EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

YOLO-Bubble: A Lightweight and High-Precision Instance Segmentation Model for Industrial Bubble Detection With BubConv Enhancement

Wen Yang, Jianpeng Li, Jindi Jin, Hui Chen, Weiting He, Wenjun Huang

IEEE Access · 2026 · 14 · 118303-118319

DOI: 10.1109/access.2026.3716943 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Wen Yang und 5 weiteren Autor:innen. Die Arbeit nennt 51 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: State Key Laboratory of Industrial Control Technology, China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Rapid and accurate measurement of bubble size is important for industrial visual inspection, because bubbles provide useful morphological cues for process control, material quality assessment, and product state evaluation. Typical scenarios include bubble distribution analysis in porous industrial materials and bubble inspection for Baijiu, a traditional Chinese distilled spirit. Although the imaging media differ, these tasks share common visual challenges, including small targets, low contrast, dense adhesion, irregular contours, and strict requirements for real-time deployment. Existing deep learning models often struggle to balance detection speed and accuracy, making their deployment on edge devices in industrial scenarios challenging. To address this issue, this study proposes a novel, lightweight, and efficient bubble instance segmentation algorithm based on the YOLO11-seg instance segmentation framework, named YOLO-Bubble. First, through feature layer redundancy analysis, a differential channel adjustment strategy is adopted to reduce the model burden. Second, BubConv is designed as a bubble-specific feature enhancement convolution module, where trainable kernels are optimized end-to-end under branch-specific center-response, contrast-enhancement, and angular-difference transformations. Unlike conventional multi-branch convolution modules that mainly increase generic feature diversity, BubConv constrains the weight search space with bubble morphological priors and reparameterizes all branches into a single standard convolution during inference. Finally, a lightweight segmentation head is proposed by reconstructing the feature processing sequence and introducing partial convolution. Experiments are conducted in two representative bubble inspection scenarios: gypsum board bubbles and Baijiu bubbles. Across five independent random seeds, YOLO-Bubble achieves an AP of $61.8\pm 0.4$ % on the self-built gypsum board bubble dataset, improving YOLO11n-seg by 3.3 percentage points, and an AP of $43.7\pm 0.6$ % on the Baijiu bubble dataset, improving YOLO11n-seg by 2.3 percentage points. Meanwhile, the parameter count is reduced by 28.4% and GFLOPs by 21.9%. On an edge industrial controller equipped with an Intel Processor N97 CPU, YOLO-Bubble reduces latency from 58.10 ms to 34.16 ms, increases edge-device FPS from 17.21 to 29.27, and lowers peak memory from 615.26 MB to 550.53 MB compared with YOLO11n-seg. These results provide a technical basis for real-time online bubble measurement across industrial production and quality inspection scenarios.

Förderung

  • State Key Laboratory of Industrial Control Technology, China

Zitieren

APA

Yang, W., Li, J., Jin, J., Chen, H., He, W., & Huang, W. (2026). YOLO-Bubble: A Lightweight and High-Precision Instance Segmentation Model for Industrial Bubble Detection With BubConv Enhancement. IEEE Access, 14, 118303-118319. https://doi.org/10.1109/access.2026.3716943

Vancouver

Yang W, Li J, Jin J, Chen H, He W, Huang W. YOLO-Bubble: A Lightweight and High-Precision Instance Segmentation Model for Industrial Bubble Detection With BubConv Enhancement. IEEE Access. 2026;14:118303-118319. doi:10.1109/access.2026.3716943

BibTeX
@article{Yang2026,
  author = {Yang, Wen and Li, Jianpeng and Jin, Jindi and Chen, Hui and He, Weiting and Huang, Wenjun},
  title = {YOLO-Bubble: A Lightweight and High-Precision Instance Segmentation Model for Industrial Bubble Detection With BubConv Enhancement},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {118303--118319},
  doi = {10.1109/access.2026.3716943}
}
Kostenlos starten →