Diffusion Model-based Method for Generating Aligned Anomaly Images and Masks
- 주제(키워드) Anomaly Inspection , Anomaly Detection , Image Generation , Diffusion , Anomaly Generation
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 Yong Seok Heo
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035482
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
In industrial manufacturing, anomaly inspection performance is frequently hampered by a lack of anomaly data. To address this issue, models have been developed that use anomaly generation techniques to create synthetic anomaly images and corresponding masks. However, these methods face significant chal- lenges, such as the frequent misalignment of generated anomaly images with their corresponding masks and a lack of realism, which lead to performance degradation in downstream tasks. To tackle this challenge, we introduce An- odapter [1], a unified few-shot anomaly generation model that uses a single diffusion process to sequentially generate anomaly masks and their correspond- ing images. Anodapter employs a cohesive strategy where the anomaly masks guide the generation of images, ensuring precise alignment and high-quality results. This model efficiently separates anomaly information into appearance and spatial components. For spatial control, we introduce a Switch Adapter that manages the spatial arrangements of anomalies. To control appearance, the model employs specialized prompts with unique identifiers, enabling selec- tive generation of anomaly images or masks. Through extensive experiments using the MVTec dataset, we demonstrate that our model can generate realis- tic and diverse anomaly datasets, significantly outperforming existing methods both quantitatively and qualitatively in downstream tasks such as anomaly de- tection, localization, and classification.
more목차
Abstract
1 Introduction 1
2 Related Works 8
3 Proposed Method 12
3.1 Preliminary 12
3.1.1 Latent Diffusion Model 12
3.1.2 Dreambooth 13
3.2 Proposed Method 14
3.2.1 Switch Adapter 15
3.2.2 Generation Control Prompt 18
3.2.3 Training objective function 21
4 Experiments and Results 24
4.1 Experiments 24
4.1.1 Experimental settings 24
4.1.2 Anomaly generation for anomaly detection and localization 27
4.1.3 Anomaly generation for anomaly classification 30
4.1.4 Comparison in Anomaly Generation 30
4.1.5 Comparison in Computational Complexity 39
4.1.6 User Study 41
4.1.7 Ablation study 44
5 Conclusions 48
Reference 50
