Submit

Login Login

Register Register

Home / Journals / CMC / Online First / doi:10.32604/cmc.2026.085242
Journal Menu
Special Issues
Table of Content

All issues

Online First

2026

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

2006

2005

2004

Open Access

ARTICLE

EFAS-YOLO: A Lightweight Edge-Frequency Aware YOLOv11 Framework for Steel Surface Defect Detection

Jiahui Liu, Longzhen Dong*, Zeling Hou
School of Computer Science and Technology, Harbin University of Science and Technology, Harbin, China
* Corresponding Author: Longzhen Dong. Email: email

Computers, Materials & Continua https://doi.org/10.32604/cmc.2026.085242

Received 07 May 2026; Accepted 31 July 2026; Published online 21 August 2026

Abstract

Detecting surface defects on steel is challenging because many defect regions are visually weak, have blurred boundaries, and contain minimal pixel information. In detectors from the You Only Look Once (YOLO) family, these subtle cues may be weakened at the early feature extraction stage and further attenuated during repeated downsampling. To improve the preservation and utilization of such defect-related details, this paper proposes EFAS-YOLO, a lightweight YOLOv11-based detection framework for steel surface defect inspection. First, an Edge-Frequency Aware Stem (EFAS) is introduced before the backbone to explicitly extract Sobel-based gradient responses and fuse them with learnable shallow texture features, allowing edge-sensitive information to be retained from the input stage. Second, the backbone channel configuration and shallow receptive field are adjusted to better match the feature distribution produced by EFAS and to avoid unnecessary channel expansion. Third, a Small Defect Enhancement Path (SDEP) is constructed to transmit enhanced P3-level high-resolution features to the corresponding neck branch through a lightweight residual path, reducing the loss of spatial details for small defects. Experiments on NEU-DET show that EFAS-YOLO achieves 79.8% mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5), outperforming YOLOv11n by 3.6 percentage points while maintaining a lightweight scale of 2.1M parameters and 158 FPS on an RTX 3090 GPU. Additional validation on GC10-DET achieves 68.3% mAP@0.5, suggesting that the proposed design maintains stable performance across different steel surface defect datasets.

Keywords

Real-time surface defect detection; strip steel inspection; convolutional neural network; YOLOv11-based object detection

  • 115

    View

  • 23

    Download

  • 0

    Like

Related articles
Copyright© 2026 Tech Science Press
© 1997-2026 TSP (Henderson, USA) unless otherwise stated

Share Link