The idea is to explore a classification problem for a single coin and a regression problem for a group of coins, trying to count how much money they sum. Idea was taken from Brazilian Coins
- Background masking
- Coin segmentation
- Image augmentation
- Train model
- Test on classification and regression datasets
Original image:
Use HSV format to select proper threshold for background masking.
Use Hough transform to find circle objects on image.
SVM on radiuses of coins shows ≈60% accuracy
- Custom CNN on edged images: ≈30% accuracy
- Custom CNN on color images: ≈85% accuracy
- Bottleneck features from VGG16: ≈97% accuracy
- Feature tuning of VGG16: ≈98% accuracy
Mean error for regression: 6.5 cents