Description
Fixed pattern noise (FPN), or fixed-structure noise, is a regular spatial structure (grid, hatching, column pattern) inherent to the sensor, always present at the same location in the image.
It originates from small gain and offset variations between pixels (and between columns or rows) of the sensor: each photosite does not respond exactly like its neighbor, which draws a fixed pattern once the signal is stretched.
Unlike random noise, it does not average out by stacking undithered frames: being fixed, it is reinforced. Dithering and proper calibration are what neutralize it.
This is an electronic defect. Distinct from banding (broad, obvious lines), from walking noise (a fixed pattern translated into diagonal streaks by lack of dithering), and from a dark mismatch.
Visual signature
A regular geometric pattern (fine grid, hatching, column or row structure) overlays the image, most visible on the sky background after stretching.
Its essential characteristic is being fixed on the sensor: it occupies exactly the same position on every frame, regardless of how the sky tracks.
It does not diminish with simple stacking if the frames are not dithered: on the contrary, the pattern consolidates from frame to frame.
It becomes flagrant with aggressive stretching and on low-signal images, where the fixed pattern represents a significant fraction of the background content.
Differential diagnosis
Do not confuse with banding: banding draws broad, obvious bands aligned on one sensor axis; FPN is a fine, repetitive grid across the whole field.
Distinct from walking noise: walking noise is precisely a fixed pattern that, without dithering, is translated into diagonal streaks during stacking, whereas raw FPN stays stationary.
Separate from color noise: color noise is random and chromatic; FPN is structured and geometric.
Related to residual hot and cold pixels: same family of fixed sensor defects, which calibration and dithering correct together.
Probable causes
- Gain and offset variations between sensor pixels
- Column or row non-uniformity (sensor readout)
- No dithering between frames
- Incomplete calibration (missing or mismatched darks or flats)
- Aggressive stretching revealing the fixed pattern on low signal
- Budget sensor or unfavorable gain settings
Course of action
- Calibrate with darks and flats matched to the light frames
- Enable systematic dithering between every frame
- Stack with appropriate statistical rejection (sigma clipping)
- Accumulate enough dithered frames to dissolve the pattern
- Avoid excessive stretching on too-weak a signal
- Choose an operating point (gain, offset) recommended for the sensor
- Supplement if needed with a light denoising pass after calibration
The Doc's advice
FPN is your sensor's signature, and the good news is that it can be tamed reliably. Two weapons. Calibration first: matched darks and flats remove the bulk of fixed gain and offset variations. Dithering next, and this is the most important: by shifting the framing by a few pixels between each frame, the fixed pattern falls at different sky positions and the stacking rejection dissolves it. Without dithering, no number of frames will make FPN disappear: it will merely transform into walking noise. Calibrate properly, dither systematically, and FPN becomes a non-issue.
Think you can see this defect in your image?
Run a diagnosisFrequently asked questions
What is fixed pattern noise?
It is a fixed-structure noise intrinsic to the sensor, appearing as a regular pattern (grid, hatching, column or row structure) always at the same location in the image. It comes from small differences in gain and offset between photosites and between readout columns or rows: no pixel responds exactly like its neighbor. Unlike random noise, it is deterministic and fixed, which completely changes how to fight it: it must be decorrelated from the signal, through dithering.
How can fixed pattern noise be suppressed?
Two complementary levers. First, careful calibration: properly matched darks and flats remove most fixed gain and offset variations. Second, and most decisively, dithering: by slightly shifting the framing between each frame, the sensor's fixed pattern lands at different positions relative to the sky, so the statistical rejection of the stacking dissolves it. With calibration plus dithering plus a sufficient number of frames, FPN becomes invisible.
Why does FPN not disappear by stacking more frames?
Because it is fixed. Stacking reduces random noise because that noise changes with every frame and is averaged out. FPN always occupies the same pixels: stacking undithered frames only confirms the pattern rather than erasing it, and may even strengthen it by improving its signal-to-noise ratio. The only way to average it out is to decorrelate it from the sensor through dithering, so that it does not land at the same sky position from frame to frame.
FPN or walking noise: what is the relationship?
They are two sides of the same problem. Fixed pattern noise is the raw fixed pattern of the sensor (or a calibration residual). Walking noise appears when that fixed pattern is not dithered and stacking, combined with a slight tracking drift, regularly translates it: the pattern "walks" and forms diagonal streaks. In other words, walking noise is FPN that has moved due to lack of dithering. The remedy is the same in both cases: systematic dithering.