Object Detection

@zander This is a simple sample of the plastic raw material that contains a manufacturing defect.

@wisam This is a good example of how an end user communicates. It is NOT what the analyst needs. WHAT is the defect, colour, thickness, surface texture, or something else? There is NO sensor/software that detects 'defect', they can detect holes, thickness, colour, and reflectivity.

I hope you are starting to understand there is a HUGE gap between where you are and where you need to get to. In code we cant say, 

if defect then reject

unless reject is maybe a function that uses a hardware sensor to determine the thickness etc.

If we look inside the defect black box, we will see something like (pseudo code)

if ((thickness > 2mm) OR (thickness < 1 mm)) then signal defect.

We do that because the user told us the thickness must be at least 1mm but no more than 2mm.

This is just a hypothetical example, the values can be whatever is pertinent to your use case.

Building a business/industrial system is all about communication and translations. The analyst translates user speak to system speak, the designer translates system speak into program speak and the programmer translates program speak into code.

The analyst would for instance when he heard the word 'defect' ask for a definition and hopefully you would tell him the specific technical details that comprise a defect.

Is this helping?

@wisam

After all those posts I still have no clear idea what the set up is or kind of defect? Is it a plastic covered spinning cylinder or a sheet of plastic being rolled out by a cylinder. I don't know what the images are really. Two still shots of the plastic, one with the kind of defect to be detected visually?

You subtract (compare) the pixel values of the image without the defect and the pixel values of any images that might have a defect.

You might direct your question to the edge impulse mob?

@wisam

Just type into google,
detecting defects in plastic sheeting
and you get plenty of hits

https://parisatechnology.com/vision-inspection-systems-for-plastic-sheet/

https://darkfield.com/industries/plastic-film-sheet/

and so on...

@robotbuilder I had to look up edge computing again, the first time all I saw was it applies to routers etc, but a second look tells me it's whatever the end user is in contact with so this applies.

As far as pixel comparing, keep in mind the object is 1 metre wide moving at 1 meter per second. There is no mention of the minimum size of a so called 'defect'. If it is as large as 1 sq cm, then there are 10,000 parts of a photo/video to be analyzed per second, but if it's more like 1 sq mm then it's a million. When is this pixel analysis going to be done if it's a continuous process?

It doesn't matter what we say, the OP has already picked the camera and software. If you go back to his first post, he just wants to know how to send a notification after all the magic has happened. The notification takes 1 line of code plus a handful of lines for setup. 

@zander 

The edge site has a forum so I would think that would be the place for answers on getting a notification of a defect.

Any defect on a sheet I imagined would just be a comparison with what it should look like just like looking for a defect in a circuit board.

For more complex requirements such as sorting fruit then colors, shape, size, markings and so on all come into play. As I have mentioned many times visual processing and object recognition has interested me for a long time 🙂

@robotbuilder Agreed, I was reacting to a continuous 1 to 2 metres per second by 1 metre wide scenario. I suspect the edge resource the OP mentioned includes a multi-thousand-dollar camera as well as a lot of software. I will bet as much as you want defect notifications are part of the software already.

Hi @wisam,

  I am unclear about the mechanical (and legal) arrangements concerning this enterprise, and tend to agree with Ron's concerns. This is only a suggestion to consider ... you are completely responsible for implementing it. You may well need professional help implementing it, which is well beyond my experience.

----

  So, just a couple of related thoughts which maybe totally inappropriate, but maybe useful.

I am assuming the film is partially transparent to light. This implies light passing through unflawed film will be partially absorbed by a constant amount, whilst flaws may absorb more light if it is thicker, less light if it is thinner.

1.   Rather than looking at an image on the roll, consider an arrangement that passes light through the film as it goes onto, or comes off the roll. I am assuming that the flaws will alter the intensity of light passing through the film, which can be detected by a row of light sensors.
2. Flatbed image scanners use a single 'component' consisting of a row of sensors to 'photograph' a single line of the page image at a time. I suspect these components may be available, so perhaps they could be adapted for this purpose? In addition to designing appropriate mechanical, and possibly optical components, this project will include relatively high speed electronics to transfer and analyse the light level measurements. Whether sufficiently high speed sensors and electronics are readily available to meet the speed that the film is passing by the sensor, I have no idea.

Best wishes, Dave

@zander @DaveE @robotbuilder

I tried to do a simple experiment using the Edge impulse, but it definitely needs further improvement and development

@wisam That's for sure.

Hi @wisam,

  Your experiment suggests that the software principle to 'analyse' the data could be very simple, as all it needs to do is to look for changes in light level, which might amount to little more than a numerical comparison and if statement, albeit practical issues like noise reduction may well add complexity. However, it may require a substantial 'low-level' software content to perform the necessary data handling involved with interfacing to sensors, motors, etc. Furthermore, depending upon the amount of film to per scanned per second, and the minimum flaw size to be detected, the speed of processing may require some specialised approaches that could vastly increase the complexity of the solution.

It is possible that components developed for edge processing could play a part, but from my simplistic viewpoint, whilst they may provide a convenient solution to one or two aspects of the overall solution, I do not see this as a problem that is generally amenable to an edge processing approach. Of course, this is only a wild speculation based on the very limited amount of information presented ... a detailed analysis and specification of the full requirements may result in a completely different conclusion.

The'obvious' tricky bit will be incorporating the simple principle of flaw detection into the overall process, including the mechanical film handling, with light sensors, electronics to clock the data out, perform comparisons, etc. possibly at very high speed, depending upon the rate the film is moving. Scanning an A4/letter size sheet in 30 seconds would probably be straight forward, though far from trivial, possibly based on modifying a desktop scanner. Adapting that to a situation where the equivalent of an A4 size page of film passes by the sensor every (say) one tenth of a second, is likely to present far more practical issues. I can only widely guess at your requirements, and I may well be orders of magnitude away from reality.

Thus, whilst I would not wish to dissuade you from continuing to investigate the possibilities, as I have seen nothing to suggest it is impossible, but I would like to warn you that this could become a complex and expensive project, including a wide range of skills, including mechanical, electronic, software and optical engineering.

Best wishes and good luck, Dave

@davee @ron

Thanks for your input.

I know that it is not easy to implement, but I will continue searching to find the best possible solution at the lowest costs.

@davee @wisam So far he has shown us a small spot in the middle of a 1 meter wide sheet flying past at 1 to 2 meters per second. Draw a picture of a camera over that sheet to scale and observe the angles involved. The ratio of mm^2 object to number of pixels will change by quite a lot and the light intensity will fall off rapidly.

There is MUCH more to this.

@davee @wisam If you read the comments, you would know the object is 1 meter wide moving at 1 to 2 meters per second so the following is indeed off by orders of magnitude.

Scanning an A4/letter size sheet in 30 seconds would probably be straight forward, though far from trivial, possibly based on modifying a desktop scanner. Adapting that to a situation where the equivalent of an A4 size page of film passes by the sensor every (say) one tenth of a second, is likely to present far more practical issues. I can only widely guess at your requirements, and I may well be orders of magnitude away from reality.

The other thing everyone is failing to take into account is if a conventional camera lens is used then the camera will see a cone and that means the parts farthest from the center of the lens will have much less light, and the area to be examined per degree will be much higher.

With 1 meter in width, I would estimate at least 10 cameras spread equally across the width and separated in the Y axis by the same amount.

Strobe lighting will be mandatory.

BTW, I have not yet seen a definition of what a defect is, looks like.

Hi Ron @zander,

  Sorry, I did miss the exact speed and width, but without knowing the size of defect, etc. that must be detected, these are not that helpful.

So instead, I used the A4 scanner as a widely available example of the principle I was suggesting, applying a rough idea of the speed achieved by common, low cost examples found in many homes and small offices.

That is, not using conventional cameras, but rather a line scanner approach. The optics are also rather different, with each of the active parts of the sensors probably within about a 1cm of the 'spot' they are scanning.

Whether the suggestion is feasible or not, I don't know ... that is why I suggested it only as a possibility for further research, while warning it was a complex project.

I hope @wisam found it thought provoking and interesting.

Best wishes, Dave