Composite material surface microscopic defect detection and classification combining diffusion models and zero-shot learning.

This research aims to address the technical challenges of data scarcity and the identification of unseen defect types in the detection of microscopic defects on the surface of composite materials. An innovative detection framework integrating diffusion model and zero-shot learning has been constructed, adopting a dual-path collaborative architecture design. In this framework, the diffusion path learns the potential distribution characteristics of defects through the conditional diffusion process, and the zero-shot path realizes cross-modal knowledge transfer through the vision-semantic joint embedding space. The framework integrates an improved denoising diffusion probability model, a zero-shot classifier based on attribute description, and a multi-level feature fusion mechanism, achieving collaborative optimization among modules through a joint training strategy. Evaluations on multiple benchmark datasets verified the effectiveness of the method. The zero-shot learning mechanism achieved detection accuracy comparable to that of traditional supervised learning without the need for labeled samples, significantly reducing labeling costs. At the same time, it demonstrated excellent generalization ability across different types of composite materials. This framework provides an efficient technical solution for intelligent quality control in the composite materials manufacturing industry, promoting the development of industrial defect detection technology towards intelligence.