Temperature Super Resolution

Learning-based super resolution of thermal images to reconstruct high-resolution temperature fields from low-resolution sensor data.

Problem

Thermal imaging systems often trade spatial resolution for cost, frame rate, or hardware constraints.
As a result, low-resolution temperature maps lack the spatial detail required for accurate analysis, monitoring, or downstream decision-making.

The objective of this project was to reconstruct high-resolution temperature distributions from low-resolution thermal inputs, while preserving both global structure and localized temperature gradients.

Key challenges:

• Strong information loss in low-resolution inputs
• Smooth but spatially complex temperature fields
• Sensitivity to noise and illumination artifacts
• Need for physically plausible reconstructions, not just visually sharp outputs

Approach

Data

• Paired low-resolution and high-resolution thermal images
• Total samples: 9,864 image pairs
• Resolutions:
  • Low-res: 120 × 160 (reshaped from 122 × 160)
  • High-res: 480 × 640
• Train / validation / test split: 80% / 10% / 10%
• Augmentation applied only to training data:
  • Flip, rotation, shift, scale
  • Elastic deformation
  • Brightness, contrast, gamma
  • Hue saturation and noise

Model

• Residual U-Net inspired by SRResUNet
• Encoder–decoder structure with:
  • Residual blocks and PReLU activations
  • Multi-scale feature aggregation via skip connections
• Upsampling performed using PixelShuffle, avoiding checkerboard artifacts common in transposed convolutions
• Final output reconstructs full-resolution temperature maps

Training

• Loss function: L2 reconstruction loss
• Optimizer: Adam
• Learning rate: 1e-4
• Batch size: 32
• Training duration: 100 epochs
• Experiment tracking using Weights & Biases

Evaluation

• Reconstruction loss evaluated on validation and test splits
• Final losses:
  • Training: 0.0114
  • Validation: 0.0146
  • Test: 0.0139
• Qualitative evaluation focused on:
  • Preservation of temperature gradients
  • Structural consistency with ground truth
  • Absence of hallucinated artifacts

Notes & Lessons Learned

• Residual connections are critical for stable convergence in thermal super resolution
• PixelShuffle upsampling produced smoother and more physically plausible outputs than naïve transposed convolutions
• L2 loss was sufficient due to the smooth nature of temperature fields; perceptual losses were unnecessary
• Data augmentation significantly improved generalization, especially under limited sensor variability
• Visual realism must be evaluated alongside numerical loss for safety-critical thermal applications

This project reinforced the importance of architecture choice and data quality when applying super-resolution techniques to physically grounded imaging modalities.