A Hgh-Senstvty Oversamplng Dgtal Sgnal Detecton Technque for CMOS Image Sensors Usng Non-destructve Intermedate Hgh-Speed Readout Mode Shoj Kawahto*, Nobuhro Kawa** and Yoshak Tadokoro** *Research Insttute of Electroncs, Shzuoka Unversty 3-5-1 Johoku, Hamamatsu 432-8011, Japan TEL:+81-53-478-1313, FAX:+81-53-412-5481, E-mal:kawahto@dl.re.shzuoka.ac.jp ** Toyohash Unversty of Technology, Abstract Ths paper dscusses the possblty of a hgh-senstvty dgtal sgnal detecton technque usng frame oversamplng and a CMOS mage sensor wth non-destructve hgh-speed ntermedate hgh-speed readout mode. Smulaton results show that the frame oversamplng s very effectve to reduce output amplfer nose and quantzaton nose of ADC. I. Introducton CMOS mage sensors are expected not only for the low-power smart mage magng devces but also for the hgh-senstvty magng devces. In the pxel, a bured photo dode enables ts random nose to be comparable to that of the CCD[1][2]. The mert of the CMOS mage sensor over the CCD on the amplfer nose s dscussed by Matsunaga et al [3]. Another possblty of the hgh-senstvty magng n the CMOS s the use of sgnal processng technque. A few technques usng oversamplng ADC for CMOS mage sensors have been proposed such as pxel-level A/D converson [4], and column parallel A/D converson usng delta-sgma modulaton [5]. Ths paper dscusses the possblty of another type of a oversamplng sgnal processng technque for ultra-low lght level CMOS mage sensors, that s frame oversamplng sgnal processng [6]. Ths technque allows us to use hgh-gan column parallel amplfer array and to reduce the nose of the output amplfer broadband nose and the quantzaton nose of the ADC. The necessary magng devce for ths sgnal processng s a CMOS mage sensor that the sgnal can be read out durng the sgnal charge accumulaton wthout destructng t. A smple three transstor APS as well as an APS specally desgned for non-destructve readout [7] can be used for ths purpose. In the followng, archtecture, prncple and smulaton results of the frame oversamplng sgnal processng usng non-destructve ntermedate magng are descrbed. Fg. 1 Archtecture of the frame oversamplng magng system. II. Archtecture and Prncple Fg. 1 shows the proposed scheme for the dgtal CMOS mage sensor system. The correspondng sgnal flow dagram s shown n Fg. 2. The mportant features are 1) each frame s read
out at M tmes hgher rate than the vdeo rate (f v ) wthout resettng the sgnal charge, 2) the read sensor sgnal s amplfed by the column amplfer wth the gan of M, 3) the dgtzed sgnal by the A/D converter s accumulated n the dgtal ntegrator, 4) the ntegrator output s converted to analog sgnal wth the gan of 1/M, 5) the analog output s subtracted from the read sensor sgnal n front of the column amplfer, and 6) a dgtal low-pass flter wth frame memory sampled at M x f v elmnates the nose component n the frequency range larger than the Nyqust frequency of the vdeo frame and the output s decmated to the vdeo rate. Fg.2 Sgnal processng dagram of the oversamplng CMOS mage sensor system. Fg. 3 Three transstor type actve pxel crcut. Ths scheme acts as a predcton coder n dfferental pulse code modulaton (DPCM). The most mportant advantage s that a large-gan column amplfer array can be used so that the nose level of the fnal amplfer s neglgble. Wthout ths predcton, the large gan cannot be used because the amplfer output wll saturate for the large nput sgnal. The quantzaton nose of the A/D converter s also reduced by ths scheme. The dgtal ntegrator and the dgtal flter ncrease the equvalent resoluton of A/D converter. If the nose s deally fltered, the theoretcal mprovement of SNR for the quantzaton nose s gven by approxmately 30 log 10 M [db] wth the frame-oversamplng rate of M x f v. For nstance, n the case of M=16 and 10bt ADC, the equvalent resoluton s 15bt. We can expect a very-wde lnear dynamc range. The actve pxel s ether smple 3-transstor type shown n Fg. 3 or n-pxel charge transfer type usng bured photo dode. In the case of 3-transstor type, the waveform of the cathode voltage of the photo dode and the readout tmng are shown n Fg. 4 n the case of M=4. After resettng, the reset level s read out, converted to dgtal, and memorzed. The ntermedate sgnal durng photo sgnal charge accumulaton s read out wthout destructng the sgnal charge and converted to dgtal. The reset sgnal s subtracted from the ntermedate sgnal n dgtal doman, and the fxed pattern nose (FPN) and ktc nose n the pxel can be cancelled out n dgtal doman. The column amplfer s useful to make sure the exact cancelng of the FPN and ktc nose. For example, n the case of 10bt ADC, the full-scale voltage of 0.8V and the column amplfer gan of 16, the equvalent resoluton n nose cancelng s 50μV. The dgtal low-pass flter further reduces the ktc nose and the pxel source follower nose by a factor of M. Another mportant feature of ths scheme s that the nose s not accumulated n the dgtal ntegrator because of the nose negatve feedback usng DAC. The nose stored n the frame memory v n (0) s gven by vn(0) = M vnp(0) + vnc(0) + q(0) where v np (0), v nc (0) and q(0) are pxel and column amplfer nose, crcuts nose, nose between the column amplfer and the ADC, and the quantzaton nose of the ADC, respectvely, at the samplng
nstance of t=0. The sgnal samplng s at the nstances of t = T0 Δ T (=1,2,...M), where s the nteger, T 0 s the oversamplng frame perod, and the ΔT s the tme dfference of the reset samplng nstance from the fnal sgnal samplng nstance, respectvely. In the sgnal samplng, the followng relatonshp s obtaned: y() = M s() + v () u() c() w () = u () + y() u () = u ( 1) + y ( 1), where v n () s vn() = M vnp() + vnc() + q() and y(), u(),c() and w() are the sgnals shown n Fg. 2. The output of the nose memory s gven by 0 (=0) c () = vn (0) ( 0). From these equatons, the fnal output at the samplng nstance of t = M T0 ΔT s obtaned as wm ( ) = M sm ( ) + vn( M) vn(0) or wm ( ) = M sm ( ) + M ( vnp ( M) vnp (0)) + vnc ( M) vnc (0) + q( M) q(0) In the usual dgtal sgnal ntegraton, the nose s also accumulated as well as the sgnal. The M tmes ntegraton leads to the ncrease of the nose power by a factor of M. In contract to such a dgtal ntegraton, the proposed method allows us to suppress the nose accumulaton because the system predcts the n-pxel sgnal charge ntegraton. n Fg. 4 Tmng dagram of the non-destructve III. Performance predcton wth smulaton Fgure 5 and 6 show the smulated predcton of the SNR mprovement versus oversamplng rato (OSR) for the quantzaton nose of ADC and the crcuts nose n front of and at the back of ADC, respectvely. A CIF(352 x 288) sze mage sensor wth the vdeo rate of 30 frames/s s assumed. In the case that the dgtal LPF s used, the SNR mprovement for the quantzaton nose at the oversamplng rato of 16 s about 27dB. Ths means that the equvalent resoluton s about 14bt f 10bt ADC s used. In the case of the quantzaton nose and the crcuts nose at the back of the column amplfer, the 6dB/octave factor of the nose mprovement s due the amplfer gan tself. The nose accumulaton s successfully suppressed.
square method. Fg. 5 Reducton of quantzaton nose of ADC usng oversamplng processng. Table 1 Assumpton of the nose performance of a CMOS mage sensor. Item Value Converson gan of the pxel 25 μv/e- Input referred nose due to crcuts 2 e- n front of column amplfer Input referred nose due to crcuts 10eat the back of column amplfer Full scale voltage of the 10b ADC 1V Input referred nose due to the 10equantzaton nose of the ADC Image sensor sze CIF (352x288) Vdeo frame rate 30 frames/s Fg. 6 Reducton of crcut nose n front of and at the back of column amplfer usng oversamplng. In the case of the quantzaton nose, dgtal flter s effectve for the reducton of the nose wth the oversamplng. In the crcuts nose n front of the column amplfer, the oversamplng s not so effectve from the smulaton results. However, the domnant nose components are the wdeband output amplfer nose and the quantzaton nose of the ADC. Therefore, the oversamplng technque s useful for the reducton of the total nose behavor. The low-pass FIR dgtal flter s not effectve for the crcuts random nose. We need another nose reducton technque such as a least Fg. 7 Total nput referred nose as a functon of the oversamplng rato. Usng the results of the smulaton shown n Fg. 5 and Fg. 6, the mprovement of the total nput referred nose as the number of electrons s estmated. The assumpton of the CMOS mage sensor performance s shown n the Table 1. Fgure 7 shows the relatonshp between the oversamplng rato and the total equvalent nose referred to the nput (electrons) under the assumpton of the mage sensor shown n Table 1. Wthout the oversamplng, the nput referred nose s about 14 electrons n ths case. At the oversamplng rato of 16, the nput referred nose can be reduced to be about 2 electrons.
IV. Conclusons Ths paper presents a sgnal processng technque to reduce the crcuts nose and quantzaton nose usng frame oversamplng and a CMOS mage sensor n non-destructve readout mode. The random nose level can be reduced the nput referred nose of 2 electrons. Another effectve nose reducton technque s to use estmaton wth least square method n the dgtal flterng secton. Ths s left as the near future subjects. References [1] I.Inoue et al, Low-dark current pnned photo-dode for CMOS mage sensor, 1999 Workshop on CCD and AIS, pp. 25-28 (1999). [2] K. Yonemoto et al, A CMOS mage sensor wth a smple FPN-reducton technology and a hole accumulated dode, Dg. Tech. Papers, ISSCC, pp. 102-103 (2000). [3] Y. Matsunaga, Y. Endo, Nose cancel crcut for CMOS mage sensors, ITE Techncal report (n Japanese), IPU 98-2. Vol. 22, No. 3, pp. 7-11(1998). [4] B. Fowler, A.E. Gamal, D. X. Yang, A CMOS area mage sensor wth pxel-level A/D converson, Dg. Tech. Papers, ISSCC, pp. 226-227 (1994). [5] J. Nakamura et al., On focal-plane sgnal processng for current-mode actve pxel sensors, IEEE Trans. Electron Devces, Vol. 44, No. 10, pp. 1747-1758 (1997). [6] S. Kawahto et al, A hgh-senstvty dgtal CMOS mage sensor usng oversamplng sgnal processng, ITE Tech. Report (n Japanese), IPU2001, pp. 25-30 (2001). [7] D. Handoko, S. Kawahto et. al., "A CMOS mage sensor for focal-plane low-power moton vector estmaton," Dg. Tech. Papers, Symp. VLSI Crcuts, pp. 28-29 (2000.6).