Step 1 - Camera Selection 0:38 Smart camera - brains and processing in the camera Dumb camera - brains and processing in separate vision processor Cost rule of thumb: if you have 3 cameras or more, it’s more cost effective to use 3 dumb cameras with an external vision processor Not all manufacturers make both types - sometimes choice of manufacturer makes this decision for you Step 2 - Questions to ask for a new vision application 2:35 Is an industrial camera capable of seeing this? What do you want to measure? What’s the tolerance range of the measurement? How good is the lighting? Step 3 - Lighting 5:25 Highlights the things we care about, and hides the things we don’t Angle: side light shows raised surfaces well, hides print Color: Using a green light means the camera can only see the red on a green and red label Shadow/backlighting: shows part edges well, hides details on the surface Step 4 - Order parts 5:10 Suppliers will help you spec parts for your application Step 5 - Program 7:07 Step 6 - Run time 7:23 Testing during run time is the only way to know it’s working Longer run times are recommended to account for unknown changes that will happen (ie: lighting changes because there is a window in the room)
Wondering what camera would be most appropriate for capturing images of large boards (as large as floorboards)? Currently, my customer is using a normal mirror camera mounted at 3m above the table. You can imagine how inconvenient it is to mount a camera, take images, unmount it, and transfer the images through USB. Afterward, the images are passed to custom OpenCV-based shape detection software. So they don't need any smart camera but just something that would work with such large boards and would connect and be controllable over a cable, so it can be mounted once and not touched anymore. I think GigE Vision with PoE might be a good choice, but I'm completely clueless about camera brands and quality. The customer is a startup, and the budget is very limited; definitely, it would be good to keep the cost of the camera setup under 500$.
The size of the part doesn't determine what camera to get. The important thing is putting the right lens on the camera to see the whole part. After that, the question is how big are the features you're looking for? You'll want a high enough resolution camera to see the important features with the required precision. I've got an article here that explains the math and some rules of thumb: www.breen-machine.com/robust-vision-inspection-in-5-steps/ Regarding GigE/PoE, that's usually a good option, and you should be able to get a camera/lens/cables for $500 if the resolution requirement is small. I don't have a specific brand recommendation, but the good cameras will typically have a brand named sensor (Sony, etc). An alternate thought, you say the customer is a startup. If they're developing a product that they'll sell a lot of, it might make sense to develop things further. Start with a proof of concept with good quality, off the shelf parts (like you're talking about), then start slimming things down. It's possible you can even get a webcam to work for the application. If there's enough budget, try a few cameras and see what works well at the best price.
@@breen-machine Thank you for the valuable advice and the article. This company now had two satisfied customers, and so they feel ready to invest in optimizations to reduce manual fiddling with the off-the-shelf camera. The recognition requirements are simple, we need only to accurately (1-2mm precision) recognize the shape of the board on solid color background in a single-shot image. So, I guess even a cheap machine vision camera with a good sensor (Sony as you mentioned) should be enough. Regarding the lenses, there is also one more important factor - it's compensating for lens distortion (usually barrel). I'm wondering how do typical machine vision cameras and their lenses deal with this? Are there any compatibility rules or correction profiles that can be set on the camera itself or is it always the job of programmers to post-process the captured image to remove the lens distortion? I know OpenCV has good calibration tools to determine the required lens correction once and then apply it to all captured images, but it would be much better to get corrected images straight from the camera. And this might be the area where some brands might be more appropriate than others to achieve a hassle-free camera+lens combination. Without knowing the typical caveats and even the right terms (besides pixel resolution and lens focus distance) it is very difficult to make the right choice immediately. Unfortunately, we are located in a small town and don't know any other factory or consultation company that deals with machine vision, so it feels somewhat scary to make the right choice on our own.
How big are the boards you're measuring? If they're small, 1-2mm precision is easy to get. If they're large (like a 4x8 ft sheet of plywood), you'll need a high resolution camera to measure that precisely. That's a good point. Compensating for lens distortion is important for measurement applications. The calibration tools you're talking about (a grid or checkerboard pattern is usually the best) will have to be used every time you change the camera mounting. It sounds like you'll be moving this camera to different job sites, maybe mounted on a tripod. If that's correct, you'll have to recalibrate every time you set up. I'm not aware of any cameras that will post process an image before sending it to the controller. I'd recommend trying the OpenCV options.
Changing the JOB on practically every Cognex camera takes a very long time (up to 10 seconds). The profinet communication frame itself is exceptionally complex and consists of numerous conditions. The cameras from the IS2800 series are a complete misunderstanding altogether. The software freezes, is not well-developed, and Cognex panels were not compatible (and probably still aren't). I definitely prefer using Keyence products (IV, IV2, IV3, and IX cameras depending on the needs). Due to all these issues with Cognex, I have developed an aversion :) @@breen-machine
Step 1 - Camera Selection 0:38
Smart camera - brains and processing in the camera
Dumb camera - brains and processing in separate vision processor
Cost rule of thumb: if you have 3 cameras or more, it’s more cost effective to use 3 dumb cameras with an external vision processor
Not all manufacturers make both types - sometimes choice of manufacturer makes this decision for you
Step 2 - Questions to ask for a new vision application 2:35
Is an industrial camera capable of seeing this?
What do you want to measure?
What’s the tolerance range of the measurement?
How good is the lighting?
Step 3 - Lighting 5:25
Highlights the things we care about, and hides the things we don’t
Angle: side light shows raised surfaces well, hides print
Color: Using a green light means the camera can only see the red on a green and red label
Shadow/backlighting: shows part edges well, hides details on the surface
Step 4 - Order parts 5:10
Suppliers will help you spec parts for your application
Step 5 - Program 7:07
Step 6 - Run time 7:23
Testing during run time is the only way to know it’s working
Longer run times are recommended to account for unknown changes that will happen (ie: lighting changes because there is a window in the room)
You are truly putting in your great efforts. I can say it from the video and this comment.
It helps a lot.
Thank you :)
Really nice video. As someone just dipping their toes into machine vision, your clear, concise presentation of the material is very useful.
Glad you enjoyed it!
Yeah straight to the point and useful!
looking forward to go through all your videos on Machine Vision.
thank you for this video are you working in other types of camera like Omron F430 or V430 ,
Thank you very much, Breen!
Very informative, and easy to follow!
Great to hear!
Wondering what camera would be most appropriate for capturing images of large boards (as large as floorboards)? Currently, my customer is using a normal mirror camera mounted at 3m above the table. You can imagine how inconvenient it is to mount a camera, take images, unmount it, and transfer the images through USB. Afterward, the images are passed to custom OpenCV-based shape detection software. So they don't need any smart camera but just something that would work with such large boards and would connect and be controllable over a cable, so it can be mounted once and not touched anymore. I think GigE Vision with PoE might be a good choice, but I'm completely clueless about camera brands and quality. The customer is a startup, and the budget is very limited; definitely, it would be good to keep the cost of the camera setup under 500$.
The size of the part doesn't determine what camera to get. The important thing is putting the right lens on the camera to see the whole part. After that, the question is how big are the features you're looking for? You'll want a high enough resolution camera to see the important features with the required precision. I've got an article here that explains the math and some rules of thumb: www.breen-machine.com/robust-vision-inspection-in-5-steps/
Regarding GigE/PoE, that's usually a good option, and you should be able to get a camera/lens/cables for $500 if the resolution requirement is small. I don't have a specific brand recommendation, but the good cameras will typically have a brand named sensor (Sony, etc).
An alternate thought, you say the customer is a startup. If they're developing a product that they'll sell a lot of, it might make sense to develop things further. Start with a proof of concept with good quality, off the shelf parts (like you're talking about), then start slimming things down. It's possible you can even get a webcam to work for the application. If there's enough budget, try a few cameras and see what works well at the best price.
@@breen-machine Thank you for the valuable advice and the article.
This company now had two satisfied customers, and so they feel ready to invest in optimizations to reduce manual fiddling with the off-the-shelf camera. The recognition requirements are simple, we need only to accurately (1-2mm precision) recognize the shape of the board on solid color background in a single-shot image. So, I guess even a cheap machine vision camera with a good sensor (Sony as you mentioned) should be enough.
Regarding the lenses, there is also one more important factor - it's compensating for lens distortion (usually barrel). I'm wondering how do typical machine vision cameras and their lenses deal with this? Are there any compatibility rules or correction profiles that can be set on the camera itself or is it always the job of programmers to post-process the captured image to remove the lens distortion? I know OpenCV has good calibration tools to determine the required lens correction once and then apply it to all captured images, but it would be much better to get corrected images straight from the camera. And this might be the area where some brands might be more appropriate than others to achieve a hassle-free camera+lens combination. Without knowing the typical caveats and even the right terms (besides pixel resolution and lens focus distance) it is very difficult to make the right choice immediately. Unfortunately, we are located in a small town and don't know any other factory or consultation company that deals with machine vision, so it feels somewhat scary to make the right choice on our own.
How big are the boards you're measuring? If they're small, 1-2mm precision is easy to get. If they're large (like a 4x8 ft sheet of plywood), you'll need a high resolution camera to measure that precisely.
That's a good point. Compensating for lens distortion is important for measurement applications. The calibration tools you're talking about (a grid or checkerboard pattern is usually the best) will have to be used every time you change the camera mounting. It sounds like you'll be moving this camera to different job sites, maybe mounted on a tripod. If that's correct, you'll have to recalibrate every time you set up.
I'm not aware of any cameras that will post process an image before sending it to the controller. I'd recommend trying the OpenCV options.
CV nxc😮
I hate cognex cameras so bad... trust me
Haha, I trust you. People on the internet are always very honest about the things they hate. ;)
What do you prefer?
Changing the JOB on practically every Cognex camera takes a very long time (up to 10 seconds). The profinet communication frame itself is exceptionally complex and consists of numerous conditions. The cameras from the IS2800 series are a complete misunderstanding altogether. The software freezes, is not well-developed, and Cognex panels were not compatible (and probably still aren't).
I definitely prefer using Keyence products (IV, IV2, IV3, and IX cameras depending on the needs). Due to all these issues with Cognex, I have developed an aversion :)
@@breen-machine
L