Pynq OCR frustration

RE: Pynq OCR frustration

dubbie — Fri, 03 Jul 2020 09:16:54 GMT

Would it be possible to try something radically different? I can remember being a student looking at case examples at how to use vision systems to recognise some piston rod things from engines that had been manufactured and dumped into a box that needed to be picked out of the box by a robot arm. There was lots of fancy image processing and lighting but eventually the company just stopped dumping them all into a box - which caused all the different orientations - and maintained alignment and stacking straight out of the manufacturing cell. Problem solved instantly, no expensive and unreliable vision system needed.

So, would you be able to add a bar code below the number, that should be much easier to detect. I one made a system detecting bar codes using a early solid state video camera and a PIC18F84 which processed live video data as it came straight out of the camera. I was surprised it worked too, but it did. We did overclock it a bit but not one of the PICs ever objected.

Dubbie

RE: Pynq OCR frustration

Fred27 — Tue, 30 Jun 2020 18:40:34 GMT

By the way, can anyone guess from those numbers what sport I usually watch? I'm also using 04 and 99 if that helps.

RE: Pynq OCR frustration

fmilburn — Sat, 27 Jun 2020 03:03:52 GMT

I wonder if using color with the numbers an unusual color in a training dataset would help speed things up and improve accuracy. Rather like differentiating between apples and oranges. Some of the training exercises In the online class I took used color that way. On the other hand, like you said, this would seem easy without the bother.

RE: Pynq OCR frustration

ralphjy — Sat, 27 Jun 2020 02:16:06 GMT

I feel your pain - I went down a different rabbit hole with Movidius and Pynq I’m probably going to abandon my project due to lack of time.

Good luck with yours - it’s frustrating when you can see the light but you’re not getting closer [View:/resized-image/__size/16x16/__key/commentfiles/f7d226abd59f475c9d224a79e3f0ec07-fd423752-d385-4e4d-ba32-6e0876bc6590/contentimage_5F00_2.png:16:16].

Think of all the image processing that you’re learning [View:/resized-image/__size/16x16/__key/commentfiles/f7d226abd59f475c9d224a79e3f0ec07-fd423752-d385-4e4d-ba32-6e0876bc6590/contentimage_5F00_1.png:16:16]. I’ve learned a lot more about Linux.

RE: Pynq OCR frustration

Fred27 — Fri, 26 Jun 2020 22:28:52 GMT

One thing I find very strange is that what looks better doesn't always work better. Sobel edge detection for instance never looks like it would help at all but sometimes it seems it does.

RE: Pynq OCR frustration

genebren — Fri, 26 Jun 2020 22:10:30 GMT

You really look like you are getting close. OCR is tough, due to size and angle. I would be tempted to use a 'sharpen' filter across the image to drive the numbers to black and white (not as munch gray).

[View:/resized-image/__size/620x360/__key/commentfiles/f7d226abd59f475c9d224a79e3f0ec07-fd423752-d385-4e4d-ba32-6e0876bc6590/5732.contentimage_5F00_191801.png:620:360]

I just ImageJ to analysis the image. I constructed a rectange around the '46' in the image and then produced a histogram of the region. The upper/left window is the result. The image appears to have fifteen small peaks with the majority of the region being black. After I ran 'sharpen' across the the image, I ran another histogram on the same region. The lower/left window is that result. The majority of the region is now either black or white, with little gray in between. This might produce a better image for the OCR to be run on.

Good luck!

RE: Pynq OCR frustration

Jan Cumps — Fri, 26 Jun 2020 21:35:06 GMT

[mention:2136c4649d3b47f78653b7e97ac61f69:e9ed411860ed4f2ba0265705b8793d05], any advise? You have experience with visual pattern recognition ...

RE: Pynq OCR frustration

beacon_dave — Fri, 26 Jun 2020 20:09:47 GMT

How about some of the automatic number plate recognition systems for this ?

Path II Programmable Project Report ( License Plate Recognition ): Finished