Few-shot cross-episode adaptive memory for metal surface defect semantic segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Materials Science

Background:

• Few-shot semantic segmentation is crucial for metal surface defect detection with limited data.
• Existing methods struggle with adaptability and segmentation granularity in few-shot scenarios.

Purpose of the Study:

• To propose an episode-adaptive memory network (EAMNet) for improved few-shot semantic segmentation in metal surface defect detection.
• To address limitations in semantic description and segmentation granularity of previous methods.

Main Methods:

• Developed an episode-adaptive memory unit (EAMU) using an adaptive factor for cross-episode semantic dependency modeling.
• Introduced a context adaptation module (CAM) for fine-grained segmentation by aggregating hierarchical features.
• Proposed global response mask average pooling (GRMAP) for direct fine-grained cue extraction.
• Implemented attention distillation (AD) to stabilize cross-episode adaptation.

Main Results:

• EAMNet demonstrated superior performance in few-shot semantic segmentation for metal surface defects.
• The proposed methods achieved state-of-the-art results on the Surface Defect-[Formula: see text] and FSSD-12 datasets.
• Significant improvements in adaptability and segmentation granularity were observed.

Conclusions:

• EAMNet effectively handles subtle variations across training episodes in few-shot learning.
• The proposed approach enhances the accuracy and granularity of metal surface defect segmentation.
• This work sets a new benchmark for few-shot semantic segmentation in industrial defect detection.