SyS-AE — Symmetric Structural Auto-Encoder

Reference implementation for the paper:

D. Staps, M. Kaden, J. Auth, F. Zaussinger, T. Villmann, "Compression of Particle Images for Inspection of Microgravity Experiments by Means of a Symmetric Structural Auto-Encoder", 2023 13th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS), IEEE, 2023, pp. 1–5. DOI: 10.1109/WHISPERS61460.2023.10431286

What this is

An auto-encoder for lossy image compression that preserves perceptual/structural properties, designed for low-power, resource-restricted encoding (motivated by microgravity plasma experiments — PK4/COMPACT — on the ISS, where images must be compressed on-board and inspected visually by experts on earth).

The Symmetric Structural Auto-Encoder (SyS-AE) encodes with an element-wise invertible non-linearity φ_θ followed by a linear down-projection A:

encoder   Φ(x) = A · φ_θ(x)
decoder   Ψ(y) = φ_θ⁻¹(A⁺ · y)      # A⁺ = Penrose pseudo-inverse of A

Only one side is trained by stochastic gradient descent; the counterpart follows in closed form (pseudo-inverse of A, analytic inverse of φ_θ). This keeps encoding cheap and makes the model structurally symmetric and interpretable. φ_θ can be the identity (→ linear AE) or a scaled/shifted sigmoid/arctan; the sigmoidal choice doubles as visual contrast enhancement.

The training loss is the structural (dis)similarity derived from SSIM (DSSIM = (1 − SSIM)/2), compared against a plain MSE loss and the classic Eigenimages baseline.

Datasets

• PK4 — 70 000 grayscale particle images, 11×11 px (55 000 train / 5 000 val / 10 000 test), gray values normalized to [−1, 1].
• MNIST and Fashion-MNIST as illustrative benchmarks.

All images are vectorized (concatenation) before encoding.

Repository layout

Path	Purpose
sisiae.py	SyS-AE model (encoder/decoder, Penrose inverse, φ and φ⁻¹)
main_sisiae.py	main training/experiment entry point (PK4)
main_sisiae_fashion.py, main_EigenImages.py, main_jpeg.py	Fashion-MNIST / Eigenimages baseline / JPEG comparison
utils/	data.py (loading), measures.py (SSIM/DSSIM), plots.py, names.py
paper_images/, test_results/, dataframes/	as-published figures, metric dumps and result tables
literature/	cited references (PDFs)

Reproduce

pip install -r requirements.txt
python main_sisiae.py            # PK4 experiment
python main_sisiae_fashion.py    # Fashion-MNIST
python main_EigenImages.py       # Eigenimages baseline

The exact state used for the paper is tagged v-paper-2023a:

git checkout v-paper-2023a

Results (paper)

The SSIM-trained non-linear SyS-AE gives the lowest DSSIM across all dimensionality-reduction ratios and all three datasets, and outperforms the Eigenimages baseline in visual inspection — reconstruction error is spread noise-like over the image rather than concentrated on the particle. See paper_images/ and Fig. 1 / Fig. 2 of the paper.

How to cite

If you use this code, please cite the paper above. A machine-readable entry is in CITATION.cff.

BibTeX:

@inproceedings{staps2023sysae,
  title        = {Compression of Particle Images for Inspection of Microgravity Experiments by Means of a Symmetric Structural Auto-Encoder},
  author       = {Staps, Daniel and Kaden, Marika and Auth, Jan and Zaussinger, Florian and Villmann, Thomas},
  booktitle    = {2023 13th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS)},
  pages        = {1--5}, year = {2023}, organization = {IEEE},
  doi          = {10.1109/WHISPERS61460.2023.10431286}
}

Acknowledgment

Developed by Daniel Staps (0009-0002-4459-4544) at the Saxon Institute of Computational Intelligence and Machine Learning (SICIM), University of Applied Sciences Mittweida. Funded by the German Aerospace Center (AIMS/DAIMLER 50WK2270A, AIMS/IAI-XPRESS 50WK2270D, AID 50WM2163/50WM2354) and the Federal Ministry for Economic Affairs and Climate Action (AI4OD 19A20023D). Repository cleanup and documentation carried out with assistance from Claude (Anthropic).

License

MIT — see LICENSE.