LIMODENet: restoring channel-degraded satellite imagery under a spiking constraint
Satellite images sent over a low-SNR DVB-S2X link at 1 dB arrive badly degraded. Every restorer here has about 0.75–0.78M parameters and was trained under the same protocol on the same data. Use this page to check two claims from the paper:
- LIMODENet restores better than the CNN-AE and the U-Net at matched parameter count. NAFNet-lite and Restormer-lite restore better still. They are included so you can see that.
- Only LIMODENet converts to a spiking network with zero blocked operations. NAFNet-lite and Restormer-lite each have more than 20.
Everything runs in your browser (ONNX Runtime Web, on the CPU). Each model is about 3 MB and is downloaded from the Hub the first time you use it. Numbers come from one seed (seed 0) on the 30 shipped images; the paper reports three seeds on the full test split, so expect small differences.
Input
Real 1 dB/q100 DVB-S2X test images, three per class, drawn with a fixed seed (not hand-picked).
Restorers
Outputs
Pick an image and press Restore.
Score every shipped sample
Runs the selected restorers on all 30 images. PSNR is pooled over all pixels, as in the paper; SSIM is averaged over images. Takes a minute or two in the browser.
Which operations block conversion to a spiking network?
This is a static audit of the model code. The result below was generated by demo/audit.py; it needs no
weights or data, so you can rerun it yourself with python demo/audit.py and check every line it reports.
Parameters, mult-adds and latency
Parameters and mult-adds (GMac, at 128×128) were counted from the PyTorch models with torch.utils.flop_counter,
attention matmuls included. Latency is measured now, in your browser (ONNX Runtime Web, CPU, batch 1, median of 5 runs). The spiking
model's numbers are T=8 dense passes on conventional hardware; this is not a neuromorphic-hardware measurement.
The spiking network at work
Spikes per layer, membrane potentials, and the restoration building up over T=8 timesteps, replayed from activity recorded from the released spiking checkpoint on real 1 dB test images. Press Play. Open it full-screen.