Organic Name Reactions for the students and aspirants of Chemistry12th.pptx
From Biological to Smart CMOS Imaging: Architectural approach
1. CMOS IMAGE SENSORS
From the Biological Eye to Smart Machine Vision
.
Fayçal Saffih, Ph.D. , Visiting scholar
Communication Research Laboratory (CRL)
McMaster University
2. Outline
CMOS Imaging Technology
New Advancements in Smart Cameras
Applications in the Medical Domain
Contributions and Future Research
3. CMOS Imaging Technology
InStat/MDR
Two main electronic imaging technologies: Complementary Metal-
Oxide-Semiconductor (CMOS) and Charge-Coupled-Devices
(CCD)
Total unit shipments of CMOS image sensors surpassed CCDs for
the first time in 2005. (In-STAT/MDR)
CMOS image sensor shipments will grow at roughly seven times
the rate of CCDs through 2008 (In-Stat/MDR)
4. Why CMOS imaging is important ?
Architectures flexibility allows a lot of flexibility of image acquisition in
CMOS imagers compared to their CCD counter parts.
CMOS Imagers are well known for their low power (20-50 mW) compared to
CCD’s (2-5W) for the same pixel throughput. This is why CMOS imagers
are dominating portable devices market and wireless medical imaging .
Integration of acquisition and data processing is only possible in CMOS
imaging technology.
Compatibility with main stream CMOS technology. As such, every CMOS
foundry has CMOS image sensor division.
5. Research Contribution and
Future Work
Foveated Sampling Architectures
for CMOS Image Sensors
Ph. D. Thesis available at:
http://www.cs.yorku.ca/~visor/pdf/FaycalPhDThesis.pdf
6. Philosophical Statements
Image Scanning Statement
Image Scanning should be more adapted to image sampling/acquisition
rather than image display compatibility.
Image Information Acquisition is the priority of the imager.
Image Sampling Statement
As image resolutions get higher and with it the amount of transmitted
image data for display or processing, new architectures are needed for
down-scaling the sampling resolution for regions of reduced interest.
Innovative architectures are also needed to exploit the Human Visual
System for transmitting only the most important regions in the acquired
image.
Only Regions of Interest are sampled at highest imaging
requirement such as dynamic range, spatial resolution ... etc.
7. Why Foveated CMOS Imagers?
Foveated image sampling architectures are needed due to:
1. Limited imaging system resources (memory, processing power & time)
2. Not all image areas are of the same degree of importance (or of interest)
3. Only area of interest are sampled at the highest resolutions, highest
dynamic range…etc.
4. Biologically inspired
“Nature does nothing uselessly.”
uselessly
–Aristotle, Politics
“The diversity of the phenomena of nature is so great, and the treasures
hidden in the heavens so rich, precisely in order that the human mind
shall never be lacking in fresh nourishment”
nourishment
–Kepler, from Carl Sagan’s book “Cosmos”
8. Why Foveated CMOS Imagers?
Ramon Cajal identified the retina basic anatomical
structure. Shown here is his sketch of the interconnectivity
configurations of photocells rodes (f) and cones (e) [4]
9. Our Research Contribution Outline
Standard CMOS Image Sensor
System Level Approach (Time Domain Fovea)
Pyramidal CMOS Image Sensor
Architecture
Physical Design
Scanning
Foveated Dynamic Range Enhancement
High-Speed Imaging of Pyramidal CMOS Imager
Low Pyramidal Imager FPN Perception by HVS
Pixel Level Approach (Spatial Domain Fovea)
Multiresolution CMOS Image Sensor
Architecture
Multiresolution Active Pixel Sensor
Results
Conclusion and Future Work
12. The Seminar Outline
Standard CMOS Image Sensor
System Level Approach (Time Domain Fovea)
Pyramidal CMOS Image Sensor
Architecture
Physical Design
Scanning
Foveated Dynamic Range Enhancement
High-Speed Imaging of Pyramidal CMOS Imager
Low Pyramidal Imager FPN Perception by HVS
Pixel Level Approach (Spatial Domain Fovea)
Multiresolution CMOS Image Sensor
Architecture
Multiresolution Active Pixel Sensor
Results
Conclusion and Future Work
13. System Level Approach
Time Domain Approach:
Because Dynamic Range enhancement can be
realised by multi-exposure concept, controlling the
topology of imager’s integration time profile may
lead to a non-uniform and interesting Dynamic
Range enhancement
Design Approach:
Keep the standard APS structure while changing the
Architectural Sampling structure
Pyramidal CMOS Image
Sensor
17. Foveated Dynamic Range Enhancement
Mathematical Foundations: Prediction…
Definitions:
Integration time of a pixel (or of RST-
SEL shared sampling entities): Time
between consecutive RST & SEL
signals.
Ts: Pixel scanning time (0.1µs ~ 1µs)
Tspl: Row sampling time row
sampling including Vs sampling, reset,
Vr sampling (1µs ~ 10µs)
r: ring order
R: Total number of rings
Tin(r): Inward scan integration time.
Tout(r): Outward scan integration
time.
r +1 r −1
Tin( r , R, Ts , Tspl ) = 2 ∑ iTs + ( R − r )Tspl + rTs , Tout ( r , Ts , Tspl ) = 2 ∑ s
iT + ( r − 1)T + rT
i =1
i = R spl s
i→i −1
i →i +1
…and after simplification we get:
(
Tin ( r , R, Ts , Tspl ) = −Ts r 2 − 2Tspl r + R 2Ts + RTs + 2 RTspl )
Tout ( r , Ts , Tspl ) = Ts r 2 + 2Tspl r − 2Tspl
18. Foveated Dynamic Range Enhancement
Mathematical Foundations: Prediction…
(
Tin ( r , R, Ts , Tspl ) = −Ts r 2 − 2Tspl r + R 2Ts + RTs + 2 RTspl )
Tout ( r , Ts , Tspl ) = Ts r 2 + 2Tspl r − 2Tspl
ToT ( R, Ts , Tspl ) = Tin( r , R, Ts , Tspl ) + Tout ( r , Ts , Tspl ) = R 2 Ts + R (Ts + 2Tspl ) − 2 Tspl
19. Foveated Dynamic Range Enhancement
Mathematical Foundations: Prediction…
max (Tin( r , R, Ts , Tspl ), Tout ( r , Ts , Tspl ) )
DRenh (r ) = 20 log
min (Tin ( r , R, Ts , Tspl ), Tout ( r , Ts , Tspl ) )
1 1 2
FOVborder (α , R ) = 1 + α R ( R + 1) + α ( R + 1) − 1 , where α = Ts T spl
α 2
20. Foveated Dynamic Range Enhancement
Mathematical Foundations: Controllability…
Pinned Foveated Dynamic Range Enhancement
21. Foveated Dynamic Range Enhancement
Mathematical Foundations: Advantage …
Dynamic Range is usually
associated to the number of
“meaningful” bits of the
digital output of an imager.
DRdB
DRbits = log 2 10 20
3D view of the pinned FDR enhancement expressed in binary bits.
22. Foveated Dynamic Range Enhancement
Experimental Setup…
CMOS Pyramidal Imager under test inside black box
23. Foveated Dynamic Range Enhancement
Experimental Results…
The model extraction will provide us:
1. Lmin: Minimum detectable light intensity
2. Lmax: Maximum saturation light intensity
3. Ss: Sensitivity of the imager
The Objective of this methodology is the extraction of:
1. The System electrical dynamic range
2. The Optical dynamic range enhancement
3. Verification of the mathematical predictions of dynamic range enhancement
24. Foveated Dynamic Range Enhancement
Experimental Results…
Vcds in RMS voltage the whole pyramidal imager for 8 Slopes of linear regions of photon transfer curve
integration times
25. Foveated Dynamic Range Enhancement
Experimental Results…
Measured and modeled values for Vcds Correlation of the model to the measured values
26. Foveated Dynamic Range Enhancement
Experimental Results…
Definitions:
Most researchers in CMOS imagers
agree that Lmin is constant with
integration time.
Lmin ≈ 5Lux
Lmax is calculated based on the model.
It is at which light intensity the pixel
will reach saturation
System and Enhancement Dynamic Ranges Extraction
of the Pyramidal CMOS imager at 1MHz (800 Fps)
CalculDR = 20 Log ( Lmax Lmin )
max( MaxDetLI inward , MaxDetLI outward )
OptDRMx _ MnDet (r ) = 20 log
min ( MaxDetLI inward , MaxDetLI outward )
max (Tin( r , R, Ts , Tspl ), Tout ( r , Ts , Tspl ) )
Theory_DRenh = 20 Log
min (Tin ( r , R, T , T ), Tout ( r , T , T ) )
s spl s spl
27. Pyramidal CMOS Image Sensor
Experimental Results: Ring Sampling & Blur Symmetrization
Motion Blur effect on rolling-shutter classical CMOS imagers Global shutter versus rolling shutter and motion blur
Rolling shutter is used to separate integration time from frame time (speed)
To cancel motion blur, Global shutter is used but mainly for still imaging applications
To have a motion-blur free video (continuous) images, global shutter is not adequate due to :
Higher transmission rate requirements to transfer high frame rate images.
Higher power consumption
Excessive heat (increasing dark current).
The above factors are more pronounced with increasing CMOS imagers resolutions.
28. Pyramidal CMOS Image Sensor
Experimental Results: Ring Sampling & Blur Symmetrization
The PixeLINK CMOS imager window (784x784) exposure (or integration) time = 43.2ms
The integration time the 64x64 pyramidal CMOS imager =
43.24ms
29. Foveated Dynamic Range Enhancement
Experimental Results…
( + )x½= ,
Inward scan Outward scan Fused Image Rolling scan
Demonstration of Foveated Dynamic Range Enhancement at the Foveal Rings
( + )x½= ,
Inward scan Outward scan Fused Image Rolling scan
Demonstration of Foveated Dynamic Range Enhancement at the Peripheral Rings
30. Pyramidal CMOS Image Sensor
High-Speed Imaging …
FRPyr R + 1.5 Inner ring
=8 R →∞ → 8
FRClass R+7
Outer ring
R=1024
The Pyramidal CMOS imager is designed primarily for parallel ring sampling with 8 output channels
This property of multi-channel readout is inherent to the pyramidal architecture in to contrast to multi-channel
technique included in high speed classical CMOS imagers.
In serial readout of sampled rings:
Pyramidal imager is faster at its fovea (central rings) than a classical CMOS imager of similar size of the
fovea.
Pyramidal imager is very close in image sampling speed with the classical imager at higher imager resolutions.
31. Low Pyramidal Imager FPN Perception by
Human Visual System
Oblique effect found in the contrast sensitivity of the HVS at relatively high frequencies [94]
Average spatial power spectrum distribution of about 500 natural scenes [100]
32. Low Pyramidal Imager FPN Perception by
Human Visual System
Classical CMOS Imager FPN
2D-Fourier
Topology
Spectrum
33. Low Pyramidal Imager FPN Perception by
Human Visual System
Pyramidal CMOS Imager FPN
Topology
34. HVS Spatial Filter Application on
Pyramidal FPN images: Model Human Visual System
Low Pyramidal Imager FPN Perception by
35. HVS Spatial Filter Application on
Pyramidal FPN images: Results Human Visual System
Low Pyramidal Imager FPN Perception by
36. HVS Spatial Filter Application on
Pyramidal FPN images: Results Human Visual System
Low Pyramidal Imager FPN Perception by
37. HVS Spatial Filter Application on
Pyramidal FPN images: Results Human Visual System
Low Pyramidal Imager FPN Perception by
38. HVS Spatial Filter Application on
Pyramidal FPN images: Results Human Visual System
Low Pyramidal Imager FPN Perception by
39. Low Pyramidal Imager FPN Perception by
Human Visual System
HVS Spatial Filter Application on
Pyramidal FPN images: Conclusion
FPN is known to increase with increasing sampling frequency
(or with decreasing integration time). This is why more
oblique stripes are more visible with increasing sampling
frequency.
The positive value of the oblique FPN stripes infers the
presence of the FPN noise stripes in the tilted HVS filter and
its absence from the normal HVS filter. Therefore, the
pyramidal imager FPN noise is less perceived by normal
human observer in contrast to the normal CMOS imager
counter part.
The importance of sampling architectures is proven again not
only to provide better imaging performances, but can make
some inherent limitation of CMOS image sensors, such as
higher FPN compared to CCD’s, less noticeable.
40. The Seminar Outline
Standard CMOS Image Sensor
System Level Approach (Time Domain)
Pyramidal CMOS Image Sensor
Architecture
Physical Design
Scanning
Foveated Dynamic Range Enhancement
High-Speed Imaging of Pyramidal CMOS Imager
Low Pyramidal Imager FPN Perception by HVS
Pixel Level Approach (Spatial Domain)
Multiresolution CMOS Image Sensor
Architecture
Multiresolution Active Pixel Sensor
Results
Conclusion and Future Work
41. Pixel Level Approach
Spatial Domain Approach:
Only regions of interest are sampled at the highest possible resolution.
Other regions of less (or no) interest are sampled at lower resolutions.
human Eye has variable interconnectivity configurations between the
photo-cells and the ganglion cells (processing neurons):
one-to-many configuration for cones (highest resolution needed for
reading)
Many-to-one configuration for rods (low resolution used for low
light vision)
Design Approach:
Keep the CMOS Image Sensor architecture while changing the APS
Sampling structure
Multiresolution CMOS Image Sensor
42. Multiresolution CMOS Image Sensor
Image Sampled at highest resolution in all pixels
Only Regions-Of-Interest (ROI) are kept at their highest resolution
(no-charge distribution)
Regions-Of-Less-Interest (ROLI) are selectively down-resolved by
a charge-sharing mechanism to get a single value representing their
average.
Previous implementations were chip and column level approaches
using the same concept.
The suggested approach is pixel-based to ensure the expandability
of the architecture to any imager size without increasing
complexity.
Further details about the mechanism implementation, layout design as
well as simulated Multiresolution foveated images are discussed in the
thesis.
43. The Seminar Outline
Standard CMOS Image Sensor
System Level Approach (Time Domain)
Pyramidal CMOS Image Sensor
Architecture
Physical Design
Scanning
Foveated Dynamic Range Enhancement
High-Speed Imaging of Pyramidal CMOS Imager
Low Pyramidal Imager FPN Perception by HVS
Pixel Level Approach (Spatial Domain)
Multiresolution CMOS Image Sensor
Architecture
Multiresolution Active Pixel Sensor
Results
Conclusion and Future Work
44. Conclusion and Future Work
Conclusion:
The architecture flexibility of CMOS imaging technology was:
Key factor:
In enhancing the optical dynamic range following a foveated profile
mimicking human eye (biological) vision.
Key factor:
In minimizing the perception of the CMOS imager FPN noise by human
observer benefiting from HVS spatial discrimination limitation (Oblique
Effect)
Key factor:
In developing similar multiresolution mechanisms as those found at retina of
the human eye which lead to an expandable multiresolution CMOS imager.
Future Work
Biological vision systems can give a great help in synthesising novel
architectures exhibiting more adaptability, to light intensity for higher ranges
and spatial resolving power. The Pyramidal CMOS imager and the
multiresolution CMOS imager are first steps towards this vision.
1. Shown is only the discharge site of the generated e-h pairs at the depletion region 2. The e-h pairs can be generated at: depletion region, peripheral regions, the substrate 3. The e-h pairs are drifted towards the discharging site by the electrical field or by diffusion
Although integration time could be controlled at the pixel level, we choose the system level (control) to minimize the drawbacks of pixel integration such as: 1. low fill factor, 2. extra noise sources, 3-decreased resolution
Why we suggested this pyramidal based on :1-our 1 st philosophical statement : Image info first 2-circular symmetry of biological vision systems (output diagonal busses are consequence of ring sampling) 3- the low dimension (1D) of the raster scan
This is the first prototype of the pyramidal CMOS imager not optimized in: a-area, b-power consumption. Layout symmetry was a very interesting/useful factor. Pixel size = 16umx16um=256um 2 Fill Factor =78% Chip size =4mmx4mm
Raster scan (also called rolling scan) is also possible in pyramidal imager Shown is : the readout path between the pyramidal and the classical CMOS imager. Bouncing scan is possible in classical imager but with being its purpose (the circular symmetry and foveated DR enhancement)
The bouncing edges are the inner ring and the outer ring. Integration time of a pixel (or of RST-SEL shared sampling entities)= time between consecutive RST & SEL signals.
Photo-Signal= Integration time x Light intensity x Sensitivity Linearity: As long as the photo-response is linear (no saturation) the fused image is: 1- also linear 2- independent on ring order r (in contrast to inward and outward scan)
The minimum of the FDR enhancement IS where Tin=Tout. This point is called the fovea border FOVborder fast reaching its limit at α = 1 and above Fovea Border :1- Is situated at (R/sqrt(2)) or (71%)R, 2-can be programmed to pin at the outer/peripheral ring for a perfect foveated DR enhancement shape, 3- will be used later on for determining the system electrical DR and the FDR enhancement.
Controllability concerns the profile of the FDR enhancement by pinning its border to the outer ring and other alternatives such as reverse-FDR
DR measures an interval with a ratio approach that can also be described by a ratio of maximal binary number to the smallest (of course 1). Dual sampling technique has inspired this FDR approach however the whole imager will benefit from the extra binary bits (DR enhancement) which also cost the resultant image to be in a large size. Hence, due to philosophy statement 2: only region of interest is given the appropriate attention and resources (memory).
The real time acquisition of the system was realized by setting Labview process to real time within windows2000 environment. Tool used is called TaskInfo.
This is very simplistic model but serving the targeted objectives.
This is for rolling scan (Non-bouncing scan). Tin=124.8 ms, 62.4ms, 49.92ms, 24.96ms, 12.48ms, 6.24ms, 2.496ms, and 1.248ms for 10KHz, 20KHz, 25KHz, 50KHz, 100KHz, 200KHz, 500KHz and 1MHz respectively or 8Fps, 16Fps, 20Fps, 40.1Fps, 80.1Fps. 160.Fps, 400.6Fps, 801.3Fps (NB-FRAMErate = SPLfreq/1248;BS-FRAMErate = SPLfreq / 2496) Model speaks through the correlation values: 1- When noise discarded the model failed miserably.
CalculDR : is the system DR which is based on reachable Lmax and detectable Lmin (shared for all Tint –see slide 21-22) OptDRMx_MnDet : 1- to show the linearity of the system Max Detectable Light Vs Tint 2-That this approach is Dynamic range enhancement towards brighter light intensities. Theory_DRenh: The formula used to expect the dynamic range enhancement (from dual sampling DR enhancement)
Raster scanning is just an extreme case of the Rolling shutter.
Tint is the same for both, therefore its should be faire parameter: Shows the impact of the sampling architecture Imagine 2 pointers reading the 2 images simultaneously. Think of it as film image integration The reason of the blur is therefore a result of the architecture sampling.
Use PowerPoint to show DR by changing the contrast or brightness The first set acquired at: 270 lux, 8 fps (10KHz) The second set sampled at: 402 lux, 400 fps (1MHz)
Because R 2 is the dominant player even without assuming Tspl=3Ts the results is likely not affected.
Because R 2 is the dominant player even without assuming Tspl=3Ts the results is likely not affected.
The present analysis shows that pyramidal CMOS image sensor is always faster than classical CMOS imager of equivalent size and scanning timing parameters. However, this high speed feature is more prominent in the case of acquisition segmentation into parallel readout channels, as originally designed, than in the case of serial readout. To conclude this section, its is worth mentioning that parallel segment readout scheme of the pyramidal imager is the readout of choice not only because of the fast frame rates its achieved, furthermore, because it is the most natural scheme to this architecture due to the fact that every segment has its own independent sample and hold banks of capacitors.
From top-right picture: Oblique bandwidth is about ½ of the cardinal bandwidth. It shows the Coincidence or Adaptation of the HVS spatial resolution power bandwidth to that its surrounding nature.
Fixed pattern noise (FPN) is the spatial noise distribution with no illumination of the image sensor array that is explicitly time independent, and hence “fixed” Causes: 1- photo-signal generation mismatch (variation of the photo-sensing areas -mask & process variations-, dark current: impurities, thermal generation, lattice defects) 2- photo-signal transportation mismatch (S&H mismatch, column amplifiers mismatch) This correlated double sampling suppresses reset noise, 1/ f noise, and Fixed Pattern Noise (FPN) due to threshold voltage variations.
Filter D and E for all 8 direction to mimic the circular shape of the filter and ADD F (half the bandwidth) to the cardinal axes to implement the oblique effect Filters D, E and F were chosen because: there expected to introduce the most significant response based on their Peak Frequency.
Using empirical data and well accepted spatial filter suggested by prof. Hugh at York University a modeled spatial filter was built. The oblique effect is introduced based on the experimental facts shown in slide 32 (bandwidth fact)
The Fourier spectrum modules were shown with modules sliced between 0 3(or4) to see Matlab was used. Inverse FFT show no discernable differences between B and C used their difference.
Color-map used full range of the difference : information extraction.
Information content in the spatial frequencies can be preserved if kernel averaging is done adequately. Could save large amounts of image data that is not needed and allocate those resources to high spatial frequency content areas.