Lightweight YOLOv8n Improvement for Steel Surface Defect Detection and TensorRT Inference Acceleration
ITM Web of Conferences · 2026 · 93 · 02001
DOI: 10.1051/itmconf/20269302001 ↗Kurz gefasst
Fachartikel in der Zeitschrift ITM Web of Conferences, online erschienen am 13.11.2026. Verfasst von Jingyu Zheng. Die Arbeit nennt 15 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Steel surface defect detection demands both reliable accuracy and real-time inference performance for practical industrial deployment. This paper investigates a lightweight improvement of YOLOv8n by replacing selected backbone downsampling convolution layers with GhostConv modules. To evaluate the effect of replacement depth, three variants (Ghost-1, Ghost-2, Ghost-4) are compared with the baseline under the same training protocol. The NEU-DET dataset is re-split with a stratified 8:1:1 ratio into 1441 training, 180 validation, and 179 test images. Experimental results show that Ghost-2, which replaces layers 5 and 7, achieves the best mAP@0.5 of 0.832, exceeding baseline by 0.008. However, the baseline remains superior on the stricter mAP@0.5:0.95 metric (0.495 vs. 0.475 for Ghost-2). TensorRT FP16 deployment provides approximately 2× speedup for both models: baseline latency reduces from 10.64 ms to 5.32 ms, while Ghost-2 reduces from 12.21 ms to 5.55 ms. Ghost-2 retains a slight mAP@0.5 advantage on TensorRT FP16 but is 4.3% slower. The results indicate that partial GhostConv replacement can improve coarse detection accuracy, yet strict localization accuracy and deployment latency require careful trade-off.
Zitieren
APA
Zheng, J. (2026). Lightweight YOLOv8n Improvement for Steel Surface Defect Detection and TensorRT Inference Acceleration. ITM Web of Conferences, 93, 02001. https://doi.org/10.1051/itmconf/20269302001
Vancouver
Zheng J. Lightweight YOLOv8n Improvement for Steel Surface Defect Detection and TensorRT Inference Acceleration. ITM Web of Conferences. 2026;93:02001. doi:10.1051/itmconf/20269302001
BibTeX
@article{Zheng2026,
author = {Zheng, Jingyu},
title = {Lightweight YOLOv8n Improvement for Steel Surface Defect Detection and TensorRT Inference Acceleration},
journal = {ITM Web of Conferences},
year = {2026},
volume = {93},
pages = {02001},
doi = {10.1051/itmconf/20269302001}
}
EUVIMED