Object Detection

I tried to find a solution using the Edge impulse, but it was a bit difficult to discover such subtle flaws.

@wisam There are solutions for that problem, but since this is obviously a work related question and not a hobbyist question, I will leave it there.

@zander Certainly this is a question related to a factory that has this problem
We are trying to install an ANPR Camera to detect manufacturing defects, but this requires issuing an alert or notification.

If you have any idea or solution please help, thanks.

@wisam Other than ethical issues, there are also huge legal liability issues with a non-employee being involved without a contract in place.

You need to break the process down into very small pieces and then get advice on those. Perhaps start with getting a list of cameras based on your proposed speed of 1mm/msec minimum. What shutter speed is needed, how much light is on the subject. More importantly, what is the area to be examined by a single camera. Will you use 1 computer per camera or more than 1 camera.

Just determining the questions to be answered could be a multi person multi month exercise. There are a lot of what ifs.

Interesting, but definitely a corporate exercise.

@wisam I just noticed that an ANPR camera is only for reading vehicle license plates. What has that to do with defect detection?

@zander 

_First, for the required camera, we chose an ANPR camera because it has a high shutter speed and high FPS, but the software is otherwise, so we are trying to find a smart auxiliary tool to alert when detected.

I think one camera is enough for this proposal.

_Regarding lighting, appropriate lighting will be highlighted under the specified area of ​​the roll, which is estimated at one square metre.

What do you think of the edge impulse platform?

@wisam If your roll is 1M wide, then if the camera has an FOV of 10 cm, you will need 10 cameras. I have no idea if that is accurate, I have nowhere near enough information. This is a SWAG, what you need to know is how many camera pixels are needed per mm^2 to resolve the defects.

Since we have no idea what the camera has to detect, the solutions are many.

If it is related to color, then since electronic cameras only see in B&W that will require a lot of additional processing. I would think along the lines of continuous spectrum analyses.

If it is surface defects where they stick above the avg surface height, an overhead camera with a mylar film on wire frame rotating mirror directing strong light can detect the moving shadows. That will require at least 100 cameras though.

AGAIN, there is a LOT of analyses to be done, picking the hardware first almost always leads to failure. The last time I saw that the company lost $1 Billion and shortly thereafter went bankrupt.

My recommendation is to call a few consultants and tell them all the exact same story. Pick the consultant that asks the most enlightening questions. Expect at least 1 person for 1 monthe to determine the scope, that will very likely result in a minimum of 3 people for 3 to 6 months. After that probably a year to build and test.

'Good luck.

@wisam Your project is NOT an 'edge' device. I have never seen a canned solution work without some amount of modification and that can be more expensive than you might think. Also, now you have to keep your mods working with the everchanging underlying product. If you say you will not upgrade, then you will lose support either explicitly or implicitly generally with decreasing support staff and the use of new hires. Either use the canned solution as intended and change your process to match or go the custom route with full control.

@zander 

Thanks for responding and trying to help.

Warm regards.

@wisam No problem. I used to know a guy who had an industrial vision system company, if I still had his contact info I would have put you in touch right away. Keep in touch and tell us how it progresses.

@zander Well,  I'll definitely share comments here when I get something new.

And if you get any new insight, keep me posted.

@wisam Hard to provide insight without knowing an awful lot more. I am afraid you are at the point of not knowing what you don't know. That is not unusual, but I can't tell you how to break free of that. I was only able to comment as I did because I had a friend that was in that business, and I am a semi-pro photographer as well as a software guy and an experienced analyst. I am not saying you need to be all those things, but without some related experience it will be a much steeper hill to climb.

When you propose some random hardware and software in the guise of a question, we are not going to get very far, so think about what you want and ask a question as the user of the possible system. At this point, we don't know if we are looking for colour changes, surface appearance, holes (requires backlight)

Here are some things to think about. what is the smallest 'thing' you want to detect. What is the resolution of the camera. Does that yield enough pixels to do an analysis. This is of course influenced by focal length, distance from sensor plane to the object, and f stop.

You will NOT want to use a rolling shutter unless it is synched to the assembly line movement. My DSLR has a shutter speed of 1/8,000 of a sec or 125 micro seconds, I assume the possibility of cameras suitable for this job might have that or better but research is needed. The data transfer speed from sensor to processor as in my camera may be much slower (I have buffers and know my limits) but you need a continuous transfer so that needs research.

You can get by with a longer shutter speed if you turn the exposure way down and use a strobe for lighting that is tuned to the shutter speed. Think of those stop action demonstrations you may have seen, they adjust the strobe until the image appears frozen. As long as your assembly line speed is consistent, or a feedback loop is used this will help. One problem with the strobe, it is by it's very nature a sampling technique, if you really want 100% continuous then the hill just got a lot steeper. Consider a 2nd processor for processing a 'hit' as the extra time could make you miss a certain amount of area.

Obviously, the time to do the analyses has to be less than the shutter speed time and must happen asynchronously. In the previous example of 1/8,000 sec the 125 microseconds is quite a long time for fast MCU like a 240gHz where 1 cycle is 4.17 nano secs Instructions take more than 1 cycle, but that info is enough to determine a lot of processing can be done in 125 micro secs since there are almost 30,000 cycles in that 125 micro secs. In other words, it's doable at 1/8,000, and if you need more, just use those numbers to scale the result. My best hunch is you could easily speed the shutter up by a factor of 3 or maybe as much as 5.

HINT: If you do end up using an Arduino or similar, learn how to use the command line and change the optimization directive from less memory to more speed (or whatever it's called these days)

Out of ideas for the moment,

All the best.

@zander First of all, thanks for your time,
On the one hand, it depends on experience, as you mentioned before,
On the other hand, I don't think it needs to be complicated.
Because if the lighting is well focused, the defect will appear somewhat clearly and will be detected by the camera.
But it remains the most difficult matter to send an alert or notification because it is not possible to hire a person to monitor around the clock and continue staring at the display screen.

Compare the difference between the two pictures with and without lighting.

@wisam Sending a notification is simple, it's the detection that is hard. I have no clue what the two pictures are.