Enhancing Few-Shot marble slab surface defect detection: A diffusion framework with knowledge distillation and semantic guidance

doi:10.1016/j.engappai.2026.114334

Enhancing Few-Shot marble slab surface defect detection: A diffusion framework with knowledge distillation and semantic guidance

2026-02-28

0

PRE

AI

DOI：10.1016/j.engappai.2026.114334

Original

Original request for help

Save

Abstract

En 中文

Surface defect detection on marble slabs is of paramount importance in industrial quality assurance, directly impacting production efficiency and economic sustainability in the stone processing industry. However, conventional Few-Shot Anomaly Detection (FSAD) techniques often suffer from limited reconstruction quality and suboptimal segmentation performance, hindering their applicability to marble surface inspection tasks. Compounding this issue is the lack of comprehensive, high-resolution datasets tailored to marble slab defects, which presents a significant barrier to research progress in this area. To overcome these limitations, we introduce a Mask-guided and Enhanced Semantic-guided Diffusion-based framework (MESD), which exploits multiple representational domains and targeted guidance strategies for robust and accurate defect detection. Specifically, MESD unifies pixel-space, feature-space, and latent-space representations, augmented by two key components: the Mask-Guided Knowledge Distillation network (MGKD), which emphasizes defective regions to refine reconstruction quality, and the Semantic-Guided Enhancement Network (SGEN), which aids in preserving semantic coherence during the reverse diffusion process. In conjunction with the framework, we present the Marble Slab Surface Defect Dataset (MSSDD), the first few-shot dataset explicitly curated for marble defect detection and segmentation. Extensive evaluations demonstrate that MESD achieves compelling performance on MSSDD, as well as on general anomaly detection benchmarks. MESD shows strong capabilities in segmentation tasks, highlighting its promise for practical deployment in industrial inspection pipelines.

Keywords:

Few-Shot Anomaly Detection

Marble Surface Defect Detection

Diffusion Models

Knowledge Distillation

Semantic Guidance

Enhancing Few-Shot marble slab surface defect detection: A diffusion framework with knowledge distillation and semantic guidance

Abstract

Journal

Organization