Related Experiment Videos
EGONet: edge guided omni-directional attention with multi-scale bilateral feature integration for surface defect
Kamal M Othman1, Faleh Alqahtani2, Mai Alduailij3
1Department of Electrical Engineering, College of Engineering and Architecture, Umm Al-Qura University, Makkah, 24381, Saudi Arabia.
Scientific Reports
|August 10, 2026
Summary
This study introduces EGONet, a novel network for surface defect segmentation in metals. EGONet effectively captures defect details and context, improving automated inspection accuracy.
Area of Science:
- Computer Vision
- Materials Science
- Artificial Intelligence
Background:
- Surface defect segmentation (SDS) in metallic materials is complex due to defect variability and intricate textures.
- Existing methods fail to integrate fine boundary details, long-range dependencies, and multi-scale context effectively.
- Automated inspection requires robust algorithms for accurate defect detection and characterization.
Purpose of the Study:
- To propose a novel hierarchical architecture, EGONet, for enhanced metallic surface defect segmentation.
- To address limitations in modeling fine boundaries, omni-directional dependencies, and multi-scale context.
- To improve the accuracy and generalization of automated inspection systems for metallic surfaces.
Main Methods:
- Developed the Edge-Guided Omni-Directional Attention Network (EGONet) with multi-level features.
- Integrated Dense Atrous Spatial Pyramid Pooling (DASPP) for multi-scale context and a Sobel-guided Edge Attention Module (EAM) for boundary cues.
- Employed Omni-Directional Attention (ODA) and Quad-Statistical Spatial Attention (SAM) for long-range dependencies and spatial sensitivity, alongside Efficient Channel Attention (ECA) and Bilateral Feature Integration Block (BFIB).
Main Results:
- EGONet demonstrated superior performance in capturing fine-grained details and semantic information.
- The architecture effectively models omni-directional dependencies and multi-scale context.
- Achieved competitive results against state-of-the-art methods on MT-Defect and SD900 datasets, showing leading performance in specific defect categories and strong generalization.
Conclusions:
- EGONet provides a robust framework for metallic surface defect segmentation, overcoming limitations of previous methods.
- The proposed modules significantly contribute to improved detection accuracy and feature representation.
- EGONet offers a promising solution for enhancing automated inspection systems in industrial applications.