SSIM, PSNR, MSE and ΔE00: How to Compare Two Images and Interpret Each Metric
What each image comparison metric measures, how it is calculated, which values are good, and which one to use for quality control, sample degradation or microscopy, with numerical examples.
Saying that two images “look alike” is not enough for a quality report or a paper. Several metrics exist to quantify it, but each measures something different: MSE and PSNR measure pixel-by-pixel intensity difference, SSIM structural similarity as perceived by the eye, ΔE00 colour difference, and histogram correlation the tonal distribution. This guide explains how to interpret them and when to use each one.
Before comparing: requirements
No metric can fix a poor acquisition. For the comparison to be meaningful:
- Same size and alignment: a shift of a few pixels tanks almost every metric even if the content is identical.
- Same illumination, exposure and white balance: otherwise you are measuring the camera, not the sample.
- Lossless format (PNG or TIFF) whenever possible: JPEG compression adds artefacts that count as differences.
- Same scale in microscopy (same objective and resolution).
MSE: mean squared error
MSE = (1 / N) · Σ (Iₐ − I_b)²The average of the squared intensity differences between corresponding pixels.
- 0 = identical images. The higher, the more different.
- Its units are intensity² and its value depends on bit depth (8 or 16 bits), so it is not comparable across different datasets.
- It is very sensitive to global brightness changes, even when the image “looks” the same.
PSNR: peak signal-to-noise ratio
PSNR = 10 · log₁₀(MAX² / MSE) [dB]where MAX is the maximum possible pixel value: 255 for 8-bit images and 65,535 for 16-bit images. It is MSE expressed on a more intuitive logarithmic scale.
| PSNR (8-bit) | Rough interpretation |
|---|---|
| > 40 dB | Differences practically invisible |
| 30–40 dB | Good similarity; subtle differences |
| 20–30 dB | Visible differences |
| < 20 dB | Clearly different images |
If the images are identical, MSE = 0 and PSNR is infinite.
Example: brighten an entire 8-bit image by 5 grey levels. MSE = 5² = 25 and PSNR = 10 · log₁₀(255² / 25) ≈ 34.2 dB — “subtle differences” according to the table, even though the structure has not changed at all. That is why PSNR alone is not enough.
SSIM: structural similarity index
Proposed by Wang and colleagues (2004), it compares luminance, contrast and structure in local windows (typically 11 × 11 pixels) and averages the result. It is designed to resemble how we perceive differences.
- Range up to 1 (identical). In practice, values between 0 and 1.
- Besides the mean value, the SSIM map shows where the differences are.
| SSIM | Rough interpretation |
|---|---|
| > 0.95 | Very similar |
| 0.80–0.95 | Noticeable differences |
| < 0.80 | Major differences |
Back to the 5-level brightening example: the luminance term at a mid-grey level of 128 is ≈ 0.999 and contrast and structure are unchanged, so SSIM is practically 1. By contrast, a slight blur can give a similar PSNR but a clearly lower SSIM, because it destroys structure. Using both together tells you what kind of difference there is.
ΔE00: CIEDE2000 colour difference
ΔE00 measures the colour difference between pixels in the CIELAB space, which is designed to be perceptually uniform. The CIEDE2000 formula corrects the uniformity deviations of earlier formulas (ΔE76, ΔE94).
| ΔE00 | Rough interpretation |
|---|---|
| < 1 | Not perceptible to the human eye |
| 1–2 | Perceptible only on close observation |
| 2–3.5 | Perceptible at a glance |
| > 5 | Clearly different colours |
Acceptable tolerances depend on the industry (printing, paints, textiles, food). Note that converting RGB to CIELAB assumes a colour space (usually sRGB with a D65 illuminant): for reliable colour measurements the camera must be calibrated and the lighting controlled.
It is the right metric for colour changes: yellowing of a polymer, fading after UV ageing, uniformity of a coating or staining.
Histogram correlation
Compares the intensity distribution of the two images. It equals 1 when the distributions are identical.
- Useful for detecting changes in exposure, contrast or tone.
- It ignores pixel position: if you shuffled all the pixels of an image, its correlation with the original would still be 1. Never use it on its own to claim two images are the same.
Which metric should you use?
| Goal | Recommended metrics |
|---|---|
| Check that compression or processing does not degrade an image | PSNR + SSIM |
| Colour change from ageing, degradation or staining | ΔE00 (mean and regions with the highest ΔE) |
| Morphological changes in microscopy | SSIM map + ROI selection |
| Exposure control between sessions | Histogram correlation |
| Visual inspection of where the difference is | Pixel-by-pixel difference map |
Common mistakes
- Comparing misaligned images or images of different sizes rescaled without control.
- Trusting a single global number: a small, localised difference (a defect, a crack) barely moves the mean. Use heatmaps and regions of interest (ROI).
- Comparing PSNR or MSE values across images with different bit depth or content.
- Measuring colour with uncalibrated cameras or changing lighting.
- Using JPEG for images you will compare quantitatively.
MIND4SCE-Image calculates MSE, PSNR, SSIM, ΔE00, histogram correlation and pixel differences, with heatmaps, synchronised views and ROI selection, plus bioimaging modes such as cell viability. Everything is processed in the browser: your images never leave your computer.