DeepIR

Deep InfraRed image processing framework

Synopsis

Thermal images captured by uncooled microbolometer cameras suffer from non-uniform noise that leads to severe loss in spatial resolution. We identify that thermal images can be factored into a scene-dependent flux that is concisely modeled as outputs of deep networks, and a scene-independent and slowly varying non-uniformity. By capturing multiple images with small camera motion, we are able to estimate the high quality thermal image.

Paper

Coming soon

Requirements

See requirements.txt.

Usage

If you want to run on simulated images, we have included two Boson images in the data/ folder.
If you want to run on a sequence of real images, please download the data from the link given in the next section and place the folders boson and lepton in data/ folder.
Once you downloaded the data, open demo.py and change the relevant variables -- all information should be available as comments. For example, if you want to run a simulation, the set of variables should be:

imname = 'test1'
camera = 'sim'
scale_sr = 1        # Set scale to 1 for denoising, higher integers
                    # for super resolution
nimg = 5            # Number of input images. For denoising, 3 - 5 suffice,
                    # but for super resolution, you may need more.

We have provided default configuration files in configs/, you may edit them as you see fit.

Real data

Download image sequences from real camera here

The data folder contains mat files from a Boson camera (640x512) and a Lepton 3.5 camera (160x120). Each image sequence was captured with small amounts of camera motion that can be modeled as affine transformation.

Citation

Vishwanath Saragadam, Akshat Dave, Ashok Veeraraghavan, and Richard G. Baraniuk, "Thermal Image Processing via Physics-Inspired Deep Networks", IEEE Intl. Conf. Computer Vision Workshop on Learning for Computational Imaging, 2021.

Acknowledgements

We thank the authors of Deep Image Prior for sharing their code. We repurposed some of their code for DeepIR.

UkcheolShin/DeepIR