Self-Supervised Industrial Image Generation for Defect Recognition Under Limited Data

Deep learning-based industrial image inspection has achieved tremendous success in recent years, while the task of learning a model from a few defect images remains unexplored. The most popular method is to directly generate defect images to augment the limited dataset before training the recognitio...

Full description

Saved in:
Bibliographic Details
Main Author: Lee, Younkwan
Format: Conference Proceeding
Language:English
Subjects:
Accuracy
Benchmark testing
Image recognition
Image synthesis
Inspection
Surveillance
Training
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Deep learning-based industrial image inspection has achieved tremendous success in recent years, while the task of learning a model from a few defect images remains unexplored. The most popular method is to directly generate defect images to augment the limited dataset before training the recognition model. To address this issue, we propose a new framework called REG-Net that combines the generation module and classification module in an end-to-end manner, with few-shot defect image generation as assistance. Specifically, we design a two-branch module with attention fusion to directly combine normal and defect features. This reduces the negative impact on the classification module when the generation module performs poorly. REGNet improves further classification performance through recognition-friendly defect generation. Extensive experiments on the MVTec-AD benchmark show that effectiveness compared to the recent state-of-the-art methods.
ISSN:2643-6213
DOI:10.1109/AVSS61716.2024.10672565