SAID: Segment All Industrial Defects with Scene Prompts.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Industrial Automation

Background:

• Image segmentation is crucial for industrial surface inspection to detect product defects.
• Existing methods often require product-specific training, limiting their generalizability.
• Foundation models like SAM offer zero-shot segmentation but struggle with specialized tasks and require manual input.

Purpose of the Study:

• To develop an automated image segmentation model for industrial defect detection that overcomes the limitations of current approaches.
• To improve the accuracy and efficiency of defect segmentation in diverse industrial settings.
• To eliminate the need for manual interaction and post-processing in segmentation tasks.

Main Methods:

• Proposed SAID (Segment All Industrial Defects) model, which encodes prompt-image pairs into scene embeddings using a Scene Encoder.
• Implemented a Feature Alignment and Fusion Module to resolve embedding alignment issues.
• Achieved automatic segmentation without manual intervention.

Main Results:

• SAID demonstrates superior segmentation performance compared to SAM across various industrial scenes.
• In one-shot target scene segmentation, SAID improved mIoU metrics by 5.79% over MSNet and 0.87% over SegGPT.
• The model effectively addresses the alignment challenge between scene and image embeddings.

Conclusions:

• SAID offers a robust and automated solution for industrial defect segmentation.
• The proposed model significantly enhances segmentation accuracy and efficiency in industrial inspection.
• SAID represents a notable advancement over existing foundation models for specialized downstream tasks.