Analog noise estimator CLI

Command line interface and results for analog-noise-estimator module for estimate analog noise on the images.

You can estimate noise in whole image (as in original idea) or estimate noise in small boxes and next merge estimations into single result using provided rating functions.

Includes tools for display estimation results, calculate metrics and plot Gauss rating function.

Results

Used Laplacian 3x3 kernel and Numpy implementation of convolve, box 25x25 pixels.

Sigma is normal distribution sigma parameter.
In "average local sigma clip" sigma is clip on value 30.

Metrics present noisy part of image.

"Global sigma" is calculated from whole image without splitting into boxes.
"Global rating" present noisy part of image, but it is calculated from global sigma.

Good square

Global sigma: 4.630285306260067
Average local sigma: 5.779785867841117
Average local sigma clip: 5.779785867841117
Global rating: 0
Constant metric: 0.04643962848297214
Linear metric: 0.02700538949831164
Gauss metric: 0.019763315201960927

Noise square

Global sigma: 46.1618466613965
Average local sigma: 56.172134778404356
Average local sigma clip: 30.0
Global rating: 1
Constant metric: 1.0
Linear metric: 1.0
Gauss metric: 1.0

NOAA control area

Global sigma: 3.0373033731974606
Average local sigma: 4.357952733368098
Average local sigma clip: 4.357952733368098
Global rating: 0
Constant metric: 0.022727272727272728
Linear metric: 0.010603128772758869
Gauss metric: 0.012523404874079712

APT - bad quality

Global sigma: 42.14491795131236
Average local sigma: 50.41452993959203
Average local sigma clip: 29.175048508970093
Global rating: 1
Constant metric: 0.9773425499231951
Linear metric: 0.9762678452411343
Gauss metric: 0.9758709874331352

APT - medium quality

Global sigma: 17.020939720120662
Average local sigma: 21.43077214287576
Average local sigma clip: 15.831956487119431
Global rating: 0.8918800746691967
Constant metric: 0.45802005012531327
Linear metric: 0.430505130732554
Gauss metric: 0.4216867574113292

How to use

1. Install Python dependencies

pip install -r requirements.txt

2. Run command line interface

python cli.py

3. Import in your code (quick path)

import cv2
import noise_estimator

img = cv2.imread("data/product1.png", cv2.IMREAD_GRAYSCALE)

noise = noise_estimator.estimate(img)

print("Noisy / all", noise)

4. Look at the cli.py code for more complex usage example

Gauss rating

This package contains function for rate noise. It returns 0 for good quality image and 1 for noisy. Rating is based on detected sigma value.

We assume that sigma less then 10 means clear image (rate 0). Different sensors may to produce different average sigma. But empirical analysis has shown that good quality images have sigma 10 or less.

Sigma greater then 20 means that image is bad (rate 1).

Values between 10 and 20 are problematic. 11-12 are quite code, 17-19 are quite bad, 13-16 are medium, but the relationship does not seem to be linear. We use cumulative (normalized) normal distribution function for determining this value.

Below shown plot with metric values for different sigma.

References

J. Immerkær, “Fast Noise Variance Estimation”, Computer Vision and Image Understanding, Vol. 64, No. 2, pp. 300-302, Sep. 1996