Incremental Learning for Defect Segmentation With Efficient Transformer Semantic Complement.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Semantic segmentation of industrial surface defects is crucial for quality control.
• Existing models struggle with new defect types and catastrophic forgetting during incremental learning.
• Low contrast between defects and backgrounds complicates segmentation.

Purpose of the Study:

• To develop an effective incremental learning method for industrial surface defect segmentation.
• To address the limitations of existing models in handling new defect classes and low-contrast scenarios.
• To improve the adaptability and performance of defect segmentation systems.

Main Methods:

• Introduced a plug-and-play Transformer-based semantic complement module (TSCM) integrating global context into CNNs.
• Proposed multi-scale spatial pooling distillation (MSPD) for preserving spatial relations during incremental updates.
• Implemented an adaptive weight fusion (AWF) strategy for balancing model stability and plasticity.

Main Results:

• The proposed method significantly outperforms existing approaches in incremental segmentation scenarios.
• TSCM effectively fuses global and local information, enhancing segmentation accuracy.
• MSPD and AWF successfully mitigate catastrophic forgetting and improve feature alignment.

Conclusions:

• The developed method offers a robust solution for evolving industrial defect segmentation tasks.
• The combination of TSCM, MSPD, and AWF provides superior performance and adaptability.
• This work advances the capabilities of automated quality inspection in manufacturing.