A Process-Informed, Photorealistic Digital Twin for Vision-Based Inline Inspection of C-SMC Strip Deposition

Patrick Flore; Fabian Röder; Kevin Chen; Andreas Gebhard

cs.CVOct 20, 2026

A Process-Informed, Photorealistic Digital Twin for Vision-Based Inline Inspection of C-SMC Strip Deposition

Authors: Patrick Flore, Fabian Röder, Kevin Chen, Andreas Gebhard

Organizations: Leibniz-Institut für Verbundwerkstoffe GmbH, Erwin-Schrödinger Straße, Geb. 58, Kaiserslautern, 67663, , Germany

Abstract

Inline quality assurance for Sheet Molding Compound (SMC) manufacturing is hindered by the lack of large-scale, accurately labeled image datasets for training machine-learning models. This work presents a process-informed, photorealistic digital twin for carbon-fiber SMC strip deposition. A rigid-body simulation reproduces the deposition kinematics (settling, friction, belt motion), while a physically grounded polarization-rendering model produces realistic appearance, so that every image carries exact, mutually consistent labels in synchronized formats: oriented bounding boxes, point annotations, density maps, and polarization renderings (Angle and Degree of Linear Polarization). Because the joint distribution of many overlapping strips cannot be specified analytically, simulation yields exact, scalable ground truth that manual annotation cannot provide. The twin produces over 16,000 images across 3,300 frames, and the synthetic data replicate near-isotropic fiber orientation (axial concentration r < 0.05) consistent with as-deposited commercial SMC. Trained purely on synthetic data, YOLOv11 and CSRNet reach detection mAP50 up to 0.91, with a dedicated oriented-bounding-box detector confirming the oriented labels are directly trainable (mAP50 = 0.83); on real manufacturing images, the synthetic-to-real difference in edge and texture structure (Kolmogorov-Smirnov distance 0.46, distribution overlap 0.56) is of the same magnitude as the difference between two real acquisition series of the same material (0.41, overlap 0.57). The approach substantially reduces the labeling burden for camera-based inline inspection of fiber orientation and volume content.

Figures & tables

Figure 1: Simulation as displayed in Blender’s GUI. (A) shows the spawning region for the outlet, (B) indicates the camera’s origin point used for rendering, (C) marks the measurement box used for analysis and calculations, (D) represents the rigid-body conveyor-belt model, (E) shows individual fiber strips, here visibly stacked on top of one another and (F) shows the illumination area.

Strips per run	Time per measurement frame	Time per run ( $33$ frames)
$1000$	$11$ s	$6.0$ min
$1250$	$11$ s	$5.9$ min
$1500$	$14$ s	$7.5$ min
$1750$	$27$ s	$14.3$ min
$2000$	$26$ s	$13.7$ min

Table 1: Wall-clock cost of the generation analysis phase (four polarization renders, density computation, and annotation per measurement frame), by deposited strip count, reconstructed from database timestamps on the hardware above. Each run comprises

33

measurement frames; each configuration comprises five seeded runs. One-time scene setup and physics baking are excluded.

Figure 2: Systematic parameter control and multi-format ground truth generation. (A) Temporal evolution of strip density during the deposition simulation. Different curves represent varying configured strip counts ( $N$ ) and belt speeds ( $v$ ), with shaded regions indicating the mean $\pm$ range across multiple random seeds. Lower belt speeds converge to a steady-state packing density after an initial transient; at higher belt speeds the measured density instead decays toward zero once spawning ends, as the accumulated strips are transported out of the measurement volume. Both behaviors are reproducible across seeds. (B) Density progression: representative rendered frames (AoLP, cyclic colormap) at increasing deposited density, each labeled with the number of strips inside the measurement box ( $n$ ) and the 2D and 3D density metrics. (C) Multi-format ground truth for a single frame (shown at moderate density for legibility; the label format is identical at all densities), all generated simultaneously from the simulation state: the input AoLP image, the oriented bounding boxes (OBB), the point (centroid) annotations, and the derived density map. The oriented boxes follow each strip’s true orientation and are exact and per-strip varying, not axis-aligned or at a constant angle. (D) Annotation self-consistency over a representative sample of scenes across all configurations: the density-map integral matches the annotated object count closely (mean $-0.29\%$ , essentially all within $2\%$ ), the small residual reflecting spatial discretization and boundary-projected strip centroids.

Figure 3: Fiber orientation distribution of synthetic datasets: (A) Orientation histogram over the full corpus of $3{,}145$ annotated frames ( $1.07$ million labeled strips), with the uniform-expectation line and a periodic kernel density estimate on the secondary density axis; the axial concentration is $r=0.047$ (statistics in the text), quasi-isotropic up to a weak deposition-induced bias and consistent with as-deposited commercial SMC sheets. (B) AoLP renderings across manufacturing stages, capturing orientation information suitable for validating detection algorithms and training orientation-sensitive networks.

Model	Precision	Recall	$\text{mAP}_{50}$	$\text{mAP}_{50\text{--}95}$
YOLOv11n	0.984	0.793	0.883	0.812
YOLOv11s	0.995	0.810	0.896	0.859
YOLOv11m	0.997	0.818	0.903	0.884
YOLOv11l	0.997	0.820	0.902	0.887
YOLOv11x	0.998	0.822	0.908	0.899

Table 2: YOLOv11 strip detection on the held-out synthetic validation set (

144

images) after

100

converged training epochs, across the five model scales.

Figure 4: Oriented-bounding-box detection across density regimes. Ground-truth (green) and predicted (red) oriented bounding boxes from the single-scale YOLO11s-OBB detector on three held-out synthetic validation frames spanning the density range: (A) low ( $48$ strips), (B) medium ( $340$ strips), and (C) high ( $660$ strips). Predictions are shown at a confidence threshold of $0.25$ with the per-image detection cap raised to accommodate the densest frames. The predicted boxes follow each strip’s position, size, and orientation, and their count closely tracks the ground truth even under the heavy occlusion of the high-density regime.

Density regime	$n$	MAE	RMSE
low ( $\leq 220$ )	32	189.7	202.0
medium ( $221$ – $375$ )	39	42.3	55.1
high ( $>375$ )	29	151.9	166.3
overall	100	121.3	149.2

Table 3: CSRNet count error on the held-out synthetic test split (

100

frames). Density regimes are empirical tertiles of the ground-truth strip count per frame.

Figure 5: Synthetic-to-real domain gap. (A) Representative synthetic (left) and real (right) frames; the panel illustrates similar overall geometry and orientation texture, not a matched scene. The visible residuals are background appearance and slightly sharper synthetic edges, quantified in Fig. 6 A. (B) Per-pixel AoLP intensity distributions of the two domains (standard deviations $\sigma_{\text{syn}}=48.6$ and $\sigma_{\text{real}}=72.7$ on the $0$ – $255$ scale; $\text{KS}=0.57$ ). As analyzed in the text, this raw-intensity difference reflects the difference in background type (a bare conveyor belt rendered at a near-constant polarization value versus the resin- and fiber-covered real background) together with a reference-axis calibration offset in the polarization channel, rather than a difference in strip appearance. (C) Structural similarity (SSIM) over unpaired frame pairs: the synthetic-to-real distribution (median $0.13$ ) is comparable to the real-vs-real reference floor (median $0.10$ ), because the strips occupy uncorrelated positions and pixel-aligned similarity is therefore intrinsically low even between two real frames. (D) Per-frame mean-intensity percent error relative to the real-set mean; the synthetic frames cluster at a systematic offset of $\approx+58\%$ (a background-type and polarization-calibration difference), against the broader frame-to-frame spread of the real set.

Comparison	KS	Wasserstein	Overlap
Synthetic vs. synthetic (two disjoint halves)	0.01	0.8	0.99
Real vs. real (two disjoint halves of one run)	0.03	11.5	0.97
Real laboratory vs. real production run	0.41	59.0	0.57
Synthetic vs. real laboratory	0.46	47.7	0.56
Synthetic vs. real production (mismatched scene type)	0.66	96.4	0.35

Table 4: Reference values for the unpaired appearance comparison, all computed on the per-pixel Sobel gradient magnitude with the identical statistic, pixel sample, and histogram grid. The Kolmogorov–Smirnov distance has no sample-size-independent acceptance threshold and is interpretable only against such references: the first two rows bound the distance attainable within one domain, and the third gives the distance between two real acquisition series of the same material.

Figure 6: Label-free synthetic-to-real fidelity and transfer evidence. (A) Unpaired distribution of the per-pixel Sobel gradient magnitude (edge and texture sharpness) for the synthetic and real image sets; the two distributions overlap by $0.56$ ( $\text{KS}=0.46$ ), matching the overlap of $0.57$ measured between two real acquisition series of the same material (Table 4 ). (B) Rendered AoLP sampled at strip centroids versus the ground-truth oriented-box angle; the mirror-symmetric trend ( $r=0.67$ ) shows that the polarization channel encodes strip orientation. (C) A real frame dot-counted by hand (one dot per visible strip): $458$ strips in a $512\times 512$ crop, with a densest-window inset showing that occlusion makes exhaustive per-strip ground truth infeasible. (D) Temporal stability of the synthetic-trained CSRNet count across consecutive and strided frames of a steady real run (coefficient of variation $0.001$ ), a label-free consistency check.

Synthetic Data Generation

Synthetic data generation creates artificial examples, simulations, and datasets to train, test, or protect machine-learning systems.

See all 623 papers

Data Collection for Training Quality-Control AI in Carpet Manufacturing
May 31, 2026Akbar ErkinovIndustrial Anomaly Detection Industrial Inspection
Independent Researcher
Yarn tracking of large-scale 3D textile reinforcements using topological material features
Sep 15, 2026Hafsa El Herichi, Arturo Mendoza, Yanneck Wielhorski +2Materials Science
LMPS, CVN · Université Paris-Saclay, CentraleSupélec, ENS Paris-Saclay, CNRS, LMPS, 4, Avenue des Sciences, 91192 Gif-sur-Yvette, France · Université Paris-Saclay, CentraleSupélec, Inria, CVN, 9 Rue Joliot Curie, 91190 Gif-sur-Yvette, France +4
Trustworthy Visual Quality Inspection under Data Scarcity in Manufacturing
Aug 22, 2026Panagiotis Sapoutzoglou, Jessy Ribaira, Martin Kanounnikoff +3Industrial Inspection Industrial Anomaly Detection
National Technical University of Athens, Athens, Greece · German Aerospace Center (DLR), Germany · Philips, High Tech Campus 52, 5656 AG Eindhoven, The Netherlands +1

May 31, 2026J/K move · Enter open · S save

Data Collection for Training Quality-Control AI in Carpet Manufacturing

Akbar Erkinov

Independent Researcher

Industrial Anomaly Detection Industrial Inspection

Open paper AskPDF arXiv

Visual inspection remains the dominant quality-control practice in woven and tufted carpet production, yet it is slow, subjective, and inconsistent at the line speeds and widths of modern looms. We present a design proposal for an in-line machine-vision system whose primary purpose is twofold: to inspect the carpet web in real time and, equally importantly, to systematically collect and label images of defect patterns so that increasingly capable quality-control models can be trained over the life of the installation. The proposal is grounded in a concrete industrial setting: a Six Sigma (DMAIC) project at a woven-carpet production facility that anticipated a production bottleneck following the installation of additional weaving machines, with a substantial baseline defect rate and significant financial exposure associated with quality failures. We describe an imaging subsystem based on synchronized line-scan cameras with combined bright-field and grazing illumination, derive the resolution and throughput requirements needed to resolve fine structural defects across a multi-metre web, and define a carpet-specific defect taxonomy. We then lay out a staged modelling strategy that begins with unsupervised anomaly detection trained on defect-free material, following the paradigm exemplified by the carpet category of the MVTec Anomaly Detection benchmark, and matures through a human-in-the-loop annotation flywheel into supervised detection and segmentation models. Finally, we connect detection performance to the DMAIC objectives, showing how reductions in escaped defects translate into improved process quality and process sigma levels. The contribution is an end-to-end, deployable blueprint that treats data collection as a first-class engineering objective rather than an afterthought.

Abstract

Synthetic Data Generation

Explore similar work

Data Collection for Training Quality-Control AI in Carpet Manufacturing

Yarn tracking of large-scale 3D textile reinforcements using topological material features

Trustworthy Visual Quality Inspection under Data Scarcity in Manufacturing