PSNR, SSIM and Your Eyes: Comparing Image Quality

Why measure at all?

When you tune an encoder, test a scaler or compare two denoising settings, you need a way to rank results that does not depend on how tired you are. Objective metrics provide that — as long as you know what they actually measure.

PSNR: counting the error

PSNR compares each pixel of a processed image with the reference and summarises the average squared difference on a logarithmic scale. Higher is better. For 8-bit video, values in the high 30s to 40s dB usually indicate good quality; below about 30 dB, differences are often visible.

Its strength is simplicity. Its weakness is that every pixel error counts equally, whether it sits in a smooth sky where you would notice it or in dense foliage where you would not.

SSIM: comparing structure

SSIM looks at small windows and compares brightness, contrast and structure. It tends to track perceived quality better than PSNR, especially for blur and ringing. Values close to 1 indicate near-identical images.

Common mistakes

• Misaligned frames. A one-frame or one-pixel offset collapses both scores. Always align first.
• Mismatched ranges or matrices. Comparing limited-range with full-range data produces large, meaningless errors.
• Averaging hides problems. A sequence with one badly damaged frame can still report a high mean. Look at per-frame values.

Where your eyes come in

Metrics tell you how much differs; visual inspection tells you whether it matters. A practical workflow:

1. Use metrics to shortlist candidates and flag the worst frames.
2. View those frames side by side and as a difference image.
3. Watch the sequence at real speed, since some artefacts only appear in motion.

The takeaway

Treat PSNR and SSIM as instruments, not verdicts. Used together with careful viewing, they turn subjective arguments about quality into informed decisions.