1
Return

EMAN: Efficient feature modulation and aggregation network for defect classification in industrial images

delete2026-06-15
delete0
PRE
AI
V
Vojtech Schiller
A
Anzhelika Mezina *
R
Radim Burget
M
Malay Kishore Dutta
DOI：10.1016/j.imavis.2026.106077delete
deleteOriginal
deleteOriginal request for help
deleteShare
deleteSave
Abstract

Abstract

En 中文
• Proposes EMAN and EMAN-S, novel lightweight convolutional architectures integrating efficient attention mechanisms for weld defect classification from industrial radiographic images. • Achieves state-of-the-art accuracy on four benchmark datasets (RIAWELC, NEU-CLS, StitchingNet, Ball Screw Drive Defect), with EMAN-S reaching 99.80% accuracy using only 145k parameters. • Validates robustness and generalization through comparative evaluation with nine state-of-the-art models, avoiding dataset leakage issues found in prior works. • Provides explainability via Grad-CAM visualizations and detailed performance analysis, demonstrating precise defect localization and effective feature learning.

Journal

Image and Vision Computing cover
Image and Vision Computing
IF:
4.2
Papers:
4.0K
Citations:
6.7K

Organization

B
brno university of technology
Scholars:
801
Papers: 320
Citations: 0
A
amity university
Scholars:
438
Papers: 258
Citations: 0
Cited Papers

Cited Papers

Citing Papers

Citing Papers