- The paper introduces a novel framework combining OASA, SCL, and an innovative weight updating strategy to prevent catastrophic forgetting during incremental learning.
- It demonstrates superior adaptability and pixel-level detection accuracy on the MVTec-AD dataset compared to state-of-the-art methods.
- The framework promises enhanced efficiency and reduced costs in industrial manufacturing by accommodating continual changes in production.
Introduction
The field of AI-driven defect inspection, commonly referred to as anomaly detection, plays a crucial role in various industrial manufacturing processes. Automated inspections utilizing AI technologies offer not only exceptional accuracy but also a reduction in labor costs. Despite such advancements, the current systems face challenges like the need to detect multiple objects and adapt to continually changing production schedules. Traditional solutions, which often train a separate model per object, lack versatility and incur high costs due to frequent equipment updates. This paper introduces a framework called Incremental Unified Framework (IUF) designed to address these issues.
Approach and Methodology
IUF enhances the capacity for pixel-precise defect detection across diverse objects without relying on embedded memory banks for feature storage. The framework tackles a common problem in incremental learning called "catastrophic forgetting," where learning new objects diminishes the ability to recall previously learned ones. To address this, three main components are proposed: Object-Aware Self-Attention (OASA), Semantic Compression Loss (SCL), and a novel weight updating strategy.
OASA leverages self-attention mechanisms to create distinct semantic boundaries for objects, reducing feature overlap during incremental learning. SCL optimizes network adaptability by compressing non-primary semantic space, facilitating the integration of new objects into the learning model. The updating strategy ensures that important features of existing objects are retained during the learning of new objects, thereby minimizing interference.
Experimentation and Findings
The proposed framework was tested against the MVTec-AD dataset tailored for small defect inspection, compared with state-of-the-art methods in both image-level and pixel-level defect inspection. The IUF outperformed these methods, providing superior adaptability to novel objects while retaining the ability to recognize previously learned ones. Various experimental protocols were set up to simulate single-step and multi-step learning scenarios, with the new approach showing state-of-the-art performance in all cases.
Conclusion
IUF offers a significant advancement in small defect inspection, proving essential for industrial settings that undergo consistent change and require scalable solutions. By combining OASA, SCL, and the new updating strategy, IUF effectively manages the challenges presented by evolving product lines and the incorporation of new objects into the learning process. The paper concludes with the promise of the framework's deployment potentially enhancing the efficiency of industrial manufacturing while curtailing associated costs.