On camera resolution and sensitivity for a PIV experiment

[Erich's Lab]

Hi all, I hope you are all having a good day. Below is a problem I am currently facing before I construct a low cost, decent accuracy planar PIV system for under $2,000 USD.

TL, DR:

I was wondering what resolution a camera should be for 200 +/- 100 μm helium-filled soap bubbles and a 15 x 10 cm area of interest. The following resolutions are of question:

 - 14040 x 1080 px @ 220 fps, ~104.2 μm/px  x ~92.6 μm/px resolution

 -  1920  x 1200 px @ 120 fps, ~78.1 μm/px x ~83 μm /px resolution

 -  2600 x 2160 px @ 60 fps,     ~57.7 μm/px x ~46.3 μm /px resolution

Detailed version:

Recently, I started researching into double and triple exposure, color PIV to circumvent the limitations of high interframe times in industrial and machine vision cameras. This requires a color image sensor which is known to be less sensitive compared to monochromatic sensors. This is due to smaller wells in the sensor along with a color filter (often a bayer filter) which rejects photons of a specific wavelength range. In essence, less photons are converted into electrons and as a result, the quantum efficiency, signal to noise ratio, and dynamic range suffers especially for blue and red wavelengths. This means that a higher powered light source is needed to illuminate the seed particles for a given exposure,. In my case, I will simply increase the amount of LEDs to 18 CBT 90 LEDs and collimated them into a 0.5 - 1 mm light sheet with moderate divergence using fiber optics and a cylindrical lens (overall optical efficiency is estimated to be ~15%, but the light output would be significantly more powerful than efficiently coupling a single LED).  However, this begs the question on how sensitive the camera sensor has to be to properly detect and resolve seed particles. In my current and ongoing experiment, two seed generators have been designed based on cavitation and helium-filled soap bubbles. The first method generates approximately 50 μm +/- 20 μm soap bubbles using cavitation via rapid and controlled depressurization of a soap-water mixture. The second method uses a nozzle constructed through additive manufacturing to controllably generate helium-filled soap bubbles of ~200 μm +/- 100 μm. The latter method was chosen due to concerns governing required resolution and the higher ability to scatter light. 

Moving forward, 3 industrial cameras were selected as possible candidates for this experiment. These cameras were selected based on cost, resolution, and apparent sensitivity. Their characteristics are detailed below.

Basler daA1440-220uc

  resolution: 1440x 1080 px

  pixel well size: 3.45 μm

  fps: 220

  monochromatic quantum efficiency (uncooled): ~62.7%

  monochromatic dynamic range: ~71.5 dB

  monochromatic s2n ratio: ~40.2 dB

Basler a2A1920-51ucBAS

  resolution: 1920 x 1200 px

  pixel well size: 3.45 μm

  monochromatic quantum efficiency (uncooled): ~62.2%

  monochromatic dynamic range: ~71.7 dB

  monochromatic s2n ratio: ~40.2 dB

Basler a2A2600-64ucBAS

  resolution: 2600 x 2160 px

  pixel well size: 2.5 μm

  fps: 64

  monochromatic quantum efficiency (uncooled): ~53.6%

  monochromatic dynamic range: ~71 dB

  monochromatic s2n ratio: ~36.7 dB

Honorable mentions:

Basler acA1920-40uc

  resolution: 1920 x 1200 px

  pixel well size: 5.86 μm

  fps: 41

  monochromatic quantum efficiency (uncooled): ~70%

  monochromatic dynamic range: ~73.5 dB

  monochromatic s2n ratio: ~45.2 dB

Basler acA2440-35uc

  resolution: 2448x 2048 px

  pixel well size: 3.45 μm

  fps: 35

  monochromatic quantum efficiency (uncooled): ~68%

  monochromatic dynamic range: ~73.4 dB

  monochromatic s2n ratio: ~40.2 dB

Note:

Monochromatic characteristics are used to give a rough idea of how sensitive the camera could be. For color sensors, these metrics do not apply..

Another honorable mention would be the Optolution OptoCam 2.3 mp camera with a 65 μs interframe time. This is the most optimal solution as it would remove the need for multiple color exposures as well as obtain a much higher sensitivity to light, but that device is sadly outside of my current budget for a camera ($850 USD).

My biggest concern is the resolution needed for properly resolving seed particles of 200 μm +/- 100 μm. Ignoring effects of light scatter enlarging sensed particles, a resolution of ~95 μm/px seems appropriate for resolving 200 μm +/- 100 μm particles as this should generate particles of 3 to 4 pixels. However, the particles are more likely to appear larger due to effects of scattered light which could double or even triple the imaged particle size, requiring larger areas of interest to keep the particles around 3 pixels (which is fine for this experiment). Another option is to use the scattered light of 50 μm particles (most optimal as they have a Stokes number of around 0.2) and rely on the camera's ability to detect particles below its resolution, but that would require possibly higher camera resolutions. This leaves me conflicted if I should use a lower resolution camera with larger seed particles or a higher resolution camera with smaller seed particles (that more faithfully follow fluid trajectories) and hope that they can be properly resolved by the camera. Of course, there are several algorithms to increase the particle size if needed, but I would assume it would be more beneficial to have particles in the range of 2 to 4 pixels before applying DSP algorithms.

All contributions are very welcomed.

Kind Regards,

Erich

[Alex Liberzon]

Hi Erich, 

You have a complex question here and a lot of helpful information. 

Regarding the bubbles - as far as I remember their images are not one bright blob (like the solid particles or small aerosol droplets), but rather a double dots (glare points?) view, like soap bubbles - you see where the light enters and leaves the bubble thin film. This means that the bubble diameter is a distance of two glare points, and the resolution has a different meaning here. also, the laser scattering pattern is not what we typically know from the book. Take a look at some recent works on HFSB to see what they recommend. 

From our side - I'd be very happy to learn about your two bubble devices, especially if these could be replicated - the price for the commercial HFSB generator is quite high. Please share (in private or publicly) what you've done, maybe a blog post with some drawings and tips? This would be a great source, I'm sure like all the other things you do. 

My experience is that color CMOS sensors are about 60% less sensitive than the same model monochromatic sensor. These are double effects—the size of the well and the filter itself. Maybe the situation has improved recently; I haven't checked the new models for a few years. 

I'd go for -  2600 x 2160 px @ 60 fps,     ~57.7 μm/px x ~46.3 μm /px resolution - the 40 fps or 60 fps versus 200 fps are meaningless if you speak of flows you do not plan to use them at this frame rate, right? I'd vote for the sensitivity first and then for resolution. 

By the way, I found a manufacturer, Viewworks, that uses OEM sensors sams as TSI's with a double-shutter option. Custom synchronization makes it possible to drive them in the PIV staggered mode. I'm working with the local machine vision company, and they prepared me a "homemade" PIV system with an 8Mp sensor. It is, however, more expensive than the Optolution option. 

Last idea: Come over for your experiment in my lab. Maybe it's cheaper than building your own lab. Of course, travel is not cheap either. 

Best regards, 

Alex

[Erich's Lab]

Hi Alex,

I finalized my camera selection to the following two cameras:

Basler acA1920-40uc

  resolution: 1920 x 1200 px

  pixel well size: 5.86 μm

  fps: 41

  monochromatic quantum efficiency (uncooled): ~70%

  monochromatic dark noise: 6.9 e-

  monochromatic saturation capacity: 32,800 e-

  monochromatic dynamic range: ~73.5 dB

  monochromatic s2n ratio: ~45.2 dB

FLIR BFS-U3-27S5C-C

  resolution: 1946 x 1464 px

  pixel well size: 4.5 μm

  fps: 95

  monochromatic quantum efficiency (uncooled): ~71%

  monochromatic dark noise: 5.8 e-

  monochromatic saturation capacity: ~25,000 e-

  monochromatic dynamic range: ~72 dB

  monochromatic s2n ratio: Not Specified

Both sensors have high quantum efficiency, low dark noise, and high well capacitance. To be honest, I am leaning towards the FLIR camera for its slightly larger resolution.

On the seeding generators, they are extremely simple to create, just hard to operate and maintain. For the micro bubble generator, it was made out of a used pressure washer with a nozzle and side discharge valve to control cavitation and fluid pressure. The nozzle is 3D printed and casted into aluminum using the lost-foam casting method. the nozzle hole is drilled using a .25 mm unbranded drill bit from an unknown origin. The system is quite crude, but is functional for the most part. I stopped investigating this method for generating seed particles as it requires a resolution of at least 50 μm to properly resolve the seed particles (e.g., a 5 mp sensor with smaller pixel sizes and well capacity). For the helium-filled soap bubble (HFSB) generator, I simply followed the current state-of-the-art modular 3D-printed HFSB nozzle designs and modified them to create slightly smaller seed particles. This method of generating seed particles is attractive to me mainly because of the use of low-pressure systems. However, it is quite finicky to setup since I have no flow meters a hand and the system is set up by observing the dispersion of the HFSB particles.

On your concern governing the halo effect on soap bubbles, this is really not an issue until the particles becomes at least 6 pixels in diameter. Below 6 pixels, they appear as a solid mass akin to micro spheres commonly used to seed water flumes.

Additionally, your remark on the sensitivity of color sensors seem correct. There is a rather significant luminosity difference between each color channel and a monochromatic sensor. Although this is not ideal, the main attractive advantage of color sensors is the ability to color-code three exposures enabling acceleration and pressure measurements regardless of the camera's interframe time. Beyond this, I would have otherwise focused on monochromatic sensors and request numerous evaluation cameras to analyze interframe and blind times for the monochromatic versions of the aforementioned cameras.

Finally, I would love to visit your lab :D

However, I would not be able to do so for some time since I work full-time and am a full-time student.

Best Regards,

Erich

[Erich's Lab]

Hi all, here is an update to my situation.

Due to some circumstances, I decided to snatch a Basler aca1920-40gc camera (used on eBay for $200 USD) for triple color, PIV. The camera has good quantum efficiency for the blue and green wavelengths, but suffers a little on the red wavelength. This effectively limits the PIV system to  two color  PIV experiments since approximately 20% or more red LEDs are needed to mostly balance the histograms of each segregated particle image pairs. However, I ran into an interesting find; the Basler aca1920-155um (~ $680 USD) supports interframe times as low as 150 microseconds with a chance of interframe times being as little as 100 microseconds. As such, I'll be forwarding some research into this camera as well as the two/three color PIV. I think comparing the multi-colored single exposure setup and the standard monochromatic setup could bring some insight into low-cost, high accuracy PIV systems for undergraduate-graduate research in limited funding environments.

Just for kicks, here are the specifications for the camera according to the EMVA1288 specifications

Basler acA1920-150um

  resolution: 1920 x 1200 px

  pixel well size: 4.8 μm

  fps: 150 (75 double-frames/second)

  interframe time: supposedly as low as 80μs, but I have not yet tested this camera

  quantum efficiency: ~50.45%

  dark noise: 10.21 e-

  saturation capacity: 5756 e-

  dynamic range: ~54.6 dB

  s2n ratio: ~37.6 dB

  linearity: horrible; do not recommend the color version of this sensor for TE-CPIV

These specifications look pitiful when compared to something like the Sony IMX series sensors  and PCO Edge cameras (I think they use GPixel GSense sensors), but keep in mind that this camera costs ~$670 USD compared to the PCO cameras (~$3900 to ~$13,000 USD each). Additionally, some commercial PIV suppliers also use this camera for specialized use cases (one used it for 4D lagrangian PTV and supposedly reported interframe times as low as 80 microseconds), so all can't be that bad ;)

To clarify the pixel resolution for the field of view, I used 15cm x 10cm as a standard to compare the different pixel resolutions, albeit ignoring the fact that there are two different resolutions for the x and y axis. The real field of view for experiment would be around 14.4 cm x  9 cm, giving a pixel resolution 75μm/px, but may change later on when I fully design and perform an experiment taking seed particle size into account. 

Best Regards,

Erich