Histogram Equalization

A photograph taken in poor light uses only part of the range it could. Its histogram bunches into one corner, and everything inside that bunch — the folds in a dark coat, the texture on a shaded wall, the writing on a faded page — is squeezed into a handful of tones your eye cannot separate. Histogram equalization spreads that bunch back out across the full range, so the differences that were already recorded become differences you can see.

What You Get

• Detail out of the shadows: the tones that were crushed together get room, and the texture inside them comes back.
• Contrast without a brightness jump: the brightness-preserving mode stretches each half of the histogram into its own range, so the picture gains separation and keeps its exposure.
• Colour left alone: the default mode changes only how light a pixel is, so hue and saturation come through the process unchanged.
• Readable scans: grayscale mode widens the gap between ink and paper, which is what makes faded text legible again.
• Numbers, not guesses: the before and after histograms, the transfer curve and the tonal statistics are all measured from your image.

Recover detail in dark and flat images with adaptive or global histogram equalization.

1. 1Add an image: Drop in a JPG, PNG, WebP, GIF or BMP, or browse for one.
2. 2Pick an algorithm: Adaptive (CLAHE) suits photographs; global is the textbook stretch; brightness-preserving keeps the overall exposure where it was.
3. 3Set tile size and clip limit: Smaller tiles work more locally; a lower clip limit holds noise down.
4. 4Compare and check the numbers: Drag the divider across the preview, and read the before and after histograms and the transfer curve.
5. 5Export: Save as PNG, JPEG or WebP.

Three algorithms, and the difference between them is how much of the picture each mapping is allowed to see.

• Global equalization builds one cumulative curve from the whole image and applies it everywhere. It is the definition of the technique and it is the most aggressive: a single bright window in an otherwise dark room dominates the histogram, so the room gets lifted and the window blows out.
• Adaptive equalization (CLAHE) cuts the picture into tiles, builds a curve for each one, and blends smoothly between the four surrounding curves at every pixel — so there are no visible tile seams. The clip limit caps how far any one tile is allowed to stretch, which is what stops flat regions like sky and paper turning into noise. This is the mode that lifts a shadow without touching the sky beside it.
• Brightness-preserving equalization (BBHE) splits the histogram at the image's own average tone and equalizes the two halves into separate output ranges. Both halves gain contrast, and because neither can cross the average, the picture's overall brightness stays where it was. Use it when a photo is already well exposed and only lacks separation.

The parameter names here are the ones the field uses, and they mean the same thing here as they do anywhere else.

• Tile size, in pixels the side of each adaptive tile, measured on the original image rather than as a count of divisions — so the same number means the same thing on a phone snapshot and a scanned poster. It can also be given as a percentage of the shorter side, which is the form to use when the same setting has to mean the same thing across images of different sizes. A tile should be larger than the detail you want to keep. The slider's own range moves with the image and with the bin count, because a tile smaller than its histogram cannot produce a smooth curve.
• Clip limit, 1 to 10 a multiple of the height a perfectly flat histogram would have. At 1 the mapping is the identity and nothing changes; 2 to 3 is the usual starting point for photographs; 5 and above pushes hard and brings grain up with the detail.
• Histogram bins, 16 to 256 how finely the tones are counted before the curve is built. Fewer bins gives a smoother, gentler curve that amplifies less noise — the standard brake on a grainy scan. 256 is the finest the format can carry.
• Strength blends the finished result back towards the original. Full equalization is often more than a picture needs; 60 to 80 percent usually looks natural.
• Smooth before equalizing runs a gentle blur before anything is measured, so it is the picture being equalized and not the grain.

Equalization computes a mapping for one quantity, and which quantity that is decides what happens to colour. The obvious shortcut — work out the new brightness and scale red, green and blue by the ratio between old and new — has two failure modes that show up on ordinary photographs: the darkest tone in the picture maps to zero, so its ratio is zero and every pixel at that tone becomes flat black, colour and all; and a saturated pixel that gets lifted has its strongest channel pushed past the maximum and clipped, which moves the hue.

This tool works in a perceptual colour space instead. Only the lightness is equalized; the two colour axes are carried through untouched, so hue is preserved by construction. Where a lifted colour ends up outside what a screen can show, its saturation is eased back along its own hue line until it fits — never clipped channel by channel. Measured on a saturated test image, that is a worst-case hue shift of 1.8 degrees against 19.7 for the clipping approach.

Three situations where equalization is the right tool rather than a brightness slider.

• Underexposed photographs Adaptive mode, tile around a tenth of the image's short edge, clip limit 2 to 3. The shadows open up and the correctly exposed parts of the frame are left near where they were.
• Scanned documents and faded text Grayscale mode with smoothing on. The gap between ink and paper widens, which is what makes faded writing readable and what character recognition needs.
• Microscopy, medical and technical images Adaptive mode with a small tile and a low clip limit. This is the setting the technique was designed for, and the reason it exists in every scientific imaging package.

Every accelerated path is checked against a plain reference implementation that has no hardware acceleration at all. The comparison runs across seven test images and fifteen parameter combinations on each processing path, and the largest difference on any colour channel is one step out of 255 — the smallest difference two ways of rounding the same number can produce. The histograms themselves match exactly. The tonal statistics beside your image are counted from every pixel, not sampled.

Hardware accelerated: a 12-megapixel photograph is equalized in about 60 milliseconds on a modern graphics card, against 2.5 seconds without acceleration. Memory use is flat across repeated runs.

• Automatic exposure, colour and detail correction in one press.

• Lift, gamma and gain wheels, curves and LUTs for deliberate looks.

• Manual brightness, contrast and saturation sliders.

• Exact histograms, palette and statistics for any image.

• Clean up grain before or after equalizing.

• Recover edge definition on a soft scan.