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

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Ceramic component authenticity relies on microscopic structures, necessitating automated identification for quality control.
• Small labeled datasets pose a significant challenge for accurate classification tasks.
• Pretrained diffusion models offer rich visual priors beneficial for small-sample learning.

Purpose of the Study:

• To investigate the efficacy of large-scale pretrained diffusion models as feature extractors for ceramic identification.
• To develop an improved feature extraction method addressing limitations in existing models for small-sample learning.
• To enhance the accuracy and data efficiency of ceramic authenticity identification.

Main Methods:

• Proposed a dual-path fusion encoder, FusionDiff, integrating CNN and adapter-enhanced DeiT paths within a frozen Stable Diffusion V1.4 framework.
• Employed a "self-supervised pretraining + supervised fine-tuning" paradigm for classification.
• Utilized a Random Forest classifier on extracted features for final classification.

Main Results:

• FusionDiff achieved a test accuracy of 99.07% on a custom ceramic dataset, outperforming SD-CNN, DeiT, and ResNet50.
• Demonstrated strong performance under extremely small-sample conditions, reaching 90.7% validation accuracy.
• Showcased competitive data efficiency and cross-domain generalization capabilities.

Conclusions:

• Pretrained diffusion models serve as powerful feature extractors for small-sample learning in materials science.
• The proposed FusionDiff encoder effectively captures both global and local features for robust ceramic identification.
• FusionDiff offers a promising solution for accurate and data-efficient quality control in ceramic manufacturing.