💻 computer science

Data-Driven Generation of Synthetic Timber for Deep-Learning-Based Wood Assessment

This paper introduces a comprehensive framework that combines LSTM-based knot generation, stochastic growth modeling, and conditional diffusion models to create realistic synthetic Norway spruce boards with known internal structures, thereby enabling the training of deep-learning systems for wood assessment tasks where real paired data is scarce.

Original authors: Osama Abdeljaber, Min Hu, Caroline Marc, Anders Olsson

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Osama Abdeljaber, Min Hu, Caroline Marc, Anders Olsson

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to be a master carpenter. To do this, the robot needs to learn how to look at a piece of wood and understand what's happening inside it—where the knots are, how the grain twists, and where the center of the tree (the pith) used to be.

The problem is that to teach a robot this, you usually need a massive library of examples where you have both the outside picture of the wood and the exact inside map. But getting that inside map is a nightmare. You'd have to cut the wood open or use expensive, slow X-ray machines to see the hidden structure. It's like trying to teach someone to guess the filling of a chocolate bar by looking at the wrapper, but you can only get the "answer key" by smashing the chocolate apart.

This paper presents a clever solution: a digital wood factory that creates perfect, fake wood.

Here is how they built this factory, step-by-step, using simple analogies:

1. The Knot Predictor (The "Storyteller")

First, the team needed a way to decide where knots (the hard, knobby parts of wood) should appear on a log. They couldn't just place them randomly; real trees have specific patterns.

The Analogy: Imagine an AI that has read thousands of real logs. It learned the "grammar" of wood. Just like a language model predicts the next word in a sentence, this model predicts the next "knot" in a log. It knows that if there's a knot here, there's a certain chance of another one appearing a few inches down the line.
The Result: It generates a realistic list of knots, just like a real tree would grow them.

2. The Growth Layer Builder (The "Architect")

Next, they needed to build the rings around those knots. Trees aren't perfect cylinders; they lean (crook), get thinner at the top (taper), and the rings get squished or stretched around knots.

The Analogy: Think of this as a 3D printer that builds the tree's skeleton. It takes the knot list and wraps layers of "wood" around them. It knows to make the layers wobble around a knot (like a bump in a road) and to make the whole tree lean slightly.
The Result: A virtual log with a mathematically perfect internal structure, including exactly where every ring and knot is located in 3D space.

3. The Fiber Mapper (The "Traffic Director")

Wood fibers are like tiny highways that run along the tree. When they hit a knot, they have to swerve around it.

The Analogy: Imagine a traffic director standing on the virtual log. As the "cars" (fibers) approach a "roadblock" (a knot), the director tells them exactly how to curve around it.
The Result: A complete map showing the direction of every single fiber inside the wood, even in the hidden parts.

4. The Realistic Painter (The "Digital Artist")

Now they have a perfect 3D map of the inside, but it looks like a wireframe diagram. They need to make it look like a real photo of wood for the robot to learn from.

The Analogy: They used a special type of AI artist (based on a technology called Stable Diffusion). Instead of painting from a text prompt like "a sunset," they gave the artist the "blueprints" (the ring patterns and fiber directions) and said, "Paint a photo of wood that matches these blueprints exactly."
The Result: A photorealistic image of the wood's surface that perfectly matches the hidden internal map they created earlier.

The "Proof of Concept" Test

To see if this fake wood factory actually works, the researchers tried a specific challenge: Can a robot learn to guess the cross-section of a tree just by looking at the four sides of a board?

The Setup: They trained a robot using only the fake wood generated by their factory. The robot saw the outside of the fake boards and was shown the "answer key" (the inside cross-section) because the factory knew it perfectly.
The Test: They then showed the robot real, actual wood boards it had never seen before.
The Outcome: The robot, trained entirely on fake data, was surprisingly good at guessing the inside structure of the real wood. It could estimate where the center of the tree was and how wide the rings were with high accuracy.

Why This Matters

The paper concludes that this "digital wood factory" is a powerful tool. It allows scientists to create unlimited amounts of training data with perfect "answer keys" for tasks where getting real data is too expensive, too slow, or destructive.

In short: They built a video game engine for wood. They taught an AI to play the game perfectly, and then showed that the AI could actually help solve real-world problems with real wood, even though it only learned in the virtual world.

Note: The authors explicitly state this is a "proof of concept." They are not claiming this replaces all real testing yet, but rather that it provides a powerful new way to train AI when real data is hard to get.

Technical Summary: Data-Driven Generation of Synthetic Timber for Deep-Learning-Based Wood Assessment

Problem Statement
The development of deep-learning (DL) models for wood assessment is hindered by the scarcity of large-scale, paired datasets where external surface scans are linked to known internal structures (e.g., knot geometry, growth layers, 3D fiber orientation, and pith location). Obtaining such ground truth for real timber is often destructive, expensive, or time-consuming, requiring X-ray CT scanning or physical cutting. Furthermore, manual annotation of wood features for supervised learning is labor-intensive. While synthetic data has been explored in other fields (e.g., autonomous driving, medical imaging) to overcome annotation bottlenecks, existing wood-generation frameworks are limited. Previous approaches either focused solely on clear wood without knots (Habite et al., 2022) or generated plausible volumetric textures without calculating 3D fiber orientations or generating scanner-like surface images for specific board dimensions (Larsson et al., 2022). There is a need for a comprehensive framework that generates synthetic Norway spruce boards with realistic surface appearances and fully known internal structures to support DL-based grading and assessment tasks.

Methodology
The authors propose a four-component procedural framework for generating synthetic Norway spruce boards. The workflow proceeds from internal structure generation to realistic surface rendering:

Data-Driven Knot Generator:
- Data Source: 97 Norway spruce logs from the open-access TreeTrace_spruce database, segmented via X-ray CT scans.
- Parameterization: Knots are extracted as point clouds and fitted to a geometric model (Foley, 2003) using 10 parameters (longitudinal position, azimuth, axis parabola coefficients, diameter coefficients, and radial boundaries for live/dead parts).
- Clustering: To address data sparsity, 16,319 unique knots are clustered into 1,000 groups based on their 9-dimensional geometric embeddings (excluding longitudinal position).
- Sequence Generation: A Long Short-Term Memory (LSTM) model is trained on sequences of these cluster tokens (derived from 1-cm discretized log slots) to predict the next knot token. The model achieves 93.2% accuracy in predicting cluster tokens and reproduces statistical properties of knot sequences (e.g., knot-run lengths, frequency) comparable to real data.
Growth-Layer Generator:
- Stochastic Modeling: Annual rings are generated using a stochastic model (Habite et al., 2022) that creates irregular cross-sectional shapes and variable ring widths based on real log measurements.
- Deformations: The nominal growth layers are modified to include:
  - Taper: Linear reduction in radius along the log length.
  - Crook: Stochastic shifts in transverse coordinates modeled as a sum of sinusoidal waves.
  - Knot Bumps: Local distortions around knots are calculated using a modified Foley model, accounting for multiple knots and the presence of crook/taper. This involves calculating influence zones and adjusting the growth layer radius based on live/dead knot boundaries.
Fiber-Orientation Model:
- 3D Assignment: Based on the generated knot geometry and deformed growth layers, a 3D fiber direction is assigned to every point in the board.
- Mechanism: The model extends Foley's single-knot flow pattern to multiple knots. It calculates a projected fiber direction around each knot (Rankine-oval flow) and enforces tangency to the local growth layer surface.
- Combination: For points influenced by multiple knots, a relevance weight is assigned based on distance to the knot axis. The final fiber direction is selected to maximize the transverse deviation score, ensuring the fiber field remains tangent to the growth layers. Inside knots, fibers follow the knot axis.
Realistic Surface Image Generator (Conditional Diffusion):
- Input: Binary ring maps and grayscale in-plane fiber-orientation maps derived from the intersection of the 3D structure with the board surfaces.
- Model: A modified Stable Diffusion XL (SDXL) model. The U-Net input channels were expanded from 4 to 8 to accommodate the conditioning maps (4 ring maps + 4 fiber maps).
- Training: The model was trained on 35,479 segments of real Norway spruce boards (45×145 mm²). Since manual ring labeling of the entire dataset was infeasible, an auxiliary SDXL model was trained to automatically label ring patterns from RGB images, creating a large paired dataset of condition maps and target RGB images.
- Output: The model generates realistic 512×512 RGB images for the four board faces, preserving the prescribed ring patterns and fiber deviations.

Key Contributions

Comprehensive Framework: The study presents the first complete framework for generating synthetic timber boards that simultaneously models knot sequences, growth-layer geometry (including taper, crook, and bumps), 3D fiber orientation, and realistic surface appearance.
Data-Driven Knot Sequencing: Unlike previous procedural methods, the knot sequence is generated by an LSTM trained on real CT data, ensuring statistical fidelity to natural knot distributions.
Integration of Physics and DL: The framework bridges analytical geometric models (for internal structure and fiber flow) with generative AI (diffusion models) to produce visually realistic outputs conditioned on physical parameters.
Open-Source Release: The entire board generator, including source code and trained models, is made publicly available to facilitate further research.

Results and Preliminary Application
To demonstrate the utility of the framework, the authors conducted a proof-of-concept application: training a model to predict the cross-sectional ring pattern (including pith location and ring width) from images of the four longitudinal board surfaces.

Training: A model was trained exclusively on 30,000 synthetic board segments generated by the proposed framework.
Testing: The model was evaluated on 76 real Norway spruce boards (152 cross-sections) not seen during training.
Performance:
- A standard U-Net produced accurate edge predictions but blurry inner regions due to the ambiguity of internal ring patterns.
- A pure diffusion model generated non-blurry outputs but "hallucinated" irregular, non-wood-like patterns.
- A combined pipeline (U-Net for initial prediction + SDXL refiner for detail) achieved high accuracy.
- Metrics: The predicted average annual ring widths showed a coefficient of determination ( $R^2$ ) of 0.93 with a mean absolute discrepancy of 0.173 mm. Pith location predictions showed mean absolute discrepancies of 3.77 mm (x-direction) and 6.08 mm (y-direction).

Significance and Claims
The paper claims that the proposed framework successfully addresses the bottleneck of obtaining large-scale paired datasets for wood assessment. By generating synthetic boards with known internal structures, the framework enables the training of DL models for tasks where real ground truth is difficult or destructive to obtain.

Modest Claims: The authors explicitly state that the results should be viewed as a "proof of concept." They acknowledge that while the synthetic data was sufficient for the cross-section prediction task, more complex tasks (e.g., full 3D knot geometry prediction) may still require fine-tuning with real data due to the domain gap between synthetic and real boards.
Utility: The generator is positioned not as a replacement for experimental data, but as a support tool for developing and testing learning-based methods, potentially reducing the volume of real paired data required for training and validation.
Limitations: The authors note limitations regarding species specificity (currently Norway spruce only), fixed board dimensions, and the computational cost of the diffusion-based refinement pipeline, which currently precludes real-time industrial application. Future work is suggested for species adaptation (e.g., via LoRA) and resolution improvements.