Comparison of Methods for Estimating the Conversion Gain of CMOS Active Pixel Sensors

The use of complimentary metal-oxide-semiconductor (CMOS) active pixel sensors (APS) in scientific applications is becoming increasingly widespread and as a result, attempts to quantify their absolute performance via the conversion gain parameter K(epsiv_/DN) are now common. The inherent non- linearity of CMOS APS response means that the gold standard "mean-variance" and "photon-transfer" methods for establishing sensor conversion gain cannot automatically be used. Recently, two alternative methods have been proposed for deriving the conversion gain of APS. These methods appear to produce good results but have yet to be directly compared using results from a single sensor. This paper uses a single data set from a prototype APS to provide an absolute comparison of the performance of these nonlinear methods with respect to each other, and to standard linear methods. It also proposes a measurement procedure to enable evaluation of any APS. The prototype APS "Vanilla" produced by the RC-UK MI-3 consortium was illuminated uniformly at 65 intensity levels between dark and saturation. Results for Vanilla from linear methods substantially overestimate the sensor full well capacity (p = 0.003) to be (94.1 plusmn 1.4) X 10³e^- and (91.2 plusmn 1.4) x 10³e^-, respectively, when compared to nonlinear analysis methods. They also slightly underestimate the sensor read noise (p = 0.016) to be 51.8 plusmn 0.7e_ and 52.3 plusmn 0.7e~, respectively. The results obtained from both nonlinear estimation (NLE) and nonlinear compensation (NLC) are in good agreement, predicting read noise of54.8plusmn0.3e^- and 55.4 plusmn 0.2e^- and full well capacity of (72.0 plusmn 0.3) x 10³e^- and (72.6 plusmn 0.3) X 10³e^-, respectively. Results indicate that nonlinear analysis should be performed for APSs.