Realistic Noise Model - Githubissues

WangLiwen1994 commented 4 years ago

Hi~

When I use the function "realistic_noise()", there is a problem of divided by 0.

M=np.sqrt( ((Ag*Dg)/(Nsat*image)+(Dg**2)*((Ag * CT1 + CT2)**2)))

How to calculate it if there is a pixel with value 0?

luixiao0 commented 4 years ago

add 1*10^-16 on every pixel may help

dahanhan commented 4 years ago

Which value range is acceptible for Ag and Dg? And where is the origination of this realistic noise model?

clausmichele commented 4 years ago

Please have a look at the paper, at page 3: This specific noise model, in equation1, is composed by two main contributions, the Photon Shot Noise (PSN) and the Read Noise. The PSN is the main noise source in low-light condition, where Nsat accounts the saturation number of electrons. The Read Noise is mainly due to the quantization process in the Analog to Digital Converter (ADC), used to transform the analog light signal into a digital image. CT1n represents the normalized value of the noise contribution due to the Analog Gain, whereas CT2n represents the additive normalized part. The realistic noisemodel, where the relevant terms for the considered Sony sensor are: Ag (Analog Gain), in range [0,64], Dg (Digital Gain), in range [0,32] and s, the image that will be degraded. The remaining values are CT1n=1.25, CT2n=1.11 and Nsat=7489. The noisy image is generated by multiplying observations of a normal distribution N(0,1)with the same shape of the reference images, with the equation2.

bfreskura commented 4 years ago

@clausmichele

Thank you very much for your work!

I understand the intuition behind the Realistic Noise model, but I can't find the origin of this equation anywhere. Can you please provide some resources from where have you derived this equation?

For example, this paper also mentions the Realistic Noise model but the maths look different. Is this the same RNM as your RNM model?

clausmichele / ViDeNN

Realistic Noise Model #9