Research on the Edge-Discrepancy Collaborative Method for Defect Detection in Casting DR Images.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Casting defects like pores and inclusions in digital radiography (DR) are challenging to detect due to their small size, varied shapes, and complex backgrounds.
• Existing methods struggle with accuracy, especially for subtle defects.

Purpose of the Study:

• To develop an advanced defect detection framework for casting DR imagery.
• To enhance the accuracy and robustness of defect identification in industrial casting processes.

Main Methods:

• Proposed MTS-YOLOv11, a framework built on YOLOv11, incorporating a Multi-Scale Edge Information Enhancement System (MSEES).
• Integrated a TripletAttention mechanism to manage channel-spatial dependencies and reduce false positives from background textures.
• Implemented a Scale-Discrepancy-Aware Gated Fusion (SDAGFusion) module for effective multi-scale feature fusion.

Main Results:

• MTS-YOLOv11 achieved mAP@0.5 of 96.5% and mAP@0.5:0.95 of 68.5% on the casting DR dataset, outperforming the baseline YOLOv11.
• Demonstrated an inference speed of 359.07 FPS with 2.72M parameters and 7.8G FLOPs.
• Showed superior cross-dataset generalization on new industrial DR data, indicating robustness.

Conclusions:

• MTS-YOLOv11 offers improved detection accuracy and robustness for casting defects in DR images.
• The framework balances high precision with computational efficiency, suitable for real-world foundry inspection.
• The proposed enhancements effectively address challenges posed by small-scale defects and complex backgrounds.