Skip to content
Back
AI · thermal imaging

Thermal Super-Resolution

Makes low-resolution thermal camera images sharper, fast enough to run on the device beside the sensor.

My partComputer vision engineer

GitHub — thermal-super-resolution ↗Read the write-up →

In short

An IMDN-derived single-channel thermal super-resolution pipeline trained with thermal-specific objectives and optimized for deployment on NVIDIA Jetson hardware.

What it is
A single-channel super-resolution pipeline that adapts an Information Multi-Distillation Network to thermal imagery for deployment near the sensor.
Why it matters
RGB-trained super-resolution models can hallucinate texture that has no thermal meaning, and higher-resolution thermal sensors are expensive relative to lower-resolution alternatives.
My part
I worked on the architecture adaptation, the training, and the edge inference.

Step by step

  1. 1

    Adapt

    Information Multi-Distillation Network layers are adapted from three-channel RGB to a single thermal channel to match sensor semantics.

  2. 2

    Transfer

    Weights pretrained on RGB provide a useful starting point in a domain where thermal training data is scarce.

  3. 3

    Retrain

    A thermal-specific objective moves the model toward gradients, contrast, and structure that belong to heat imagery, rather than photographic texture priors.

  4. 4

    Deploy

    Inference is moved toward FP16 and INT8 execution and measured as a complete pipeline on named NVIDIA Jetson hardware, not as a bare model file.

Results

34.2 dB / 0.840
×2 quality
PSNR / SSIM
31.0 dB / 0.757
×3 quality
PSNR / SSIM
29.6 dB / 0.713
×4 quality
PSNR / SSIM
~45 FPS
Edge inference
NVIDIA Jetson Orin, after quantization

From the project

Side-by-side comparison of a low-resolution thermal frame and its ×2 super-resolved output.
×2 super-resolution — a low-resolution thermal frame beside the reconstructed output on the same scene.
Low-resolution thermal frame and its ×3 super-resolved reconstruction placed side by side.
×3 comparison — the task becomes less constrained as scale grows.
For engineersHow it is built, its limits, and its keywords.

How it works

The IMDN architecture is adapted to a single channel, pretraining is transferred from RGB data, a thermal-specific training objective shifts attention toward heat-relevant gradients and contrast, and the model is optimized toward FP16 and INT8 execution on NVIDIA Jetson hardware.

Limits

  • Reported quality numbers are dataset-specific and become less constrained at larger enlargement factors.
  • “Real time” is a property of a full pipeline on named hardware at a named input size, not of a model file. The ~45 FPS figure is on NVIDIA Jetson Orin after quantization; that Jetson benchmark is not in the public repository yet.
  • Speed figures measured elsewhere, such as the desktop-GPU run in the repository’s evaluation report, used different hardware and protocols, so I keep them apart from the Jetson figure.
  • The next real test is task-based: does downstream detection actually improve?

Keywords

  • thermal super-resolution
  • IMDN
  • Jetson Orin
  • edge inference
  • single-channel super-resolution

Read the write-up →

BackAll projects →