Like a lot of true collectors here I use a flatbed scanner to scan my graded cards for documenting as well as for uploading pictures to my eBay store, However my OCD and thirst for efficiency would always present me with the following issues:
- I scan four cards at a time on pokedreamer’s scanner frame and cropping all four images is tedious
- slabs are rounded and crops are square, so you end up with little corner slivers of black or white around your card and it’s just not aesthetic
- eBay uses square ratio for all it’s thumbnails, so a vertical card crop ends up having that boring default light grey border around your card
So instead I used a combination of photoshop templates and manual work to cut each graded card out with rounded corners, place it on my preferred backdrop color of Eigengrau, and center it within a square aspect ratio. The problem is this took way too long and has little to no effect on my sales.
After enough rationalizing to myself that this was “worth it” I decided to try and automate this process as both a learning experience and way to stop wasting hours in photoshop.
First I made three folders of images holding specific datasets:
- raw images of my scans, using a variety of card orientations and styles to get a smarter model
- “ground truth” alpha maps of those scans where the cards should be cut out
- another set of raw scan images not in the first set to validate the machine learning model
Which basically gives the algorithm a cheat sheet to learn what it’s supposed to do - this is what they looked like:
For training the machine learning model, I used this open source library which only took me a full day of fussing with my development environment to get running but once it did it was beautiful:
After a couple of hours cooking on my video card, the model was finally ready to work! Amazingly, the first model produced basically perfect results:
But of course I need every card by itself as it’s own square ratio image on a nice dark grey, so next I wrote a python script using Pillow and OpenCV to programmatically detect the contours of a card, read the cert number for its filename, and then save that card out with my desired aspect ratio and background image. Shoutout to Chat GPT which helped me write most of this script. I repeat this process for the scan for the back of the cards and the end result is an output folder which I can easily upload to my eBay listings:
I exclusively use this script to prepare slab scans now and it’s taken my time to process a PSA return from an hour or two down to seconds (minus physical scan time). After a couple dozen more PSA submissions I might have saved enough time to make up for what I spent writing this script and thread. Cheers!
. I also do something incredibly similar, with a pricing database that uses ML to predict price points for grades which don’t exist. It’s fun, but it’s also my day job so I try not to spend too much time doing this stuff outside of work 
.
My post was a joke by the way 