Building a Pokémon card grading system as a personal project

General
1

Hi everyone,

I’m currently building a Pokémon card grading system as a personal project and would love to get feedback from people with grading experience.

My setup uses an Epson V600 scanner and computer vision techniques to analyze:

  • Centering (using SIFT feature matching against reference images)
  • Corners (shape analysis, circularity, whitening detection)
  • Edges (whitening, edge geometry, chip detection)
  • Surface (scratches, print defects, texture anomalies)

One thing I’ve learned is that comparing scans against online “master” images is not reliable enough due to color differences, compression artifacts, and source inconsistencies. Because of that, most of my grading logic is becoming reference-free, except for centering.

My goal is not to perfectly replicate PSA (I know that’s unrealistic without access to PSA’s internal standards and huge datasets), but to get as close as possible using consumer hardware and software.

A few questions for those with grading experience:

  1. Which defects do you believe PSA weighs most heavily when differentiating between a PSA 9 and PSA 10?
  2. For corners and edges, do you think geometric shape analysis is more important than color-based whitening detection?
  3. Has anyone here built or experimented with automated grading systems before?
  4. If you had to prioritize one category (corners, edges, surface, centering) for predicting PSA grades, which would it be?

Any advice, experiences, or technical insights would be greatly appreciated.

Thanks!

Efour Digest - Daily summary of activity
2

Quick question, whats your goal with this?

5 Likes
3

The goal isn’t to replace PSA or claim I can grade cards as accurately as a professional grading company.

I’m a software engineering student and Pokémon collector, and I’m building this as a computer vision project. My aim is to see how close I can get to estimating card condition using consumer hardware (an Epson V600 scanner) and image analysis.

At the moment I’m analyzing centering, corners, edges, whitening, surface defects, scratches, and print quality. I know a perfect PSA prediction is unrealistic without PSA’s internal standards and massive dataset, but I’m interested in learning which factors experienced collectors and graders consider most important.

So the project is part learning experience, part research experiment, and hopefully a useful pre-grading tool that can help identify cards that might be worth submitting.

1 Like
4

But why tie yourself to psa at all? Just make it independently based on condition? Im having a hard time understanding the goal behind your questions besides it being for starting a pregrade service. It feels like what you are grading is secondary to the grading portion.

3 Likes
5

What sort of volume are you wanting to achieve? My V600 takes about 6-10 minutes to scan front and back. If you use a flatbed guide/frame to align things with precision you can use tools like ImageMagick and ffmpeg to do quite a bit of heavy lifting without spending money on AI to process.

4 Likes
6

Imagine orienting this towards an identification system…

1 Like
7

I don’t have a scanner, but aren’t they awful at capturing subtle surface defects? If so, doesn’t this basically invalidate the whole system?

Also, how are you going to validate the model? You’d have to run your scans through your pregrading rubric, then send all the scanned cards in for grading to get an idea of predicted vs actual grades. To pinpoint the accuracy of the rubric, you’d have to do this over a large number of cards. And if you aren’t using it for a large number of cards, why not just use your eyes?

8

You aren’t going to get a consistent standard of grading in comparison to psa to determine whether something is worthy of grading, because psa themselves aren’t consistent. There’s really no point on pregrading/predicting if all it is is an unreliable estimate due to the grading companies themselves being unreliable, not through any fault of your or your program’s own. It isn’t just one person doing all the grading. You would have to know who was grading it and how they felt at that particular moment to what that grader prioritizes and criticizes the most on a specific card, and even then it’s still unreliable. Unless they switch to fully automatic processes within their business model, no system is going to be better than your own eyes. That’s not to discourage you from doing so, but more of a heads up on what you are actually facing when considering the hurdles of developing a prediction algo.

2 Likes