Meritcs 日本語

Background

What Is SSIM?

SSIM (the Structural Similarity Index) is the metric at the center of everything Meritcs publishes. Before getting into our own findings, it's worth explaining plainly what it is, where it came from, and how it's normally used.

Who developed it

SSIM was introduced in 2004 by Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli, in a paper titled "Image Quality Assessment: From Error Visibility to Structural Similarity", published in IEEE Transactions on Image Processing, volume 13, issue 4, pages 600–612. It has since become one of the most widely cited and widely used image quality metrics in the field.

What it measures

Earlier metrics like PSNR (Peak Signal-to-Noise Ratio) simply measured the average pixel-by-pixel difference between two images. SSIM's core idea was different: instead of comparing pixels independently, compare local statistics that reflect how the human visual system perceives structure. It combines three separate comparisons, multiplied together:

L

Luminance

Are the two regions similarly bright?

C

Contrast

Do they have similar contrast — similar spread of pixel values?

S

Structure

Independent of brightness and contrast, do the patterns of variation correlate?

The final SSIM score for a region is the product of these three terms:

SSIM = L · C · S
Ranges from −1 to 1, with 1 meaning the two regions are identical.

Where it's used

SSIM is used anywhere the quality of a reconstructed, compressed, or generated image needs to be scored against an original:

It's also one of the standard benchmarking metrics reported alongside PSNR in computer vision and image processing research generally.

How it's typically computed — and why that matters

In practice, SSIM is rarely computed over a whole image at once. It's computed inside a small sliding window (commonly 8×8 or 11×11 pixels) that moves across the image, producing one local score per window position, and those local scores are then averaged into a single number for the whole image. That final average is almost always the only number anyone reports or compares.

Meritcs's own reports compute SSIM over 8×8 non-overlapping blocks (a fixed grid partition of the image) rather than a sliding window — see Technology for why.

How the windowed SSIM score becomes one final number An image is divided into a grid of small windows. Each window gets its own local score; most score well, but one scores badly. All the local scores are averaged into a single final number, which reads as if everything scored well. Local windows across one image one window scores badly average one number 0.94

The one bad window is still in there — it's just outvoted by the good ones, and the final average can't tell you it happened.

That averaging step is exactly where Meritcs' own research begins. A single average, by construction, cannot distinguish between an image that scored well everywhere and an image that scored very well in most places but badly failed in a few — the two can produce the identical number. Our Approach page explains how we investigate that gap, and our Publications page collects what we've found so far — including a pair of tools that flag this kind of hidden, badly-scoring window without needing the original image to compare against.