EUVIMEDEUVIMED entdecken
Crossref

LCS-DETR: a lightweight transformer framework for real-time surface defect detection of ceramic flat sheet membranes

Ning Ma, Yuanyu Chen

Measurement Science and Technology · 12.06.2026 · 37 (24) · 245104

DOI: 10.1088/1361-6501/ae761d ↗

Kurz gefasst

Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 16.06.2026. Verfasst von Ning Ma und 1 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Surface defect detection is a critical task in the quality control of ceramic flat sheet membranes. Existing object detection models frequently exhibit high complexity, limiting their practical deployment on edge devices with constrained computational resources. To address this challenge, a lightweight Transformer framework for real-time surface defect detection of ceramic flat sheet membranes (LCS-DETR) is proposed. Based on RT-DETR-R18, a lightweight backbone network (LCPNet) is first designed, which dynamically adjusts the receptive field of convolutional kernels through partial convolution (PConv) and compensates for the feature sparsity issue of PConv using cheap operations. Subsequently, a lightweight network structure optimization strategy is proposed. By pruning redundant feature layers and introducing an efficient ADown module, the detection model becomes more suitable for deployment on resource-constrained edge devices. Finally, a novel self-attention and convolutional fusion-based intra-scale feature interaction module is introduced. Through the fusion of self-attention mechanisms with convolutional operations, complementary integration of global and local features is realized, further optimizing model performance under lightweight constraints. Experimental results demonstrate that compared with RT-DETR-R18, LCS-DETR reduces Params and floating-point operations by 71.0% and 61.4%, respectively, while improving detection speed by 31.0%. LCS-DETR achieves 98.8% detection accuracy on the self-constructed CS-DET dataset and 70.1% detection accuracy on the public NEU-DET dataset. Furthermore, LCS-DETR achieves 98.4% detection accuracy and 23.3 FPS on Jetson Orin NX Super 8 G, demonstrating its practical value for resource-constrained edge devices.

Zitieren

APA

Ma, N., & Chen, Y. (2026). LCS-DETR: a lightweight transformer framework for real-time surface defect detection of ceramic flat sheet membranes. Measurement Science and Technology, 37(24), 245104. https://doi.org/10.1088/1361-6501/ae761d

Vancouver

Ma N, Chen Y. LCS-DETR: a lightweight transformer framework for real-time surface defect detection of ceramic flat sheet membranes. Measurement Science and Technology. 2026;37(24):245104. doi:10.1088/1361-6501/ae761d

BibTeX
@article{Ma2026,
  author = {Ma, Ning and Chen, Yuanyu},
  title = {LCS-DETR: a lightweight transformer framework for real-time surface defect detection of ceramic flat sheet membranes},
  journal = {Measurement Science and Technology},
  year = {2026},
  volume = {37},
  number = {24},
  pages = {245104},
  doi = {10.1088/1361-6501/ae761d}
}
Kostenlos starten →