2/5/2025 | 9:30 AM - 10:30 AM | Grand Peninsula C

A simple model for characterizing and simulating dark current in CMOS sensors

Author(s)

Steve Wang | Omnivision Technologies, Inc.
Eiichi Funatsu | OmniVision Technologies Inc
Boyd Fowler | OmniVision Technologies Inc.

Abstract

A normal distribution is typically used for characterizing and modeling dark current in CMOS sensors.  Although it provides simplicity and speeds needed in real-time applications, it is usually not a very good representation of the real dark current characteristics observed in real devices.  The statistical distribution of CMOS sensor dark noise is typically right-skewed with a long tail, i.e. with more “hot” pixels than described in a normal distribution.  Furthermore, the spatial distribution in real devices typically exhibit a 1/f-like power spectrum instead of a flat spectrum from a simple Gaussian distributions model.  When simulating sensor images, for example generating images and videos for training and testing image processing algorithms, it is important to reproduce both characteristics accurately.  We propose a simple convolution-type algorithm using seed images with a log-normal distribution and randomized kernels that is able to reproduce both the correct statistical and spatial distributions, and can be easily matched to measured noise data.  It can also be parallelized with CUDA acceleration to support real-time execution.