- Project Takeaway
- Description of the Data
- Methods Overview
- Project Summary
- Descriptions of Files in Repo
For my masters capstone I built a neural network to automate the manual card grading process done by companies such as PSA. Card holders now no longer need to ship their valuable cards off and pay top prices for a manual professional grade. This model allows users to get instanteneous card grades from just a snap of their phone camera.
- The model performed well, assigning new cards a grade score with 76% accuracy, with a variance of +/- 1.2 grades.
- The model was trained on 10k labeled images with professional scores from PSA, which underwent rigorous cleaning: image selection, image cropping, and image augmentation.
- Neural network design was essential. Built on top of RESNET-50, many decisions were made: number of dense layers, pooling method, activation functions, freezing layers while training, batch size, learning rate, dropout rate.
Using this model, I founded Card Grade AI, LLC Our app allows users to get instant card grades, and host a virtual portfolio of thier collection. Tt is currently in progress. For this reason the training data and final ML model are not on this repo, and are considered proprietary.
All 10k cards were scraped from Collectors.com using beautiful soup
- Exploratory Data Analysis
- Data Collection
- Data Cleaning
- Data Standardization and Augmentation
- Neural Network Model Building
- Model Evaluation
- ResNet-50 feeds into three fully connected ReLU dense layers [1024, 512, 128]
- The dense layers feed into a single softmax prediction layer
Two rounds of training were done:
- First Round: 15 epochs, freeze first 20 layers of ResNet-50
- Second Round: 15 epochs, unfreeze all layers
The condition of a trading card significantly influences its worth. However, the current methods to assess a card's condition are insufficient. Non-professionals often lack the accuracy required for card grading, and obtaining professional evaluations is too expensive and time-consuming.
This project aims to automate the card grading process using machine learning. Our neural network utilizes the pretained model ResNet-50 combined with custom classification and output layers.