CMOS Image Sensor Design_h20_8_characterization2.pdf
- 1. IN5350 – CMOS Image Sensor Design Lecture 8 – Characterization
- 2. Project schedule Task/milestone Start Finish Chose topic/scope 1-Sep 8-Sep Create project plan (tasks, milestones, schedule) 8-Sep 15-Sep MS1 – project plan approved by Johannes 15-Sep 22-Sep Study literature on the topic 22-Sep 29-Sep Design/simulation 29-Sep 13-Oct Write up prelim report (inc references, design, results) 13-Oct 20-Oct MS2 – submit preliminary report to Johannes 20-Oct 20-Oct Design/simulation 20-Oct 27-Oct Write up final report (incl references, design, results) 27-Oct 3-Nov MS3 – submit final report to Johannes & presentation 3-Nov 3-Nov MS4 – grading (pass/fail) by Johannes & Tohid 10-Nov 10-Nov Exam 18-Nov 2020 06.10.2020 2 2 ✅ ✅ ✅ ✅
- 3. Contents • Measurement tools • QE – Quantum efficiency • CG – Conversion gain • RN - Readnoise • DC – Dark current • DSNU – Dark signal non-uniformity • PRNU – Photoresponse non-uniformity • VFPN – Vertical fixed pattern noise • HFPN – Horizontal fixed pattern noise 06.10.2020 3
- 4. Recommended reading • EMVA 1288 Standard for image sensor characterization – https://www.emva.org/wp-content/uploads/EMVA1288-3.1a.pdf • SPIE book by Jim Janesick: Photon Transfer – Available at UiO • PhD thesis on CIS characterization – file:///Users/eier/Downloads/Utsav_jain_thesis_report.pdf • SPIE paper on Raspberry Pi based camera characterization – https://www.spiedigitallibrary.org/journalArticle/Download?fullD OI=10.1117%2F1.JEI.26.1.013014 06.10.2020 4
- 5. Integrating sphere • Illuminating a sensor uniformly on all pixels – Can be combined with a filter to select specific wavelength(s) 06.10.2020 5
- 6. Optoliner 06.10.2020 6 • Integrating sphere + filters + optics • Illumination w/test pattern
- 7. Light box with Macbeth colour checker chart • Illumination w/colour temperature settings • Ref: RGB coordinates of Macbeth colours 06.10.2020 7
- 8. Light spectrum from a blackbody is determined by its body temperature 8 ( ) 1 1 2 5 2 − ⋅ = kT hc e hc T B λ λ λ B λ(T)=spectral energy (J/(s sr m3)) h=Planck’s constant (6,6 x 10-34 Js) λ=wavelength (m) c=speed of light (3x108 m/s) k=Boltzmann’s constant (1,38x10-23 J/K) Planck’s radiation law: ( ) T Bλ More details: J Nakamura, Appendix-1.
- 9. Monochromator • Purpose: Narrowband wavelength selection • Used for: Spectral response measurement (QE) 06.10.2020 9 Monochromator principles: (i) prism, (ii) grating Reference detector DUT
- 10. Quantum efficiency • Definition: Probability of a pixel detecting a photon of a given wavelength (spectral sensitivity) • Method: With a monochromator, step through each wavelength and measure average output pixel value, calculate back to #electrons captured, then divide by #photons incident on the pixel to get QE value 06.10.2020 10
- 11. QE remarks • QE influenced by angle of incidence – Wide angle => more crosstalk to neighbour pixels – Using small array in the centre minimizes crosstalk • Interference oscillations from optical stack • Latest innovation to boost QE in NIR • Ref: Scientific Reports, Vol. 7, Article No. 3832, June’17 06.10.2020 11
- 12. Photon Transfer Curve (PTC) • PTC: plot of variance vs mean (in DNs, Volts, or electrons) 06.10.2020 12 ... 2Nbit-1 0 0 Mean pixel value, (DN) 𝜎𝜎 𝑡𝑡𝑡𝑡𝑡𝑡 2 (DN 2 ) Readnoise Slope = CG (in DN/e-) Convert from DN/e- to uV/e- by multiplying with: 𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉/2𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁 � 1/𝐴𝐴𝐴𝐴𝐴𝐴 Curve bends due to clipping near saturation
- 13. Photon Transfer Curve 06.10.2020 13 𝑆𝑆𝑜𝑜𝑜𝑜𝑜𝑜 = 𝛼𝛼 � 𝑁𝑁𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒 + β , where 𝑆𝑆𝑜𝑜𝑜𝑜𝑜𝑜= output pixel value (DN) 𝛼𝛼 = conversion factor (DN/e-) 𝑁𝑁𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒= number of electrons captured (e-) 𝛽𝛽 = black level offset (DN) Let (1) From (1), the noise output variance (𝜎𝜎𝑜𝑜𝑜𝑜𝑜𝑜 2 ) can be expressed by 𝜎𝜎𝑜𝑜𝑜𝑜𝑜𝑜 2 = 𝜎𝜎𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒 2 + 𝜎𝜎𝑅𝑅𝑅𝑅 2 = 𝛼𝛼2𝑁𝑁𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒 + 𝜎𝜎𝑅𝑅𝑅𝑅 2 (2) , where 𝜎𝜎𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒 2 = electron shot noise (DN2) 𝜎𝜎𝑅𝑅𝑅𝑅 2 = readnoise floor at zero illumination (DN2)
- 14. Photon Transfer Curve (cont.) 06.10.2020 14 𝑆𝑆𝑜𝑜𝑜𝑜𝑜𝑜 − β = 𝛼𝛼 � 𝑁𝑁𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒 Replacing into (2) gives From (1) we have (3) 𝜎𝜎𝑜𝑜𝑜𝑜𝑜𝑜 2 = 𝛼𝛼 𝑆𝑆𝑜𝑜𝑜𝑜𝑜𝑜 − β + 𝜎𝜎𝑅𝑅𝑅𝑅 2 (4) Deriving (4) with respect to Sout gives 𝑑𝑑𝜎𝜎𝑜𝑜𝑜𝑜𝑜𝑜 2 𝑑𝑑𝑆𝑆𝑜𝑜𝑜𝑜𝑜𝑜 = 𝛼𝛼 (5) Conclusion: Conversion gain (in DN/e-) is slope of output variance versus mean curve. Can be referred back to FD node by dividing by ADC gain and SF gain. (DN/e-)
- 15. PTC remarks • Vary light level by changing integration time – High precision from crystal oscillator – Tint=0 gives readnoise value (RN) • To measure mean value (x-axis), prudent to capture dark frame (black level) for each setting • To measure variance (y-axis), convenient to use difference between two subsequent captures – Removes black level and fixed pattern noise – Remember to divide by 2 since noise variance is additive – Pixel values start to clip near saturation (measurement error) • For extra precision calculate variance for each pixel independently, i.e. by capturing 100-1000 pictures for each light level setting – Avoids the influence of PRNU at high signal levels 06.10.2020 15
- 16. PTC example 06.10.2020 16 1 10 100 1000 10000 100000 1000000 1 10 100 1.000 10.000 100.000 Noise variance (e- ^2) Mean signal (e-) RN (e- ^2) Shot noise (e- ^2) PRNU (e- rms) Total noise (e- rms)
- 17. Dark current (DC) measurement 06.10.2020 17 Sources of DC DC captures at constant temp. Example of bright pixels Source: Harvest Imaging blog μ, pix (DN) Tint (s) Slope=DC (DN/s or e-/s) @T=60C
- 18. Dark current measurement remarks • Ensure sensor has reached a stable temperature since DC is highly temperature dependent (DC doubles every 6-8℃) • Avoid light exposure (sometimes challenging) • Beware of DC shading. Could be due to non-uniform doping or contaminant levels and/or heat glow from circuits nearby (e.g. high power voltage driver) • Plot DC vs Tint. Slope gives DC in e-/sec. – DC can be calculated as average per frame OR individually for each pixel and then averaged – In the latter case, you can calculate the RMS variation of the DC. This is called the DSNU (dark signal non uniformity). – DSNU sometimes artificially large due to ’outliers’ in the DC distribution (ie bright pixels or black pixels). Could be filtered out. 06.10.2020 18
- 19. Photoresponse non-uniformity (PRNU) • As the name suggests, PRNU is a measure of the variation in responsivity from pixel to pixel • It’s definition varies in the literature, but the most common is as follows • Where σ50% is the rms variation of pixel mean values across the frame at 50% saturation (i.e. all the temporal noise is removed by averaging multiple frames, e.g. 100-1000 frames) 06.10.2020 19 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 = 𝜎𝜎50% 2 − 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷2 𝜇𝜇50% − 𝜇𝜇𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 (% rms)
- 20. PRNU remarks • Convenient to re-use data from PTC measurements (assuming 100-1000 frames were captured for each integration time) 06.10.2020 20 Stack of captures each with identical setting (values vary due to temporal noise, only) Remove temporal noise by calculating mean (μ) for each pixel position. Use it to calculate PRNU which is spatial and fixed noise.
- 21. • VFPN induced by col-to-col variations in Vth and parasitic couplings causing signal DC offset variations • Measured by (i) capturing dark frame, (ii) averaging all rows to form one single row without temporal noise, (iii) calculate rms value of row • VFPN value should be 10x smaller than RN to be invisible in image Vertical Fixed Pattern Noise (VFPN) 06.10.2020 21
- 22. Hnoise (rownoise) 06.10.2020 22 • Temporal noise on all pixels along one row induced by spikes on array signals or on VDD or GND during CDS readout (same noise spike on all pixels along one row) • Hnoise measured similarly to VFPN, ie average all columns to remove temporal noise, then calculate rms value