ISET Simulations


(Difference between revisions)
Jump to: navigation, search
Rjpatruno (Talk | contribs)
(Created page with "See [ Farrell et al 2006] for an ISET simulation designed to evaluate image quality trade...")

Latest revision as of 13:00, 10 August 2015

See Farrell et al 2006 for an ISET simulation designed to evaluate image quality tradeoffs between resolution and sensitivity.

The study we document below is an extension of this previous work. In this study, we use multispectral images of faces as the scene data for the ISET simulations. We use ISET to create and process sensor images. We use the Psych Toolbox to present images in a pairwise comparison task. We use the Display Simulation Toolbox to predict the displayed radiance of the images as they are rendered on a calibrated LCD display in our laboratory. And we use the S-CIELAB color difference metric to predict the visibility of uncorrelated image noise.

Joyce wrote a draft of the paper and Jiajing will present it at SPIE 2009

Comments about this study

  • In this study, we investigate the effect that pixel size, read noise and scene luminance have upon perceived image quality. We see that for most sensors, image quality is limited by read noise at low light levels and photon noise at higher light levels. If read noise is zero, then image quality is determined by shot noise. The contributions of dark voltage, dsnu and prnu are typically negligible.
  • The simulations will keep die size fixed. So decreasing pixel size means that we will increase the number of pixels. We will downsample the images with more pixels so that they are the same size as the image captured with the fewest number of pixels. We will keep exposure duration fixed.
  • Under these conditions, why does read noise have a bigger effect on image quality when both pixel size and scene luminance decrease?
    • The reason that pixel size has a bigger effect when read noise is high is because smaller pixels capture fewer photons, generate fewer electrons, and the ratio of captured electrons to read noise (SNR) is lower As pixel size increases, more photons generate more electrons and the ratio of captured electrons to read noise (SNR) is higher.

Downsampling increases sensor SNR, so the decrease in SNR with scene luminance, pixel size and read noise will be somewhat mitigated. There will be some level of read noise for which downsampling does not completed offset the effects of pixel size, scene luminance and read noise on sensor SNR. The impact of read noise, scene luminance and pixel size depends on how much sensor SNR is increased with downsampling.

  • Will image quality be affected by pixel size when read noise is zero?
    • SNR decreases with decreasing pixel size, but increases with downsampling. So, the effect that pixel size has upon image quality depends on the effect that downsampling has on SNR.
    • Read noise has a bigger effect on SNR than pixel size at low light levels. So when we delete read noise, SNR is already higher. Downsampling may then completely offset the effect that pixel size has on SNR.

We need a way to calculate SNR after downsampling

Plot SNR versus lux-sec. Note the lux-sec at which SNR transitions from being determined by read noise to photon noise.

Personal tools