EEVblog Electronics Community Forum
Products => Thermal Imaging => Topic started by: publioelon on August 02, 2021, 08:27:57 pm
-
Hi, I'd like to ask for help in choosing the best dongle-type thermal camera for android devices. I would like to detect face temperature changes, specifically using dongle-type cameras. From what I've been looking, there are options like HTi-201, Hi-301, Infiray TS3 Pro, Seek Compact, and other variations. For the specific task of choosing a dongle-type thermal camera, which camera is the most reliable and in addition will be good for face temperature inspection?
-
Some initial questions ....
Do you want to measure absolute accuracy, or just to compare ?
How accurate do you need to be ?
Do you want to measure 'face skin surface temperature' or are you really hoping to measure body (medical) temperature ?
As a summary most cameras cannot measure better than 2°C, and dongles can be twice that or worse.
There is no valid relationship between apparent skin surface temperatures and body temperature, too many variables of sweat, air flow etc. At best you get 'I might want to measure that hot person properly' but nobody owns up to how many with internal high temperatures get past a surface temperature check.
Bill
-
Hi, thanks for your reply. I need more image quality and resolution than temperature values, I am planning to use a thermometer for referencing. I plan to use it for feature extraction using CNNs, therefore i will stick to just grabbing frames from the face. If possible, maybe i'd like to observe some veins on the face too.
-
If you want high resolution, both in terms of pixels and dynamic range, look at the higher end: ThermApp Pro, Thermal Expert V1, You can get VGA resolution but it comes at a hefty price. And you should really question why you are using a dongle in the first place here.
If you want to do a learning based project, there is a database available of a few hundred faces from a selection of angles. But I am not sure if that is radiometric.
-
If you want high resolution, both in terms of pixels and dynamic range, look at the higher end: ThermApp Pro, Thermal Expert V1, You can get VGA resolution but it comes at a hefty price. And you should really question why you are using a dongle in the first place here.
If you want to do a learning based project, there is a database available of a few hundred faces from a selection of angles. But I am not sure if that is radiometric.
I'd like to observe features related to fever symptoms. The Hti-301 option looked interesting it's very high resolution and definition to a very reasonable price. I was wondering if the images are in form of raw data, radiometric and wether they are 8 bit or 14 bit. I am very inclined towards this option. I wasn't able to find any camera that surpasses the HTI-301 quality , to the best of my knowledge there isn't criticism around this camera for some reason.
I wonder if the calibration shows accurate temperature readings?
-
9 pages of discussion of the HTI here:
https://www.eevblog.com/forum/thermal-imaging/hti-ht-301-384-x-288/50/ (https://www.eevblog.com/forum/thermal-imaging/hti-ht-301-384-x-288/50/)
Includes many real user comments and images. Plus points are resolution, minus the software and oversharpening. Accuracy is what we'd expect, around +/-3°C and drifts about as the dongle warms up.
Bill
-
This thread and its links may be of interest....
https://www.eevblog.com/forum/thermal-imaging/thermal-image-dataset-for-fever-and-human-temperature-measurement/msg2981274/#msg2981274 (https://www.eevblog.com/forum/thermal-imaging/thermal-image-dataset-for-fever-and-human-temperature-measurement/msg2981274/#msg2981274)
The use of thermal imaging cameras for fever detection has been discussed previously. It is not a simple task for a thermal camera due to the accuracy of measurement required. Unless a thermal camera is described as fully radiometric and inetended for accurate temperature measurement it is unlikley to be adequate for the task. Many generic dongle type thermal cameras suffer from less than stellar measurement accuracy and no claims of great accuracy are made by the manufacturer. Most fever detection systems require the use of a knowon accurate emission black body reference that the camera monitors within the scene to improve measurement calibration and resultant accuracy.
The known IR measurment target points for human head temperature measurements are , Tympanic membrane in ear canal, inside the mouth, specifc temporal arteries near the surface of the skin near the temple, the inner corner of the eyes adjacent to the bridge of the nose nose. These test points are known to have an excellent 'fresh' blood supply from the bodies core. The forehead is far from ideal as a reference point but is commonly used for IR thermometer tests. It is a large surface area, with minor blood flow from the core and subject to temperature variation due to sweat and other unpredictable influences. When requiring a likley core temperature reading, the measurment system of the thermal camera needs to take account of the difference between the measured temerature at the location on the head and true core temperature.
Fraser
-
Discussion of the Seek Scan fever detection camera solution....
https://www.eevblog.com/forum/thermal-imaging/seek-scan/msg2989554/#msg2989554 (https://www.eevblog.com/forum/thermal-imaging/seek-scan/msg2989554/#msg2989554)
-
Discussion of the Seek Scan fever detection camera solution....
https://www.eevblog.com/forum/thermal-imaging/seek-scan/msg2989554/#msg2989554 (https://www.eevblog.com/forum/thermal-imaging/seek-scan/msg2989554/#msg2989554)
Thanks for the links! I've been carefully reading all the materials posted in this thread so far, I feel more educated about the technical details. I've come up with question in regards to the Seek Compact Pro device and the Infiray line.
From what I could gather from the HTI-301:
- It needs a 10 minute time to warm up to avoid temperature decay overtime.
- It freezes every 3 minutes to calibrate the temperature.
- The resolution is not raw when the image is opened in a PC.
- The bolometer used is from Infiray company.
- It uses to much shaperning to make improve the image quality based on the FLIRs MSX.
- Accuracy is precise.
- Good image quality.
- Stores image to the phone embedded memory.
- The app is mediocre, needs an alternative solution.
- The bolometer is actually 128x160 but the image is processed to increase the resolution (can anybody confirm this?).
Seek Compact Pro:
- constant Flickering
- Low image quality despite resolutin
- Manual focus adjustment
- Reasonable price for resolution
So, apparently a popular choice is the HTi-301, but the Infiray t3pro and T3S are also powerful options. Out of these cameras which one can be used with android and is more suitable for temperature? The more i dig, the more questions I have. The seek modules are often returned from the topics i read. Should I go for a Hti-301 or Infiray line? For some reason the HTi-301 image looks better, But I wonder if the Infiray modules have the same problem as HTi-301 given that they are the same bolometers.
-
HTI-301
- It needs a 10 minute time to warm up to avoid temperature decay overtime.
Many modern thermal imaging cores have a warm up time to reach thermal equilibrium as they are not temperature stabilised. The measurement accuracy is better after a decent warm-up time. 10 minutes sounds about right for a dongle camera. Do not think for a moment that just because of a warm up period, a dongle camera provides good measurement repeatability however. I do not recall the HT-301 being decribed as a very accurate measurement device. This is especially so where any medical imaging is involved!
- It freezes every 3 minutes to calibrate the temperature.
Normal behaviour of a microbolometer based thermal imaging camera system that uses a FFC flag for self calibration and FFC. Any camera that does not use an FFC flag for calibration, uses software techniques to achieve good FFC but the measurement accuracy is often compromised due to the loss of a calibration reference flag.
- The resolution is not raw when the image is opened in a PC.
The image is what is produced by the Host software, either Radiometric or non Radiometric, depending upon the specific Donge camera. RAW data where thermal imaging is concerned is a complex topic. A truly RAW image would be pretty ugly and not easily used. The image undergoes a minimum of dead pixel correction, Non uniformity correction and flat field correction. Once these processes have been completed the image becomes useable but is not truly "RAW".
- The bolometer used is from Infiray company.
Yes, IRay.
It uses to much shaperning to make improve the image quality based on the FLIRs MSX.
Nothing like FLIR's MSX that uses a visible light edge captured overlay. Image enhancement is common on modern thermal imaging cameras and can often improve the image in terms of interpretation by the user and 'artistic' appearance. Over processing can make the thermal image look strange however. Be aware that image enhancement can effect radiometric accuracy in some cameras as the pixel data is being artificially manipulated and much depends upon how the measurement system operates. Advanced noise reduction is also employed in modern microbolometer based systems and this can produce very clean images but can also reduce image sharpness and update rate in some cases.
- Accuracy is precise.
Accuracy and repeatability of measurement are are a complex subject in thermal imaging cameras. You normally get exactly what you pay for ! I cannot cover the whole topic here but Dongle cameras are not well known for their accuracy or repeatability of measurement and are not normally considered a precision temperature measurement tool, especially if used in medical work and not 'tuned' for the task. Remember, most thermal imaging cameras state a measurement tolerance of +/-2C or +/-2% (whichever is greater) and many dongles are +/-3C or +/-3% with some being +/-5C or +/5%. Not really what you would call a precision measuring system in the conventional sense of the word. The tolerance is often stated at a specific ambient temperature.
- Good image quality.
That depends upon your needs as pretty pictures are not always what you need for acurate radiometric imaging. It is the accurate pixel data that counts. That is not to say that you cannot have a pretty picture and accurate pixel measurement data as well.
- Stores image to the phone embedded memory.
This is normal behaviour for a dongle type phone camera. The actual dongle is like a software defined camera and the Host does the heavy lifting and storage.
- The app is mediocre, needs an alternative solution.
Sadly it is true that many quite good thermal imaging dongles suffer from less than wonderful Host software. Third parties and enthusiasts often write their own software using an SDK or reverse engineering of the communications protocol. Note that Dongle cameras do not always provide deep access into the image data and present a 'processed' set of data out of the provided hardware-host driver.
- The bolometer is actually 128x160 but the image is processed to increase the resolution (can anybody confirm this?).
I have never before heard this and it does not reflect what I have seen of the HT-301 imagery. Interpolation has been used to increase displayed image resolution but I do not see this with the HT-301 and do not believe this to be the case. It can be challenging to determine true resolution when heavy image processing has been employed but I cannot see that reducing apparent resolution. What is your source for this claimed 128 x 160 pixel resolution please ?
-
HTI-301
- It needs a 10 minute time to warm up to avoid temperature decay over time.
Many modern thermal imaging cores have a warm-up time to reach thermal equilibrium as they are not temperature stabilized. The measurement accuracy is better after a decent warm-up time. 10 minutes sounds about right for a dongle camera. Do not think for a moment that just because of a warm-up period, a dongle camera provides good measurement repeatability, however. I do not recall the HT-301 being described as a very accurate measurement device. This is especially so where any medical imaging is involved!
- It freezes every 3 minutes to calibrate the temperature.
Normal behavior of a microbolometer-based thermal imaging camera system that uses an FFC flag for self-calibration and FFC. Any camera that does not use an FFC flag for calibration uses software techniques to achieve good FFC. Still, the measurement accuracy is often compromised due to the loss of a calibration reference flag.
- The resolution is not raw when the image is opened in a PC.
The image is what is produced by the Host software, either Radiometric or non Radiometric, depending upon the specific Donge camera. RAW data where thermal imaging is concerned is a complex topic. A truly RAW image would be pretty ugly and not easily used. The image undergoes a minimum of dead pixel correction, Nonuniformity correction and flat field correction. Once these processes have been completed, the image becomes useable but is not truly "RAW".
- The bolometer used is from Infiray company.
Yes, IRay.
It uses to much sharpening to make improve the image quality based on the FLIR MSX.
Nothing like FLIR's MSX that uses a visible light edge captured overlay. Image enhancement is common on modern thermal imaging cameras and can often improve the image in terms of interpretation by the user and 'artistic' appearance. Over-processing can make the thermal image look strange, however. Be aware that image enhancement can affect radiometric Accuracy in some cameras as the pixel data is being artificially manipulated and much depends upon how the measurement system operates. Advanced noise reduction is also employed in modern microbolometer-based systems and this can produce very clean images but can also reduce image sharpness and update rate in some cases.
- Accuracy is precise.
Accuracy and repeatability of measurement are are a complex subject in thermal imaging cameras. You normally get exactly what you pay for ! I cannot cover the whole topic here but Dongle cameras are not well known for their Accuracy or repeatability of measurement and are not normally considered a precision temperature measurement tool, especially if used in medical work and not 'tuned' for the task. Remember, most thermal imaging cameras state a measurement tolerance of +/-2C or +/-2% (whichever is greater) and many dongles are +/-3C or +/-3% with some being +/-5C or +/5%. Not really what you would call a precision measuring system in the conventional sense of the word. The tolerance is often stated at a specific ambient temperature.
- Good image quality.
That depends upon your needs as pretty pictures are not always what you need for accurate radiometric imaging. It is the accurate pixel data that counts. That is not to say that you cannot have a pretty picture and accurate pixel measurement data as well.
- Stores image to the phone embedded memory.
This is normal behaviour for a dongle type phone camera. The actual dongle is like a software defined camera and the Host does the heavy lifting and storage.
- The app is mediocre, needs an alternative solution.
Sadly it is true that many quite good thermal imaging dongles suffer from less than wonderful Host software. Third parties and enthusiasts often write their own software using an SDK or reverse engineering of the communications protocol. Note that Dongle cameras do not always provide deep access into the image data and present a 'processed' set of data out of the provided hardware-host driver.
- The bolometer is actually 128x160 but the image is processed to increase the resolution (can anybody confirm this?).
I have never before heard this, and it does not reflect what I have seen of the HT-301 imagery. Interpolation has been used to increase displayed image resolution, but I do not see this with the HT-301 and do not believe this to be the case. It can be challenging to determine true resolution when heavy image processing has been employed, but I cannot see that reducing apparent resolution. What is your source for this claimed 128 x 160 pixel resolution please ?
Thanks again for the reply, Just adding some information. I've researched more information about the Seek Compact Pro, found some of your old posts: https://www.eevblog.com/forum/thermal-imaging/whats-going-on-with-seek-thermal-compact-pro/ (https://www.eevblog.com/forum/thermal-imaging/whats-going-on-with-seek-thermal-compact-pro/)
I absolutely did not know that FLIR sued SEEK and every 15th pixel was blinded, what happens in practice to the bolometer and the image quality? Forgive my ignorance, I am still learning about thermal cameras. Another relevant information is the pixel size, from what I was reading
, those smaller types of bolometers are prone to get noisy details in the image. In this instance, the HTI-301 has a bigger pixel unit of 17u m when compared to the other bolometers like the Seek Compact Pro and the HTI-201, does this means the HTI-301 bolometer will suffer less noise in comparison to the Seek Compact Pro? To be honest, I thought the smaller the pixel, the more details the image would have. I read a paper a while ago, it said that the bolometer size is inversely proportional to the pixel size, meaning that bolometers with higher sizes have small pixels and better image definition.
Going back to your posts, your points are extremely interesting and have very helpful information for my goals. Please allow me to ask some questions.
About Accuracy and repeatability of measurement, do you have any materials on this trade-off that may help me with my decision? I couldn't completely understand in practical terms what is the repeatability of measurement other than the literal meaning.
Also, is it possible to detect every temperature value for each pixel in the HTI-301 image using the Hi app, or do I need an alternative, as it is in the FLIR One Pro?
From the looks of it, an Infiray thermal camera is more reliable than the HTi-301 line, then?
Also, as for the bolometer size: I found it on this thread, but apparently, it was a question: https://www.eevblog.com/forum/thermal-imaging/hti-ht-301-384-x-288/200/. (https://www.eevblog.com/forum/thermal-imaging/hti-ht-301-384-x-288/200/.) The 3rd post.
-
I absolutely did not know that FLIR sued SEEK and every 15th pixel was blinded, what happens in practice to the bolometer and the image quality? Forgive my ignorance, I am still learning about thermal cameras. Another relevant information is the pixel size, from what I was reading
, those smaller types of bolometers are prone to get noisy details in the image. In this instance, the HTI-301 has a bigger pixel unit of 17u m when compared to the other bolometers like the Seek Compact Pro and the HTI-201, does this means the HTI-301 bolometer will suffer less noise in comparison to the Seek Compact Pro? To be honest, I thought the smaller the pixel, the more details the image would have. I read a paper a while ago, it said that the bolometer size is inversely proportional to the pixel size, meaning that bolometers with higher sizes have small pixels and better image definition.
The blinded 15th was only on early SEEK devices, anything new now should be normally functional, although 1% dead is not unusual. The dead pixels are substituted in software.
Generally readout and signal are getting better and that is allowing smaller pixels to reduce die size and so cost. It is almost a case of constant noise as no-one could sell say a 8um pixel detector that had an NETD (nosie) of 200mK.
It is not a case of constant die size so more pixels 'come for free'.
Bill
-
I absolutely did not know that FLIR sued SEEK and every 15th pixel was blinded, what happens in practice to the bolometer and the image quality? Forgive my ignorance, I am still learning about thermal cameras. Another relevant information is the pixel size, from what I was reading
, those smaller types of bolometers are prone to get noisy details in the image. In this instance, the HTI-301 has a bigger pixel unit of 17u m when compared to the other bolometers like the Seek Compact Pro and the HTI-201, does this means the HTI-301 bolometer will suffer less noise in comparison to the Seek Compact Pro? To be honest, I thought the smaller the pixel, the more details the image would have. I read a paper a while ago, it said that the bolometer size is inversely proportional to the pixel size, meaning that bolometers with higher sizes have small pixels and better image definition.
The blinded 15th was only on early SEEK devices, anything new now should be normally functional, although 1% dead is not unusual. The dead pixels are substituted in software.
Generally readout and signal are getting better and that is allowing smaller pixels to reduce die size and so cost. It is almost a case of constant noise as no-one could sell say a 8um pixel detector that had an NETD (nosie) of 200mK.
It is not a case of constant die size so more pixels 'come for free'.
Bill
Cool, so let's say I want to classify temperature changes using CNNs, from thermal images specifically acquired by a mobile phone. Do I need a more robust camera with more pixels like a Infray T3 Pro/Hti-301 or can I use cheaper models such as Hti-201 and Infiray Search P2?
-
It's now about the number or quality of pixels. You need (a known source and) a histogram. You can easily program an alarm function and don't need any learning based models.
Most radiometric cameras will have such a feature built in already.
If you could explain more specifically what you want to achieve, it might help with better advice.
-
It's now about the number or quality of pixels. You need (a known source and) a histogram. You can easily program an alarm function and don't need any learning based models.
Most radiometric cameras will have such a feature built in already.
If you could explain more specifically what you want to achieve, it might help with better advice.
Well, to be more specific. I have a budget of 800$ and I'd like to get a camera that gives me exactly what is promised by the device specs. Initially, i was confident tempeature could be measured by the camera with a reasonable accuracy, now I know I was dreaming, its hard to get a calibrated measurement, and on top of that it is even harder to have precision. My end goal precisely is to use machine learning for specific symptons observed from the temperature changes. It is more important to me that i have feasible dongle camera for smartphones that gives me good features for deep learning models. In other words, I'd like to observe temperature detection in the images and only if it was possible using the radiometric data. I apologize for not being specific at the first post.
-
For 800$ you can consider thermal expert q1, I still happen to have it with both lenses as well as seek compact pro therm-app and hti-301 :palm:. about lower resolution in HTI its a lie, I'm 99% sure. you just put q1 13mm lens on hti and if it has same angle and depth of field, sensors are the same(+-algorithms) - i'm too sure and lazy to conduct this experiment for now, and especially just because of someone's even more lazy assumption. Just believe that making fast-booting 38(34)/25m sensors, obtaining great enough sharpness from them to fake 17/12m and writing on a website that you succeeded in making 10-8m sensors sounds more than crazy even for Russian))
upd: checked thermal expert website, they added te-q2 for fever scan and te-sq1 for working without smartphone, no price info though
-
To take an off-the-shelf camera and attempt to use it for fever-screening is asking for a lot of problems. I know - I designed one at the start of Covid season using a crappy thermal module, and still managed to get the accuracy/repeatability good enough to get FDA approval! (Personally I still think thermal camera fever screening is mostly :bullshit:)
The accuracy/repeatability and noise of most cameras just isn't good enough, but there are things you can do to improve your data if you have access to the data at the correct level. These aren't the sorts of things you can do if you only get a colour jpeg out of the camera!
The easiest is to reduce noise by performing temporal averaging across several frames. This presents its own problems as your target is now no longer in sharp focus if they are moving - something that can be partially remedied with a big sign that says 'Stay Still'! Higher the camera frame rate the more averaging you can do without too much image blurring. The averaging should be done in either the energy or temperature domains.
To improve repeatability, put a fixed reference in the scene. It’s a *lot* easier to keep a block of metal at a fixed temperature using a thermistor than it is to calibrate out all the possible conditions that will affect the bolometer! Even a simple single op-amp based control circuit will be able to make a heater that holds it temperature within a few 10ths of a degree. If you trust the scaling of the camera (eg that a 1C difference is actually a 1C difference) then you can get away with a just a single reference in the scene. If you don’t believe the scaling, then you could use two different temperatures in the scene to give you both reference and scale. Make sure the FFC is nice and clean too!
You can probably find cheap 'body temperature' black-body references on ebay by now!
When is comes to the CNN, think carefully about how to normalise the data. I published a paper last year on using CNNs with thermal images for feature detection, but I’m no NN expert so it was a steep learning curve, and I went though many iterations of data normalisation to get the best results. Most image processing ‘examples’ will give you the standard ‘R/255, G/255, B/255’ to give you a 0-1 scaling for the NN input since almost all examples work with colour images as the input. Ideally use the radiometric data for the NN input too, don’t use a colourised image – you probably loose useful data by colourising.
-
I narrowed the options to the Hti-301, Infiray T3S and T3S Pro. I would like to know which one of these is the most reliable for good image quality and has nearly as many options as the FLIR devices