose a dual-path fusion encoder, FusionDiff. Within a frozen Stable Diffusion V1.4 framework, CNN and adapter-enhanced DeiT paths operate in parallel and are deeply integrated via feature gating. Following a “self-supervised pretraining + supervised fine-tuning” paradigm, classification is performed using a Random Forest classifier. On our custom ceramic dataset, FusionDiff achieves a test accuracy of 99.07%, outperforming SD-CNN (97.44%), DeiT (96.30%), and ResNet50 (97.00%) under a unified self-supervised evaluation protocol. Even under extremely small-sample conditions ( $$n = 50$$ ), the model attains 90.7% validation accuracy, demonstrating competitive data efficiency and cross-domain generalization capability.