Skip to content

Peak Signal-to-Noise Ratio (PSNR)

Key: psnr · Family: signal fidelity · Better: higher ↑ · Per-lead: yes

PSNR asks: how strong is the signal compared with the error? It turns the mean squared error into decibels relative to the real signal's peak-to-peak range, so a value of 30 dB means the error power is 1000 times smaller than the squared range.

Try it

Interactive demo loading… (needs JavaScript)

What to notice:

  • Identical beats give ∞ dB (zero error).
  • Every halving of the RMS error adds about 6 dB. The scale is logarithmic, so the first bit of noise costs many decibels and further noise costs fewer.
  • 20 ms late drops PSNR sharply, as with every point-wise metric.

How it works

  1. For each real recording, find the nearest synthetic recording (smallest Euclidean distance over all leads).
  2. Per lead, compute the MSE between the pair and the peak-to-peak range of the real signal.
  3. Convert to decibels and average; the score is the mean over leads.

The math

\[ \mathrm{PSNR} = 10\,\log_{10}\frac{\mathrm{MAX}^2}{\mathrm{MSE}}, \qquad \mathrm{MAX} = \max_t x_r[t] - \min_t x_r[t], \qquad \mathrm{MSE} = \frac{1}{T}\sum_t (x_r[t]-x_s[t])^2 \]

Reading the score

  • In dB, higher is better, unbounded above. Because it is normalised by the signal's own range, it is less sensitive to absolute units than MAE.
  • A single number on a log scale hides where the error is; inspect waveforms too.
  • Same caveats as other point-wise metrics: timing and baseline offsets dominate.

Use it with ECGEN-Eval

ecgen-eval eval --real real.npy --synthetic synth.npy -m psnr --output report.html

from ecgen_eval import load_ecg_npy, AVAILABLE_METRICS

real = load_ecg_npy("real.npy", fs=500)
synth = load_ecg_npy("synth.npy", fs=500)

result = AVAILABLE_METRICS["psnr"]().compute(real.data, synth.data, lead_names=real.lead_names)
print(f"PSNR = {result.score:.1f} dB")

Parameters are in the Metric Reference.

Learn more

  • Huynh-Thu & Ghanbari, Scope of validity of PSNR in image/video quality assessment, Electronics Letters 2008.