Steel Surface Defect Detection Based on Improved GCHS-YOLO Algorithm

Steel Surface Defect Detection Based on Improved GCHS-YOLO Algorithm

In this paper, we address challenges in steel surface defect inspection, such as missed detections and false detections, by proposing the GCHS-YOLO detection algorithm. Built on YOLOv8s, our approach replaces the traditional Feature Pyramid Network (FPN) in the NECK section with a multi-scale fusion...

Full description

Saved in:
Bibliographic Details
Main Authors: Ruiqiang Guo, Peiyong Ji, Yapin Zhang, Jingqi Hu, Wenlong Liu, Xuejian Li, Min Li
Format: Article
Language:English
Published: IEEE 2024-01-01
Series:IEEE Access
Subjects:
Steels
GFPN
CA
HWD
SPDConv
Online Access:https://ieeexplore.ieee.org/document/10798110/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In this paper, we address challenges in steel surface defect inspection, such as missed detections and false detections, by proposing the GCHS-YOLO detection algorithm. Built on YOLOv8s, our approach replaces the traditional Feature Pyramid Network (FPN) in the NECK section with a multi-scale fusion network (GFPN). This change improves the model’s adaptability to targets of varying scales and enhances its ability to extract key defective features.Continue to introduce the Coordinate Attention (CA) mechanism, which integrates location information to improve feature extraction, aiding in better regression and localization. This boosts the network’s ability to identify and detect defects with greater accuracy. Additionally, the Haar Wavelet Downsampling (HWD) module is incorporated to reduce the spatial resolution of feature maps while retaining important information. This not only decreases the model’s complexity but also reduces uncertainty in the extracted information.Finally, the Spatial Pyramid Dilated Convolution (SPDConv) expands the model’s perceptual field, improving feature extraction, especially for small defects. Experimental results using the NEU-DET dataset—after applying noise and Gaussian filtering—show that the GCHS-YOLO algorithm improves the mean Average Precision (mAP) by 1.2%, Precision by 0.8%, mAP@0.5:0.95 by 4.4%, and Recall by 2.8%, compared to the original YOLOv8s model.
ISSN:2169-3536

Similar Items