PaperlayerView exact revision on arXiv

Source-linked AI summary

Seeing the Heat: Synthesizing High-Resolution Wood Thermal Responses from Optical Imagery

Jingren Xie

arXiv:2609.31737v1cs.CVcs.LG

TL;DR

Pixel-level wood thermal analysis is constrained by low-resolution, noisy infrared measurements. The paper maps optical intensity to thermal conductivity in a finite-element data engine and uses the synthesized responses to train DINOv3-based surrogates. It reports morphological similarities under uniform or linear thickness-direction conductivity, alongside strong surrogate results and a mesh-convergence-supported discretization.

  • Problem

    Pixel-level wood thermal analysis lacks high-fidelity measurements because infrared thermography is low-resolution and noisy.

    Source: §1 · p. 1

  • Method

    A physics-based FEM data engine maps RGB-derived optical intensity to thermal conductivity and synthesizes responses used to supervise DINOv3-based neural decoders.

    Source: §1 · p. 2

  • Results

    Wood RGB images and thermal responses show morphological similarity with uniform or linear thickness-direction conductivity, while the smallest-backbone R-CNN achieves MAE 0.0702 and RMSE 0.0995.

    Source: §6.2 · p. 10, §6.3.1 · p. 11

  • Takeaways & Limitations

    The Patch-to-Pixel Refiner addresses standard-upsampling blur by anchoring thermal predictions to complex wood anatomy.

    Source: §7 · p. 13

    The conductivity mapping assumes locally uniform conductivity in a micro-neighborhood and uses constant anisotropy ratios, setting α = β for simplicity.

    Source: §4.1 · p. 3

Abstract

from arXiv · show

The thermal behavior of wood is a critical factor in advanced material assembly. However, pixel-level thermal analysis remains fundamentally constrained by the low resolution and noise inherent to infrared thermography. To address this, we introduce an end-to-end computational framework that synthesizes high-resolution thermal responses directly from wood RGB images. We first establish a core physical linkage: because spatial color variation in natural wood is driven by cellular anatomy, optical intensity serves as a reliable geometric proxy for the localized solid volume fraction. By leveraging this theoretical insight, we develop an automated finite-element-method data engine that maps pixel-level optical intensity to a 3D thermodynamic voxel grid, generating high-fidelity synthetic thermal responses. We find that 1) when the thermal conductivity along the thickness direction is uniform or linear, wood RGB images and their corresponding thermal responses exhibit extreme morphological similarities, and the lateral thermal diffusion acts as a low-pass filter that smooths out high-frequency details; 2) when the thermal conductivity along the thickness direction is random, such morphological similarities are destroyed, and wood's 3D structure dominantly governs its thermal response. We further utilize these synthetic thermal responses to supervise a neural surrogate model built upon the DINOv3 foundation model. Our results demonstrate that the neural surrogate model successfully internalizes the governing thermodynamic laws, thereby bypassing computationally expensive simulations and enabling high-resolution thermal inference. This methodology effectively bridges the semantic and thermodynamic domains, unlocking systematic, pixel-level analysis of fine-grained wood thermal responses. Project: https://zekifayes.github.io/seeheat

1 Introduction

Pixel-level wood thermal analysis lacks high-fidelity measurements, while infrared thermography is low-resolution and noisy. The paper proposes a physics-based framework that maps RGB imagery to thermal responses and uses them to train DINOv3-based neural surrogates.

1 Introduction

  • The framework addresses the lack of high-resolution, high-fidelity pixel-level thermal data, which limits systematic analysis of wood responses.Infrared thermography can capture thermal data, but it is low-resolution and noisy.Source: §1 · p. 1
  • Cellular anatomy drives natural wood’s spatial color variation, supporting optical intensity as a geometric proxy for local solid volume fraction.This physical linkage motivates mapping optical information to thermodynamic behavior.Source: §1 · p. 1
  • The proposed physics-based data engine maps wood RGB images to thermal responses, and the synthesized pairs supervise DINOv3-based neural decoders.The surrogates are designed to predict pixel-level responses while bypassing computationally expensive finite-element simulations.Source: §1 · p. 2, Figure 1 · p. 2

2 Related Work

The related work describes self-supervised visual models as general-purpose feature learners, while noting that their visual representations are detached from thermodynamic reality. This motivates investigating whether DINOv3 features can regress thermal responses.

2 Related Work

  • Self-supervised models such as DINO learn patch-level visual representations from large, diverse datasets without human annotation.The section also describes masked autoencoders and multimodal models including CLIP and SigLIP.Source: §2 · p. 2
  • The related work presents DINO-series models and masked autoencoders as examples of self-supervised visual representation learning.It contrasts these with CLIP and SigLIP, which align visual and textual representations through contrastive objectives.Source: §2 · pp. 2–3
  • Foundation models capture geometric and structural context, but DINOv3 features are optimized for visual semantics rather than thermodynamic reality.The paper investigates their capacity to regress complex, high-frequency thermal responses in computational thermodynamics.Source: §2 · p. 3

3 Wood Thermodynamic Modeling

Wood’s internal temperature is modeled through steady-state heat conduction over a thin spatial domain with spatially varying conductivity. The model specifies boundary conditions for contact with the testbed, convection at the top surface, and the remaining lateral surfaces.

3 Wood Thermodynamic Modeling

  • The model describes temperature within a thin wood-sample domain using spatially varying conductivity along all three directions.The domain dimensions satisfy Lx = Ly ≫ Lz.Source: §3 · p. 3
  • The sample’s contact with the testbed is represented by a Dirichlet boundary condition.The passage introduces this condition for the sample’s direct contact with the testbed.Source: §3 · p. 3
  • Heat exchange between the top surface and ambient air uses a Robin boundary condition, while the remaining lateral surfaces use Neumann conditions.The top-surface condition uses the top conductivity, convective coefficient, surface temperature, and ambient temperature.Source: §3 · p. 3

4 Visual-Thermal Computational Framework

The framework maps optical intensity to thermal conductivity and uses finite-element simulation to produce surface thermal responses. Its conductivity mapping relies on assumptions about the available 2D imagery, while the mesh study supports an efficient thin-plate discretization.

4 Visual-Thermal Computational Framework

  • The pipeline’s stated purpose is to synthesize high-resolution wood thermal responses computationally.It bridges optical intensity with a 3D thermodynamic solver.Source: §4 · p. 3, §4.2 · p. 4

4.1 Visual-Thermal Conductivity Mapping

  • A Gaussian kernel approximates lateral thermal diffusion in the surface response.The kernel is applied to account for the lateral diffusion term α∇2T.Source: §4.1 · p. 4
  • The 2D optical intensity is converted into a solid-volume-fraction proxy, yielding a conductivity mapping because solid cell walls conduct more than internal air.The mapping assumes locally uniform conductivity in a micro-neighborhood and constant anisotropy ratios, with α = β for simplicity.Source: §4.1 · pp. 3–4
  • The fitted system parameter η varies by only about 10^-3 across multiscale image-processing settings for the reported datasets.The reported datasets are Poplar, Grandis-CC, and Grandis-RC.Source: §4.1 · p. 4, Figure 2 · p. 5

4.2 Visual-Thermal Computational Framework

  • The FEM data engine reconstructs a 3D mesh, projects the optical proxy through its thickness, and solves for a surface thermal response.The mesh’s in-plane resolution matches the input image pixels, while stretched elements reduce computation in the thin-plate regime.Source: §4.2 · pp. 4–5
  • Relative error against the nz = 20 baseline falls below 10^-2%, supporting stretched elements such as nz = 10 for efficiency.The passage reports that the deviation is several orders of magnitude below the 0.5% tolerance threshold; Figure 3 presents the convergence analysis.Source: §4.2 · p. 5, Figure 3 · p. 6

5 Wood Thermal Prediction

The framework uses frozen DINOv3 features to predict wood thermal responses from RGB images, with a progression of decoders mapping semantic features to dense outputs. Patch-to-pixel refiners project token features into pixel blocks, while other heads rely on interpolation or optical-input fusion.

5 Wood Thermal Prediction

  • To avoid expensive dense PDE computations, the framework uses pretrained DINOv3 features to regress thermal responses from wood RGB images.The encoder-decoder is supervised with the synthesized paired dataset.Source: §5 · p. 6, Figure 4 · p. 7

5.1 Foundation Feature Extraction

  • Learnable softmax weights fuse representations from all DINOv3 transformer layers into a spatial feature map used by subsequent decoders.The frozen encoder converts RGB images into nonoverlapping patch tokens, which are unflattened into the fused map.Source: §5.1 · p. 6

5.2 Decoders

  • The decoder progression compares a linear head, a two-layer nonlinear head, and an optical-incorporation head against patch-to-pixel refinement.The optical-incorporation head concatenates the interpolated semantic map with the raw RGB input before fusion.Source: §5.2 · pp. 6–7
  • Patch-to-pixel refinement projects DINOv3 features directly into dense pixel blocks, with local convolutional and global attention variants.The refiner expands channels to subpixel features and uses PixelShuffle; the global variant applies self-attention before projection.Source: §5.2 · pp. 7–8

6 Experiments and Results

The experiments evaluate synthetic thermal responses and DINOv3-based decoders. They find that diffusion smooths fine detail, thickness-wise conductivity changes thermal patterns, and patch-to-pixel refiners outperform the baselines.

6.1 Experimental Setup

  • The synthetic dataset contains 1,503 paired 512×512 RGB images and thermal responses, split 8:1:1 into training, validation, and test sets.Training uses horizontal and vertical flips and rotation to reduce overfitting.Source: §6.1 · p. 8

6.2 Results on Visual-Thermal Computational Framework

  • Lateral diffusion acts as a low-pass filter: restricted diffusion preserves sharp thermal profiles, while large diffusion dampens localized peaks and fine gradients.Figure 5 shows RGB images alongside simulated surface-temperature maps under varying diffusion coefficients.Source: §6.2 · pp. 8–9, Figure 5 · p. 9
  • Uniform or linear thickness-wise conductivity preserves RGB–thermal morphological similarity, while random conductivity disrupts direct heat paths and makes 3D structure govern the response.Random stratification produces the lowest surface temperatures and nonlinear drops from the 50°C bottom boundary; in-plane anisotropy also alters lateral diffusion.Source: §6.2 · pp. 9–10, Figure 6 · p. 10, Figure 7 · p. 11

6.3.1 Main Results

  • The R-CNN and R-Attn refiners outperform every baseline across all evaluated metrics and backbone scales; ViT-S R-CNN achieves MAE 0.0702 and RMSE 0.0995.The ViT-S R-CNN also surpasses the LH baseline with ViT-L; ViT-L refiners reach δ01 of 99.84%.Source: §6.3.1 · pp. 10–11

6.3.2 Ablation Study

  • Adding one hidden layer and GELU reduces ViT-S MAE from 0.1150 with LH to 0.0840 with TLH, addressing the linear projection’s underfitting.The passage attributes the gain to greater capacity to transform semantic tokens into local energy density and thermal gradients.Source: §6.3.2 · p. 11
  • OIH’s raw-RGB shortcut preserves sharp cellular boundaries, while R-CNN’s patch-to-pixel upsampling avoids interpolation; learned fusion weights favor early transformer layers.OIH reaches ViT-S RMSE 0.1266; the learned weights support the role of low-level geometric features in predicting thermal boundary conditions.Source: §6.3.2 · pp. 11–12, Figure 9 · p. 13
  • At ViT-S, R-Attn lowers MAE from 0.0702 with R-CNN to 0.0652, but at ViT-L R-CNN lowers MAE from 0.0592 with R-Attn to 0.0556.The authors attribute the reversal to global attention helping with the smaller backbone but becoming redundant with larger backbones.Source: §6.3.2 · p. 12
  • Scaling the frozen DINOv3 backbone from ViT-S through ViT-B to ViT-L consistently improves performance across decoder topologies.The authors associate larger backbones with more physically coherent, noise-resilient representations of grain direction and density.Source: §6.3.2 · p. 13

7 Conclusion

The conclusion presents a framework that connects wood RGB imagery to high-resolution thermal analysis. It combines physics-based synthetic data generation with a DINOv3 surrogate, and identifies 3D micro-CT data as a future direction.

7 Conclusion

  • Optical intensity serves as a proxy for cellular solid volume fraction in an FEM engine that synthesizes high-fidelity, pixel-level thermal datasets.The framework maps wood imagery into a 3D thermodynamic representation.Source: §7 · p. 13
  • A DINOv3 neural surrogate predicts thermal responses directly from RGB images to bypass the computational bottlenecks of dense simulations.The Patch-to-Pixel Refiner addresses spatial blurring from standard upsampling by anchoring predictions to wood anatomy.Source: §7 · p. 13
  • Future work will combine 2D images with 3D micro-CT data to characterize wood thermal behavior more accurately.The conclusion also proposes exploring multimodal large models to make the prediction framework more practical.Source: §7 · p. 13

A Heat Transfer and Boundary Condition Verification

The appendix models through-thickness conductivity as a depth-dependent function and introduces effective conductivity and aggregated heat. In the homogeneous case, it sets the aggregated heat to zero and simplifies the boundary equation.

A.1 Thermodynamic Modeling along the Thickness Direction

  • For a single pixel, the model assumes through-thickness conductivity kz varies with z as kz = g(z).This is the stated assumption for modeling thermal behavior along the thickness direction.Source: §A.1 · p. 16
  • The appendix defines effective thermal conductivity keff and aggregated heat Φ along the z axis.The supplied passage introduces these quantities but does not provide their full definitions.Source: §A.1 · p. 16
  • For homogeneous conductivity, Φ = 0 and kz equals the constant keff = ktop, reducing the boundary equation.This case assumes Tbed is uniform across the xy plane.Source: §A.1 · p. 16

A.2 Verification of the Heat Transfer Model and Boundary Conditions

The model’s boundary-condition estimates align with expected indoor temperatures, show spatial consistency across simultaneously measured samples, and remain stable across three statistical estimators.

Validation against Physical Ranges

  • Across all four Oak samples and estimators, T∞ ranges from 21.8 to 25.0°C, matching standard indoor ambient conditions.This supports the model’s anchoring to the physical environment.Source: §A.2 · p. 16

Validation through Spatial Consistency

  • Closely grouped parameter estimates across four simultaneously measured Oak samples support spatial consistency and an almost homogeneous internal medium.The authors take this consistency as validation of the model and boundary-condition equations.Source: §A.2 · pp. 16–17

Stability of Statistical Estimators

  • OA, ME, and IQM yield estimates with small variance, often within fractions of a degree or coefficient unit.IQM trims the upper and lower 25% of spatial data yet aligns nearly perfectly with OA.Source: §A.2 · p. 17

B Derivation of Visual-Thermal Modeling

The derivation connects optical intensity to a spatially smoothed thermal response under thin-plate assumptions. Across datasets, the fitted system parameter and correlation distributions show distinct means and strong spatial mappings.

Derivation of Visual-Thermal Modeling

  • Assuming locally uniform conductivity and equal global anisotropy ratios simplifies the heat equation for a thin plate with Lz = 3 mm.The derivation treats conductivity as uniform through thickness for each pixel while allowing variation across the xy-plane.Source: B Derivation of Visual-Thermal Modeling · p. 17
  • A second-order expansion along z and boundary conditions isolate the temperature difference between the bed and top surfaces.The derivation drops the bed-surface Laplacian term because its thin-plate-scaled contribution approaches zero.Source: B Derivation of Visual-Thermal Modeling · p. 18
  • For small λ, the inverse resolvent approximates spatial diffusion, equivalent to Gaussian convolution with variance σ^2 = αL_z^2.The derivation defines λ^2 = αL_z^2 and maps σ^2 = 2λ^2.Source: B Derivation of Visual-Thermal Modeling · p. 18
  • A first-order approximation expresses inverse temperature difference as a Gaussian-smoothed conductivity field derived from optical intensity and scaled by η.The conductivity mapping uses k_top(x, y) = k_max[1 − I(x, y)], and η is defined as k_max/(L_zq).Source: B Derivation of Visual-Thermal Modeling · pp. 18–19

Distributions of the System Parameter η

C Derivation of the Weak Formulation for the 3D Heat Equation

The appendix formulates steady-state anisotropic heat conduction as a weak problem over a 3D wood domain with specified boundary conditions. It also defines uniform, linear, and layered-random conductivity profiles through the thickness.

C Derivation of the Weak Formulation for the 3D Heat Equation

  • The 3D steady-state model uses a conductivity tensor K and imposes bottom Dirichlet, top Robin, and lateral Neumann boundary conditions.The corresponding trial and test spaces preserve the known bottom temperature.Source: C Derivation of the Weak Formulation… · p. 19
  • Multiplying the strong equation by a test function and integrating over the domain converts the formulation toward a boundary integral.The bottom and lateral boundary terms vanish under the Dirichlet and zero-flux Neumann conditions; the top term uses the Robin condition.Source: C Derivation of the Weak Formulation… · p. 20

D Thermal Conductivity Distribution along the Thickness Direction

Related papers

Loading 2609.31737v1…