`
`TA 1.2: Camera on a Chip
`
`Btyan Ackland and Alex Dickinson
`
`AT&T Bell Labs, Holmdel, NJ
`Introduction
`
`Development of low-cost video camera technology has, for many
`years, been driven almost exclusively by the camcorder (3 mil-
`lion units per year in the U.S.) and security (1 million units per
`year) markets. The typical product is a multichip camera sub-
`system consisting of CCD sensor, clock drivers and analog sig-
`nal-processing devices to provide color balance and exposure con-
`trol. Output is analog NTSC or PAL. This picture is changing.
`
`Recent advances in video compression and digital networking
`technology, combined with the ever increasing power of PCs and
`workstations, are creating enormous opportunities to develop new
`multimedia products and services built upon sophisticated voice,
`data, image and video processing. This will create a si&cant
`demand for compact, low-cost, low-power electronic cameras for
`video and still image capture. These cameras will be a standard
`peripheral on all PCs bundled for multimedia applications. Given
`that in excess of 60M PCs will be sold this year, a sizable new
`market for electronic cameras is being created.
`
`Present NTSC cameras are not well suited to digital multimedia
`applications: (1) They are too expensive - OEM cost typically
`greater than $100. Experience with the development of CD-ROM
`players suggests that as the OEM cost of a peripheral falls below
`about $50, the vast majority of PCs will ship bundled with the
`peripheral. (2) They are too large to be mounted inconspicuously
`on (or in) a PC monitor. (3) They consume too much power for
`portable applications. CCD sensors require high-voltage, high-
`current clocks - a CCD-based camera subsystem will typically
`consume 1-2W. (4) Video data is output in encoded analog raster
`scan format. This limits both the flexibility of the camera and
`the cost of the overall system. In multimedia systems based on
`camcorder type cameras, additional circuitry must be included
`to convert from the NTSC style output of the camera to the digi-
`tal format required by the application, as shown in Figure 1.
`This circuitry frequently costs more than the camera itself.
`
`What is needed is a camera technology that can be customized to
`particular applications such as desktop video, still scene, and
`document imaging. Cost, size and power constraints require in-
`tegrating the image sensor along with analog and digital signal
`processing and interfacing elements onto the same die, as shown
`in Figure 2. Input to the chip is an image focused onto a sensor
`array. Output is a (possibly compressed) digital stream that con-
`nects seamlessly to the specified multimedia platform.
`
`NTSC
`Signal
`
`Encoder
`
`~
`
`L_ ,,.,,,. _._._ _..~ ..
`
`
`NTSC Camera Board
`
`Bryan Ackland
`
`Bryan Ackland received the BSc in Physics from Flinders University in 1972 and
`the BE and PhD EE from University of Adelaide, Australia in 1975 and 1979. In
`1978 he joined AT&T Bell Labs as Member of Technical Staff, In 1986 he became
`Head of the VLSl Systems Research Department at Holmdel, NJ. His interests
`included raster graphics, symbolic layout and verification tools for full-custom VLSI,
`MOS timing simulation and VLSl layout synthesis.His interests now are VLSl ar-
`chitectures and design tools for high-performance signal processing and commu-
`nications, particularly multimedia. He is an AT&T Bell Labs Fellow and an IEEE
`Fellow.
`
`CCD Sensors
`
`Any solid-state imaging device consists of an array of sensing
`elements combined with some form of transport mechanism to
`deliver these sensor outputs to the periphery of the die. Sensors
`used in commercial devices include photodiodes, MOS capaci-
`tors, charge injection devices and bipolar phototransistors. All
`of these devices use essentially the same light sensing mecha-
`nism. Photons penetrating a depletion region generate electron-
`hole pairs. These are swept away by the electric field across the
`depletion region and generate a small photocurrent.
`
`Except under very bright light conditions, it is not possible to
`use this photocurrent directly. Even at 100% conversion efficiency,
`a lOpm sensor illuminated at llux will generate a photo current
`of only 70fA. Tc achieve reasonable signal-to-noise ratio, these
`currents are usually integrated (typically for 15ms) to produce
`an accumulated charge output. For the above example a charge
`of
`coulombs (6000 e-) would be accumulated in this time.
`
`The CCD provides a simple mechanism for transporting these
`small charge packets out of the array. A CCD is a linear array of
`
`To PC Bus.
`Video Codec,
`etc.
`I ...........................................................
`
`4 ................I
`
`I
`
`i _______________,__,,,,. .............................................................
`Camera Die
`
`I
`
`Figure 1: Conventional multimedia camera.
`
`Figure 2: Camera on a chip multimedia camera.
`
`22
`
`* 1996 IEEE International Solid-state Circuits Conference
`
`0-7803-3136-2 J 96 I$5.00 I O IEEE
`
`Magna 2024
`TRW v. Magna
`IPR2015-00436
`
`0001
`
`
`
`ISSCC96 / February 8,1996 / Presidio / Sea Cliff / Buena Vista / 10:30 AM
`
`MOS capacitors that function as a charge-domain shift register
`when driven by a set of multiphase clocks. In low-cost cameras,
`interline transfer, as shown in Figure 3, is the most commonly
`used architecture. Adjacent to each column of sensors is a verti-
`cal CCD. Accumulated charge is transferred from the sensor to
`the corresponding CCD bucket. These charges are then shifted
`vertically, one line at a time into a horizontal shift register. A
`single line is scanned out from the horizontal shift register onto
`a capacitor which converts each charge packet to a voltage for
`subsequent amplification and buffering.
`
`CCD sensors have improved dramatically since introduction in
`the 1970s. Scientific arrays of 4096x4096 resolution with dynamic
`noise levels of 3-5 electrons and a dynamic range of over 80dB
`have been demonstrated. Low-cost commercial devices typically
`provide 640x480 pixel resolution with a signal-to-noise ratio of
`45dB. One advantage of CCD sensors is the high fill factor (ratio
`of light sensitive area to total pixel area) that can be obtained,
`even for small pixels. This is because the only extra area re-
`quired is that of the CCD register. State of the art sensors use
`pixel sizes of 5x5pm, which approaches the optical diffraction
`limit. A second advantage of these systems is their lack of pat-
`tern noise. This is caused by variations in offset and gain from
`one pixel to another. In a CCD, all image information travels
`through closely matched paths in the charge domain, and then
`shares the same charge-to-voltage output stage.
`
`CCDs, however, have a number of disadvantages in multimedia
`applications. Most of these stem from the need to maintain high
`charge-transfer efficiency (CTE) within the CCD shift register.
`CTE is a measure of the percentage of electrons that are suc-
`cessfully transferred from one bucket to another in one CCD shift
`cycle. In a 640x480 sensor, a single charge packet may be shifted
`1100 times. Even with a CTE of 0.9995, this will result in a 40%
`loss of charge by the time the packet reaches the output stage.
`
`One way of increasing CTE is to use 12-15V on the CCD clock
`lines. This, in turn, leads to high power dissipation. There has
`been much work recently in trying to develop low-voltage CCD
`processes better suited to multimedia applications. Low-power
`consumer-grade sensors have been reported in which most of
`the high-speed clock circuitry is driven at 3.3V [ l , 21.
`
`Another approach is to tune doping profiles to maximize charge
`carrying capacity and minimize charge loss during transfer. This
`has lead to processes that provide excellent CCD performance
`but are not at all suitable for producing standard digital VLSI
`circuits. Digital camera proposals based on CCDs [31 typically
`require at least three separate die: one for the sensor, a second
`for analog clock drivers a n d m and a third for the digital signal
`processing components [31. However, a recently reported a 2pm
`process supports high-quality CCDs, along with npn bipolar and
`
`pepeeee5_
`
`column decode
`
`Figure 4: Active pixel sensor array.
`conventional CMOS devices [41. Such a process allows for the
`possibility of integrating all camera functions onto a single chip
`at a cost of four extra mask layers and three extra implants.
`CMOS Sensors
`
`An alternative approach is a two-dimensional addressable array
`of sensors [5-81. The architecture is similar to that used in con-
`ventional random-access memories. A bit line is associated with
`each column of sensors as shown in Figure 4. A row-enable line
`allows each sensor in a selected row to place its output onto its
`bit line. A multiplexer at the end of the bit lines allows for indi-
`vidual column addressing. This is an old idea dating back to the
`early 1970s, but one that, until recently, has not found commer-
`cial application. The difficulty arises from the very small charge
`generated at each pixel. From our previous example, a charge of
`10-l6 coulombs placed on a 2pF bit-line, generates a voltage change
`of 5OpV Such a signal is susceptible to noise. What is needed is a
`simple amplifier at each pixel to provide buffering as shown in
`Figure 4. This is referred to as an active pixel sensor (APS) array.
`
`In the early 70s, with sensor sizes of 25pm and design rules of
`5pm, it was not possible to include an amplifier at each pixel.
`Today, with sensor sizes of 1Opm and process design rules a t 0.5pm,
`a three transistor pre-chargelamplifylselect circuit, as shown in
`Figure 5, can be built and still achieve a fill factor of over 25%.
`One advantage ofAPS sensors is that they can be powered from a
`single 3.3V (or lower) supply and do not require external
`multiphase clock generators. This leads to reduced system costs
`and significantly reduced power dissipation. Integrated CMOS
`cameras with a total power dissipation under lOmW have been
`reported. Silicon processing costs are also reduced because of the
`huge volumes associated with generic digital CMOS production.
`Another advantage is the flexibility that is provided by a random
`access sensor addressing. No longer is one constrained to access
`pixels in serial order determined by the architecture of the CCD
`
`verti
`CCD
`
`\
`
`Horironlal
`CCD
`
`- -
`
`Figure 3: Interline transfer CCD.
`
`t
`
`output
`
`Figure 5: Active pixel circuit.
`
`DIGEST OF TECHNICAL PAPERS
`
`23
`
`0002
`
`
`
`ISSCC96 I SESSION 1 I PLENARY SESSION I PAPER TA 1.2
`
`pipeline. This simplifies introduction of alternate forms of pixel
`access, such as that required for electronic pan & zoom.
`Arguably the most important advantage of the CMOS APS ap-
`proach, however, is the ability to integrate much of the camera
`timing, control and signal processing circuitry onto the same
`silicon die. A CIF (352x288) sensor array can be built in 0.5ym
`CMOS in under llmm2 active area. Even assuming a 30mm2
`die, this leaves area for these system functions, plus extra cir-
`cuitry to customize the camera for a particular application.
`One problem associated with APS arrays is the existence of sig-
`nificant levels of fixed-pattern noise. Unlike CCDs, in which all
`charge packets travel essentially the same signal path and use
`the same output amplifier, each pixel in an APS array has its
`own amplifier. Gain and offset variations between these ampli-
`fiers lead to a static pattern noise which appears as a background
`texture on the image. The eye is very sensitive to small amounts
`of pattern noise (1-2% is clearly visible), particularly if i t is
`aligned along vertical or horizontal lines, as is the case with
`noise contributed by column output stages.
`
`Fortunately, the ability to integrate signal processing electron-
`ics onto the same die provides a number of solutions to this prob-
`lem. Fixed-pattern noise can be significantly reduced by the use
`of a simple correlated double sampling (CDS) technique as shown
`in Figure 6. Each column output stage contains two sample and
`hold capacitors. One is used to sample the pixel reset level, one
`is used to sample the signal level (after integration). By sub-
`tracting the reset level from the signal level, much of the pixel
`amplifier offset is removed. A second level of CDS can be applied
`to eliminate offset in the column amplifier. Acombination of these
`two techniques can reduce fixed pattern noise from 5% to 0.1%.
`Or digital techniques can be used to reduce the level of pattern
`noise. The offset level for each column, for example, could be
`stored in a RAM and digitally subtracted from the output signal
`column by column. These techniques can reduce the pattern noise
`to a level imperceptible under the room-light (or brighter) levels
`of illumination found in most multimedia applications.
`A Single-Chip Multimedia Camera
`
`The basic architecture of a camera on a chip for multimedia ap-
`plications is shown in Figure 7. The detailed functionality of
`each module depends on the nature of the application. Of par-
`ticular significance is the degree of autonomy required of the
`camera. At one extreme, represented by a conventional video
`camera, the camera operates in a stand-alone mode, calculating
`exposure times and color balance and producing a stream ofvideo
`information at a predetermined frame rate.Amore flexible model,
`however, is one in which camera functionality is partitioned be-
`7 Vdd
`
`tween the camera hardware (simple, per-pixel operations) and
`soRware (complex, per-frame operations) in an intelligent host
`(e.g., PC), thereby reducing the cost of the camera hardware and
`increasing the functionality of the overall system. The camera is
`no longer autonomous, instead having a symbiotic relationship
`with the host and the host application as shown in Figure 7.
`Exposure control may be partitioned between the camera that
`maintains summary statistics on intensity derived from exam-
`ining each pixel and the host that uses the summary statistics to
`calculate exposure time that is passed to the camera.
`The sensor array is comprised of a two-dimensional pixel array
`that can be randomly addressed through adjacent row and col-
`umn decoders. An emerging standard for video-telephony is CIF
`at 288 by 352 pixels, somewhat less than NTSC resolution, but
`sufficient for many compressed video and snap-shot applications.
`
`The standard video frame rate is 30 frames per second. Multi-
`media applications, however, typically operate at 10 - 15 frames
`per second. These numbers suggest per frame exposure times of
`approximately 60 to 100ms. These are however maximum expo-
`sure times that are only used in moderate to low-light condi-
`tions. As lighting intensifies, it is necessary to reduce the effec-
`tive exposure time to ensure the sensors do not saturate (exceed
`their maximum charge capacity). Long exposure times may also
`lead to high levels of dark current pattern noise. The exposure
`control block generates timing signals to set the interval between
`resetting a row of pixels and reading the accumulated charge.
`Exposure time is calculated to maintain a specified distribution
`of pixel brightness levels. This calculation can be performed ei-
`ther by on-board circuitry, or software on the host.
`
`A specific region of the sensor may be read simply by presetting
`the counters that drive the row and column decoders to values
`that represent the origin of the region of interest at the start of
`every new frame. Electronic panning may then be implemented
`by altering the preset values to the new window origin under
`user (or application) control.
`
`Traditionally, digital cameras have used a single AID converter
`operating at the pixel rate (3Mpixels/s for CIF, 30frames/s). Eight-
`bit resolution is sufficient provided that AGC and gamma cor-
`rection have already been performed in the analog domain. Hav-
`ing the converter on the same die as the sensor, however, allows
`alternative A/D solutions. A CMOS camera in which a ZA AID
`converter has been effectively integrated into each pixel, allow-
`ing digital readout at the cell level [91. The sensor capacitance
`conveniently performs summing and integrating. This simpli-
`fies signal read-out but leads to larger pixel sizes and reduced
`sensitivity since the raw sensor current is now the A/D input.
`
`Camera Chip
`
`From Pixels
`
`4
`
`Host
`
`Software:
`* Exposure control
`*White balance
`-Pan control
`
`- Matrix elements
`
`Figure 6: CDS fixed pattern noise reduction circuit.
`
`Figure 7: Camera on a chip hardwarelsoftware partitioning.
`
`24
`
`1996 IEEE International Solid-state Circuits Conference
`
`0-7803-3136-2 I 96 i $5.00 I O IEEE
`
`0003
`
`
`
`ISCC96 I February 8,1996 I Presidio / Sea Cliff I Buena Vista / 10:30 AM
`
`Alternatively, a simple low-speed single-slope converter may be
`placed at the output of every column as shown in Figure 8. A
`comparator compares the column output against a reference. The
`reference voltage is a ramp derived from a counter feeding a DIA
`converter. When the column comparator senses equality between
`the column signal and the reference, the counter output is re-
`corded in a register associated with each column. At the end of a
`single conversion cycle (one row of video) the registers contain
`digital values representing the analog output of each column,
`and may be sequentially (or randomly) addressed to generate a
`digital output signal from the chip. Additional signal processing
`(AGC, gamma correction) may be performed by modifying the
`slope and the shape of the reference ramp.
`
`Additional signal processing is usually required to compensate
`for variations in processing and operating conditions. These may
`be performed in either the analog or the digital domain. Auto-
`matic gain control (AGC) is required in moderate to low-light
`conditions where the camera is running at its maximum expo-
`sure time. Gain may be added into the signal path to increase
`the apparent brightness of the scene. This has the effect, how-
`ever, of also amplifying noise, so the total gain will be limited by
`the signal-to-noise ratio of the array. Gamma correction is re-
`quired to compensate for display non-linear response. One imple-
`mentation simply uses the digital value of the video signal to
`address a lookup table containing the corrected signal values.
`
`It is usually necessary to color correct (white balance) to com-
`pensate for non-ideal response in the sensors and color filters.
`This may take the form of a simple gain control on each compo-
`nent channel. Digital processing, however, provides a more flex-
`ible solution in the form of a color transformation matrix:
`x , , x , , x , ,
`R'
`
`kI=k;: ;:: :;jEl
`
`This transformation is applied to each N G B sample to allow for
`crosstalk and response variation between the three color chan-
`nels. It can be used to correct significant errors in sensor spec-
`tral and color filter responses and therefore simplifies color fil-
`ter array specification. In addition, such a matrix can generate
`alternative color space formats (e.g., YCrCb, YW).
`To maximize the value of integration it is necessary to select an
`appropriate digital interface and implement it on-chip. Clearly
`the interface is highly application dependent, with choices from
`a simple proprietary unidirectional data stream for connection
`directly to a video codec, to more complex bi-directional stan-
`dard such as IEEE 1394 ("Firewire") for connection to a PC.
`
`Back-End Manufacturing Steps
`
`Although the concept of a single chip camera is based on fabri-
`cating the entire device on a single die in a standard CMOS pro-
`cess, a number of additional or back-end steps must be included
`to complete the overall manufacturing process.
`
`Chip packaging costs are a significant portion of the complete
`costs of a sensor. A sensor package must: (1) have a transparent,
`hermetically sealed lid, (2) be able to adequately dissipate sen-
`sor power, (3) be manufactured (including chip placement) to tol-
`erances sufficient for inclusion in an optical system.
`
`The relatively high power dissipated by CCD sensors and their
`sensitivity to thermal stress often requires use of costly ceramic
`or precision plastic packages. The much lower power require-
`ments of CMOS arrays permits the use of conventional plastic
`packages, considerably lowering the cost of the completed part.
`
`Testing optical devices requires more complex facilities than those
`used for standard digital parts. Additional test issues include:
`(1) providing a controlled optical source for photodetector stimu-
`lation, (2) digital tester inputs capable of verifying that a good
`output lies in a range of values rather than a single value.
`
`To construct a single-sensor color camera it is necessary to pat-
`tern the sensor surface with a suitable mosaic of color filters.
`There are two color systems available: additive (red, green, blue)
`and subtractive (magenta, cyan, yellow). Subtractive filters have
`the advantage that they let more light onto the sensor and pro-
`vide greater sensitivity. Additive filters lead to simpler color pro-
`cessing. Filters are typically made from dyed polyamide, each
`color being lithographically defined and etched in turn to create
`an individual filter square over each pixel. Various patterns, such
`as the one shown in Figure 9, have been proposed. Typically pat-
`terns are chosen to emphasize the luminance (or green) resolu-
`tion while chrominance (or rediblue) resolution is sampled at
`lower resolution. Depending on the pattern chosen, various
`amounts of buffering (typically one to two lines) and processing
`(additiodsubtraction) may be needed to derive the required per-
`pixel output data from the raw array data.
`
`Because the fill factor of a pixel is typically around 3096, consid-
`erable gains in sensitivity can be achieved by constructing a lens
`over the entire pixel to focus light onto the active area. These
`microlenses may be fabricated by patterning and etching polya-
`mide to form a cylinder over each pixel, then heating the mate-
`rial until it flows to form a spherical lens over the pixel.
`
`In addition to the single silicon die, a complete camera will re-
`quire a simple PC board (perhaps including a voltage regulator),
`a plastic housing, lens, cabling and connector. Recent develop-
`ments in the production of precision injection molded plastic
`asphere lenses suggest that a high-quality multi-element lens
`can be produced at a fraction of the cost of glass spherical lenses.
`For simple fixed-focus applications these may be attached di-
`rectly to the camera chip package.
`
`Evolution of Single-Chip Cameras
`
`The concentration of work on single-chip cameras is presently
`aimed at the production of low-cost color video cameras for mul-
`timedia applications. As the CMOS sensor array technology
`evolves, and process line widths continue to decrease, we expect
`to see: (1) CMOS image sensor arrays will become standard cells
`much like any other chip layout component. For example an im-
`age sensor and a video encoder may be combined with some
`ASIC glue to create an application-specific product. (2) more so-
`phisticated use of applicatiodcamera interaction resulting in in-
`creased image quality and production values (e.g., automatic
`head tracking supported by electronic pan and zoom). ( 3 ) higher
`resolution sensors for document image capture applications such
`as fax. (4) high-resolution sensors for all-electronic consumer still
`camera applications. (5) low-cost, low-resolution sensor3 that al-
`low intelligent machine vision functions to be added to consumer
`items such as automobiles and home appliances.
`
`The product life-cycle of the digital clock may provide some indi-
`cation of the future of cameras. Initially a relatively costly stand-
`alone device, digital clocks became cheap enough to be combined
`with radios, and then eventually became standard features of a
`wide range of products from microwave ovens to VCIls. Simi-
`larly, the high levels of integration enabled by the development
`of CMOS image sensor technology will drive consumer electronic
`cameras from their present stand-alone (camcorder) form to be-
`ing a ubiquitous feature of everyday life.
`Figures 8 and 9 and References: See page 412.
`
`DIGEST OF TECHNICAL PAPERS
`
`25
`
`0004
`
`
`
`References:
`
`TA 1.2 Camera on ti Chip
`(Continued from page 25)
`.
`
`[l] Fujikawa, F., et al., “A U3-inch 630kpixel IT-CCD Image Sensorwith
`Multi-Function Capability,” IEEE ISSCC Digest of Technical Papers, pp.
`218-219, Feb., 1995.
`[2] Bosiers, J., et al., ”Design Options for l/4”-€T-CCD Pixels,” I”.
`IEEE Workshop on Charge-Coupled Devices and Advanced Image Sensors,
`April, 1995.
`[31 Wang, S., et al., “A Real-Time Signal Processor for use with the
`Interline Transfer Color CCD Imager,” Proc. IEEE Workshop on Charge-
`Coupled Devices and Advanced Image Sensors, April, 1995.
`
`141 Guidash, R, et al., “A Modular, High Performance, 2pm CCD-BiC-
`MOS Process Technology for Application Specific Image Sensors and
`Image Sensor Systems on a Chip,” Proc. of IEEE Intl. ASIC Conf. C
`Exhibit, pp. 352-355, April, 1994.
`
`[51 Renshaw, D., et al., “ASIC Vision,” Roc. IEEE Custom Integrated
`Circuits Conf., pp. 7.3.1-7.3.4, 1990.
`[61 Fossum, E., “Active Pixel Image Sensors - Are CCDs Dinosaurs?,”
`h c . SPIE, vol. 1900, pp. 2-14,1993.
`[71 Jansson, C., et al., “An Addressable 256x256 Photodiode Image Sensor
`Array with an 8-bit Digital Output,” Analog Integrated Circuits and Si&
`F’rocessing, pp. 37-49, Vol. 4, 1993.
`181 Dickinson, A., et al., “Standard CMOS Active Pixel Image Sensors for
`Multimedia Applications,” Proc. Conf. on Advanced Research in VLSI, pp.
`214224, March, 1995.
`191 Fowler, B., et al., “A CMOS Image Sensor with Pixel Level AD
`Conversion,” ISSCC Digest of Technical Papers, pp. 216-217, Feb., 1994.
`
`row
`deccdw
`
`counter
`
`P pixel
`L
`latch
`>
`comparator
`
`column decoder
`
`Figure 8: Per-column, single-slope AID converter.
`
`Figure 9: Color mosaic pattern.
`TP 2.1: A 3.3V 128Mb Multi-Level NAND Flash Memory for Mass Storage Applications
`(Continued from page 33)
`
`Figure 6: l2SMb NAND flash memory chip micrograph.
`
`412
`
`1996 IEEE International Solid-State Circuits Conference
`
`0005