Thursday, November 19, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News Technology and Engineering

Standalone Vision Transformers Outperform Multi-Stage Pipelines in Two-Phase Bubble Segmentation

November 19, 2026
in Technology and Engineering
Denise Maddox
By Denise Maddox Scienmag Editorial Profile - Mechanical Engineering
Reading Time: 7 mins read
0
Standalone Vision Transformers Outperform Multi-Stage Pipelines in Two-Phase Bubble Segmentation

Standalone Vision Transformers Outperform Multi-Stage Pipelines in Two-Phase Bubble Segmentation

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

In quantitative fluid mechanics and nuclear thermal-hydraulics, measuring the rapid birth, deformation, and interaction of individual bubbles is essential for predicting heat removal and coolant dynamics during normal operations and accident transients. In a comprehensive investigation evaluating how training dataset volume and model structure govern the measurement of dispersed gas-liquid interfaces, researchers have demonstrated that standalone transformer-based instance segmentation architectures substantially outperform multi-stage pipelines that combine object detectors with large promptable vision foundation models. Specifically, the end-to-end vision transformer RF-DETR-Seg-2XL achieved an average precision across intersection over union thresholds from 0.50 to 0.95 of 0.887, accompanied by an instance recall of 0.973, an F1 score of 0.983, and a mean matched boundary overlap of 0.935 when tested on an independent 600-image benchmark. By contrast, the highest average precision attained by sequential pipelines pairing trained detectors with Segment Anything Model variants was 0.821, produced by combining an RF-DETR-2XL detector with SAM 3. Beyond delivering higher interface fidelity, the standalone transformer operated with an average processing latency of 193.7 milliseconds per frame, running markedly faster than sequential foundation-model workflows that required anywhere from several hundred milliseconds to multiple seconds per image.

The physical motivation behind resolving individual bubbles stems directly from nuclear engineering, where interfacial area concentration dictates mass, momentum, and energy transport across phase boundaries in core boiling channels and steam-water mixtures. Historically, experimentalists examined bubbly flows using intrusive instrumentation, such as dual conductivity probes or wire-mesh sensors, which can physically disturb local flow fields and yield measurements only at discrete spatial coordinates. Non-intrusive high-speed photographic imaging circumvents intrusive sensor disruption by offering continuous spatial observations across optical windows, but extracting quantitative data from raw frames introduces severe analytical hurdles. Bubbles continually undergo violent shape deformations, coalesce, collide, and visually overlap along the line of sight, while illumination gradients, optical refraction through channel boundaries, and motion blur frequently obscure interface perimeters. Conventional image analysis relies on rigid multi-step deterministic algorithms that combine intensity thresholding, watershed partitioning, morphological filtering, and geometric curve fitting. However, these classical pipelines demand extensive empirical parameter tuning and often fail when encountering non-spherical boundaries, dense bubble swarms, or irregular interfacial curvature, prompting the transition toward modern deep learning models that learn representations directly from annotated imagery.

To build a robust training foundation without relying exclusively on laborious manual labeling, the research team assembled a diverse collection of 2,000 annotated images spanning three distinct image domains: experimental high-speed recordings, semi-organic scenes synthesized via a generative adversarial network, and fully synthetic computer-generated bubble distributions. Experimental images were captured using a precision polydimethylsiloxane microfluidic test facility featuring a rectangular observation channel measuring 5 millimeters in width by 2 millimeters in depth. Air and deionized water were injected through dedicated syringe pumps over superficial gas velocities ranging from 0.1 to 9 centimeters per second and superficial liquid velocities from 1 to 40 centimeters per second. The transparent channel was illuminated by diffused light-emitting diode backlighting and recorded by a high-speed camera fitted with a 180-millimeter macro lens at frame rates reaching up to 6,250 frames per second and an optical resolution of approximately 90 pixels per millimeter. Because initial deterministic edge-detection and watershed routines introduced errors on overlapping and deformed bubbles, all 1,200 experimental frames underwent systematic visual inspection and manual polygon correction in the Computer Vision Annotation Tool to establish true interface boundaries.

To complement the experimental data, the researchers integrated 500 semi-organic images generated by BubGAN, a conditional deep convolutional generative adversarial network trained on 10,000 individual experimental bubble patches. BubGAN synthesizes realistic bubble templates conditioned on physical aspect ratio, orientation angle, circularity, and edge contrast, which are subsequently assembled into full channel scenes with known coordinates. The remaining 300 images were created using a custom MATLAB synthesis algorithm that placed spherical bubbles with diameters ranging from 0.1 to 2.5 millimeters into a simulated 5-millimeter by 2-millimeter channel domain at a spatial resolution of 93.6 pixels per millimeter. This simulation modeled optical superposition along the depth axis by allowing bubbles to overlap whenever the sum of their diameters remained below the 2-millimeter physical channel depth. It rendered bubble interiors with realistic radial grayscale profiles, bright internal highlights, dark interface perimeters, and standard deviations of Gaussian blurring up to 5 pixels to mimic experimental optical defocusing and contrast gradients. Combining these three domains ensured that the models encountered diverse boundary sharpness, optical artifacts, and bubble clustering scenarios while providing exact, automated ground-truth masks for the synthetic and generative datasets.

To evaluate how model performance scales with the volume of available training data, the 2,000 images were structured into seven nested subsets containing 500, 750, 1,000, 1,250, 1,500, 1,750, and 2,000 images, maintaining strict 70:15:15 partitions for training, validation, and testing across all iterations. The nested hierarchy ensured that larger datasets expanded upon smaller collections by adding newly sampled images rather than replacing existing instances. Across these datasets, the researchers evaluated standalone instance-segmentation networks alongside sequential two-stage pipelines. The standalone group comprised Mask R-CNN with a ResNet-50 feature pyramid network, single-stage YOLO26x-Seg featuring a task-specific segmentation head, and transformer-based RF-DETR-Seg evaluated at Small, Large, and 2XL scales incorporating DINOv2 vision backbones. In parallel, the sequential detection-segmentation approach decoupled bubble localization from boundary segmentation by utilizing trained YOLO26n, YOLO26x, RF-DETR-Nano, or RF-DETR-2XL bounding-box detectors to supply spatial prompt coordinates to frozen, pretrained foundation models, specifically Segment Anything Model 2.1 Hiera-L and Segment Anything Model 3. All trainable architectures were fine-tuned on an NVIDIA Tesla T4 graphics processing unit using early stopping criteria based on validation average precision.

The generalization capability of all trained configurations was systematically benchmarked against an independent 600-image dataset that was strictly excluded from training and validation, comprising 200 experimental, 200 semi-organic, and 200 synthetic images containing a total of 6,689 annotated bubble instances. Performance was assessed through standard computer vision metrics, including mask average precision across IoU thresholds from 0.50 to 0.95, precision, recall, and maximum F1 score, complemented by domain-specific measurement metrics directly relevant to fluid mechanics. Predicted and ground-truth instances were matched one-to-one using the Hungarian optimization algorithm based on mask intersection over union, filtering valid matches at an overlap threshold of 0.50. From these matched pairs, the researchers calculated the mean absolute percentage error in equivalent spherical diameter derived from segmented mask area, the Euclidean distance error in centroid localization, and the image-level mean absolute error in total bubble instance count. This rigorous evaluation framework ensured that geometric boundary distortions and instance omissions were captured alongside conventional detection statistics.

Examining performance as a function of training dataset size revealed distinct architectural behaviors and scaling trajectories. Mask R-CNN demonstrated remarkable data efficiency at small sample sizes, achieving a combined average precision of 0.873 with 500 training images and peaking at 0.887 with 750 images, where it briefly outperformed the transformer architectures on experimental frames. However, this early advantage plateaued and subsequently receded to 0.865 when trained on the full 2,000-image dataset, indicating that the convolutional baseline could not effectively leverage larger data volumes under the tested optimization parameters. In contrast, the RF-DETR-Seg transformer family exhibited steady, consistent improvements as dataset size expanded. Combined average precision for RF-DETR-Seg-2XL rose monotonically from 0.870 at 500 images to 0.887 at 2,000 images, with the Small and Large variants progressing from 0.850 to 0.870 and from 0.861 to 0.878, respectively. In the sequential pipelines, performance was dominated by upstream detector capacity rather than the choice of promptable segmentation model; switching between SAM 2.1 and SAM 3 shifted average precision by a negligible 0.002, whereas replacing a lightweight YOLO26n detector with RF-DETR-2XL produced substantial detection gains across every image category.

In head-to-head comparisons among models trained on the complete 2,000-image dataset, RF-DETR-Seg-2XL established clear superiority in overall boundary delineation and localization fidelity. It attained the top overall average precision of 0.887, matched precision of 0.994, recall of 0.973, F1 score of 0.983, and mean matched intersection over union of 0.935, while achieving an exceptionally low centroid error of 1.118 pixels and a bubble-count error of 0.368 instances per image. Mask R-CNN retained a slight advantage in equivalent diameter estimation, achieving a diameter mean absolute percentage error of 2.363 percent compared to 2.631 percent for RF-DETR-Seg-2XL, but suffered higher centroid displacement at 1.581 pixels and a larger bubble-count error of 0.547 per frame. YOLO26x-Seg lagged behind the standalone group with an average precision of 0.785, exhibiting irregular boundaries and elevated count errors of 0.892 bubbles per image. Although the best sequential pipeline, RF-DETR-2XL paired with SAM 3, achieved a low bubble-count error of 0.317 bubbles per frame, its overall average precision remained capped at 0.821, its centroid error reached 2.385 pixels, and its equivalent diameter error stood at 3.404 percent, underscoring the inability of unadapted foundation models to match dedicated end-to-end instance segmentation.

Crucially, breaking down model accuracy across individual benchmark domains exposed substantial performance disparities between synthetic and physical imagery. When evaluated across specific domains, RF-DETR-Seg-2XL achieved an average precision of 0.962 on semi-organic BubGAN images and 0.906 on synthetic images, but dropped to 0.792 on experimental recordings. While instance detection remained reliable across all domains—evidenced by an experimental precision of 0.984, recall of 0.946, and an initial AP at an IoU threshold of 0.50 of 0.945—performance degraded sharply under stringent overlap criteria. As the required overlap threshold increased from 0.50 to 0.95, experimental average precision plummeted by 0.535 down to 0.410, whereas semi-organic and synthetic precision experienced far smaller declines of 0.147 and 0.353, settling at 0.843 and 0.631. This steep drop reflects subtle interfacial ambiguities, optical refraction through channel walls, and manual annotation variance in experimental frames, demonstrating that while models reliably detect bubble presence, extracting ultra-precise boundary coordinates from physical high-speed photography remains far more challenging than processing clean synthetic interfaces.

Computational efficiency measurements further cemented the practical superiority of standalone architectures for high-throughput laboratory analysis. On the benchmark hardware, RF-DETR-Seg-Small processed frames in just 54.8 milliseconds, while the Large and 2XL variants required 75.0 and 193.7 milliseconds per image. By comparison, sequential pipelines incurred heavy computational penalties due to decoupled detector passes, image encoding, and batched spatial prompting: SAM 2.1 pipelines required between 293.6 and 486.8 milliseconds per image, while SAM 3 workflows consumed several full seconds per frame without providing commensurate accuracy improvements. The authors synthesize these findings into concrete experimental recommendations: Mask R-CNN is highly effective when manual training data is strictly limited to 500 to 750 images or when bubble diameter minimization is paramount; RF-DETR-Seg-Small offers optimal throughput for resource-constrained inference; and RF-DETR-Seg-2XL stands as the premier overall architecture for high-accuracy bubble tracking and spatial localization. Nonetheless, the authors emphasize that these conclusions reflect two-dimensional optical projections within an optically accessible 5-millimeter by 2-millimeter rectangular microchannel, and extending deep learning segmentation to complex three-dimensional bubbly flows or dense churning regimes will require further validation under broader operating conditions.

Subject of Research: Deep learning instance segmentation for bubbly two-phase flow measurement

Article Title: Effects of training dataset size and model architecture on deep learning instance segmentation for bubble measurement in two-phase flow

Article References: Abrar, K. N., & Liu, Y. (2026). Effects of training dataset size and model architecture on deep learning instance segmentation for bubble measurement in two-phase flow. Frontiers in Nuclear Engineering, 5, Article 2028300. https://doi.org/10.3389/fnuen.2026.2028300

Image Credits: AI Generated

DOI: 10.3389/fnuen.2026.2028300

Keywords: bubble instance segmentation, two-phase flow, high-speed imaging, RF-DETR, Mask R-CNN, YOLO26, Segment Anything Model, SAM 2.1, SAM 3, BubGAN, synthetic image generation, nuclear thermal-hydraulics, microfluidics, average precision, intersection over union, equivalent diameter, centroid error

Cite Scienmag News

Denise Maddox. (November 19, 2026). Standalone Vision Transformers Outperform Multi-Stage Pipelines in Two-Phase Bubble Segmentation. Scienmag. https://scienmag.com/standalone-vision-transformers-outperform-multi-stage-pipelines-in-two-phase-bubble-segmentation/

Denise Maddox. "Standalone Vision Transformers Outperform Multi-Stage Pipelines in Two-Phase Bubble Segmentation." Scienmag, 19 November 2026, https://scienmag.com/standalone-vision-transformers-outperform-multi-stage-pipelines-in-two-phase-bubble-segmentation/. Accessed 19 November 2026.

Denise Maddox. "Standalone Vision Transformers Outperform Multi-Stage Pipelines in Two-Phase Bubble Segmentation." Scienmag. November 19, 2026. https://scienmag.com/standalone-vision-transformers-outperform-multi-stage-pipelines-in-two-phase-bubble-segmentation/

Tags: average precisionbubble instance segmentationBubGANcentroid errorequivalent diameterhigh-speed imagingintersection-over-unionMask R-CNNmicrofluidicsnuclear thermal-hydraulicsRF-DETRSAM 2.1SAM 3Segment Anything Modelsynthetic image generationtwo-phase flowYOLO26
Share26Tweet16
Previous Post

Biomass Pretreatment Trade-Offs Dictate Silicon Anode Performance and Battery Durability

Next Post

Preoperative Glasgow Prognostic Scores Linked to Microvascular Invasion in Hepatocellular Carcinoma

Related Posts

Survey Examines Applications of Diffusion Models in Meteorology and Oceanography
Technology and Engineering

Survey Examines Applications of Diffusion Models in Meteorology and Oceanography

November 19, 2026
Fungal Extract Enables Green Synthesis of Multifunctional Gold Nanoparticles in Laboratory Study
Technology and Engineering

Fungal Extract Enables Green Synthesis of Multifunctional Gold Nanoparticles in Laboratory Study

November 19, 2026
3D Finite Element Simulations Accurately Predict Subsurface Residual Stresses in Deep Rolled Shafts
Technology and Engineering

3D Finite Element Simulations Accurately Predict Subsurface Residual Stresses in Deep Rolled Shafts

November 19, 2026
Digital Twin Adaptive Control Boosts Precipitator Efficiency While Cutting Energy Use
Technology and Engineering

Digital Twin Adaptive Control Boosts Precipitator Efficiency While Cutting Energy Use

November 19, 2026
Wavelet Multilayer Perceptron Architecture Enhances Multi-Label Medical Inspection Image Classification
Technology and Engineering

Wavelet Multilayer Perceptron Architecture Enhances Multi-Label Medical Inspection Image Classification

November 19, 2026
Compact Aerial Vision Model Refines Features to Improve Small Object Detection
Technology and Engineering

Compact Aerial Vision Model Refines Features to Improve Small Object Detection

November 19, 2026
Next Post
Preoperative Glasgow Prognostic Scores Linked to Microvascular Invasion in Hepatocellular Carcinoma

Preoperative Glasgow Prognostic Scores Linked to Microvascular Invasion in Hepatocellular Carcinoma

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Cryosphere Hazards Mobilize Substantial Organic Carbon Overlooked in Mountain Budgets
  • Targeted Mental Health Programs Enhance Hospital Violence Interventions for Trauma Victims
  • Macular Perfusion Varies by Retinopathy Phenotype After Photocoagulation in Diabetic Eyes
  • Kozhikode Coastline Shifts from Millennia of Accretion to Widespread Modern Erosion

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading