WAY‑2C's main strength comes from its speed and ease of training, and its ability
to reliably identify items with complex shape and/or
complex color distributions. However, once in a while we're presented with
potential applications where there's no need for training and/or inspection times
measured in milliseconds, and where
color distributions are relatively simple.
Here's one such example.
Not long ago we were contacted by an engineer looking for a system to inspect
fiber optic ribbon cable during manufacturing. They wanted to insure that
the correct number of strands were being included and that all were in the right
As initially presented, the need was to continuously monitor the cable
during production at a rate of about 500m per minute. Our experiments
showed that at this rate one of our traditional camera-based inspection systems could be
expected to complete approximately one inspection per meter.
After further study the engineer reported that there would be no need for such
frequent inspection as there was no physical way for the strands
to change order once a production run had started.
It would only be necessary to inspect once at the beginning of each run.
This engineer's need may have have disappeared. However, we realized that
whether or not the application was real, it provided an excellent example of the
potential for WAY‑2C to do cloud-based inspection.
Not having any fiber optic ribbon in house, we quickly tested the concept by scanning
an available piece of copper ribbon cable with a normal office scanner.
The first image on the right, cri_01r2.jpg, is a reduced scale version
of that scanned image which we used to train the system to recognize the
classes of interest. These were ten strand color classes and a "Margin" class.
Actual training was based on scripts automatically created by our standard
WAY‑2C Assembly Inspection Package together with a knowledge of the strand
count and the specified order of the strand colors.
The system was originally conceived as one that could be used either
for continuous camera-based on-line inspection, or as a one-at-a-time
scanner based system. It was recognized that because of the small diameter
of optical fiber strands, it might be difficult to consistently
maintain ribbon position in the image within a fraction of a strand width.
Therefore the inspection needed to be based only
on a comparison of the color class determined at each y-location with
the expected strand color class order and width
but not the exact y position. Also allowance
had to be made to handle occasional misclassifications such as might be expected
in the binding region between strands.
Once the system was trained, we used standard WAY‑2C assembly inspection
tools to verify its ability to reliably differentiate the classes of interest.
As a final test we inspected
the training image itself. One product of the inspection was image
cri_01r2_PASS.jpg. It shows the middle of the original image
overwritten by color strips, each a single pixel high. The color of each strip
is coded to represent the color class most likely to be the source
of the underlying ribbon color distribution. It is clusters of such strips
that the system uses to determine whether the strand order is correct.
The green vertical bar along the left side and "_PASS" in the image name both indicate
that no problems have been found during the inspection.
A second result of this inspection is the report file cri_01r2.txt confirming that no problems
were detected in the ribbon. The same results are also appended to a log file containing
the results of all inspections.
Training Image. cri_01r2.jpg
Returned processed training image. cri_01r2_PASS.jpg
Report for "Pass" ribbon. cri_01r2.txt
Since we did not have an actual incorrectly formed ribbon to test,
we tried to simulate one. We used a simple image editing program
to interchange the
red and orange strands in the second color cycle from the top in
the training image. To do this we cut out the orange/red strand
portion of the image, flipped it vertically, and pasted it back
into the region it was taken from -more about this below.
On the right is a reduced scale version of the resulting test image
This time, instead of inspecting the image using the computer
to which the scanner was attached, we scanned it into a Dropbox™
folder shared with a remote server running the ribbon inspection software.
The server sensed when
a new image was deposited in its shared folder,
automatically processed it and then returned
the results to the client in the same shared folder.
How long did all this take? From the time the scanned image to be inspected
was deposited in the client computer's shared folder
until the results appeared back in that
folder averaged about 20 seconds. Actual time to process the image at the
server and generate the response was about a second.
To the right are the returned processed image (cri_01r3_FAIL.jpg)
and the accompanying text report.
The red vertical bar along the left side and "_FAIL" in the image name
are hard-to-miss warnings
that a failure was detected during the inspection.
The short green vertical bar indicates the extent of the "good" ribbon
above where the failure was detected. The bottom edge of the bar
shows the exact position of the problem.
The reported "FAIL" designation and description of the failure:
"Invalid color sequence" in cri_01r3.txt were both correct.
The location, cycle 2, item 5, was also correct.
However, what initially appeared surprising to us were
the details of the reported invalid sequence: "Item 5 should be Red but is Brown"
did not agree with the error we had intended to introduce, a simple interchange
of the red and orange strands. Further careful examination of the test image
showed that our editing had been less than perfect. We had actually left what
appeared to be a sliver of orange just above the top boundary of the flipped region,
and a sliver of brown appearing material just below that boundary. So the inspection
system had not only accurately identified where the problem occurred,
but also that the failure we had introduced was not exactly the one we intended.
In addition to verifying correct strand order the system will detect
too many and too few strands, and can provide a ribbon width measurement.
Image to be inspected. cri_01r3.jpg
Returned processed image. cri_01r3_FAIL.jpg
Report for "Fail" ribbon. cri_01r3.txt
Inspecting Real Fiber
Having verified that the system would work, we tested it on a sample of
12‑fiber ribbon. For this we used a commercial photo scanner at a resolution
of 3200 pixels per inch. At this resolution each strand was about
35 pixels wide. The software is set up to be able to select either
conventional color specs for 10 colors per cycle with the standard
color order for copper ribbon, or the standard color specs for
12 colors per cycle fiber optic ribbon. There is no practical limit
on the number of fibers.
A reduced scale version of the training image for the colors is shown on the right.
Here we used
a light bluish green paper background to be easily differentiable from the fiber
coating colors. Note that the latter are much more muted
(lower saturation) than the vivid copper ribbon colors.
To accomodate this we increased the WAY‑2C sensitivity setting a notch.
Standard WAY‑2C tools confirm negligible statistical uncertainty in
differentiating the color classes.
As might be expected, the system found no ribbon faults when inspecting
the training image -so we haven't shown its returned processed image or
Again, without a sample of an incorrect ribbon, we had to resort to image
editing to create a test ribbon that would fail inspection. In this case
we swapped the "White" and "Slate" fibers. The result is shown as the image
to be inspected on the right. Once more, because of the slight tilt of the
fibers in the image, our editing was less than perfect. While the software
handles this tilt with no problem, the editing introduced the narrow white
band detected just below the slate. As shown in the returned processed image,
this was detected by the classification step in the inspection software.
The order checking portion of the software contains a feature to ignore
such very narrow strips.
The returned processed image and corresponding inspection report both show
that the system has correctly located and properly identified the nature
of the problem.
We're looking forward to demonstrating the system on a ribbon with many more fibers.
Fiber Training Image. fi1212_1c.jpg
Image to be inspected. "FI1212_1r.jpg
Returned processed image. FI1212_1r_FAIL.jpg
Report for ribbon. FI1212_1r.txt
What We've Learned
So what have we learned? We've learned that for applications of this type,
where only intermittent inspections using off-line images are needed,
cloud-hosted inspection may be a reasonable option.
In this ribbon inspection application, after initial setup, a user can simply:
Clip a sample at the beginning of a run,
Scan the sample with a high resolution photo scanner directly
into a Dropbox™ or similar folder shared via internet with
a remote inspection server.
Within seconds receive, in the same shared folder, the results of the inspection.
These results include: a short text report describing the problem (if any),
a new image with "PASS" or "FAIL" in its name showing where the failure (if any) was detected
and an updated log of all inspection results to date.
We'd be happy to work with your
vision system integrator on any kind of color-based on-site inspection system
with its attendant specialized lighting, cameras and software.
We'd now be equally happy to work with them on a cloud-hosted system.
If there really is a need for this type of fiber optic ribbon cable inspection,
and you are interested in
a no-charge, no-obligation test, contact us. We'll give you
instructions on sample image requirements and system setup procedures.
For more information: