H1merka/efficientnetb0-finetuned-banknote_classifier

Banknote Classifier

Overview

This repository contains a fine-tuned EfficientNetB0 image classifier for banknote recognition. The model is designed to classify cropped banknote images into one of nine denomination classes.

Model Card Summary

• Base architecture: EfficientNetB0
• Task: image classification
• Input size: 224 x 224 x 3
• Training framework: TensorFlow / Keras
• Training strategy: transfer learning with a frozen backbone, then fine-tuning of the upper EfficientNet blocks

Labels

The classifier predicts the following denomination labels:

10, 100, 1000, 200, 2000, 5, 50, 500, 5000

Training Data

The training notebook expects a directory structure similar to:

dataset_rub/
  train/
  val/
  test/

During training, class folders are read from train and val. The notebooks also use a separate test split for visual checks.

Training Procedure

The model was trained in two stages:

1. Feature extraction with EfficientNetB0 initialized from ImageNet weights.
2. Fine-tuning of the deeper EfficientNet blocks with a lower learning rate.

Preprocessing and augmentation used in the notebook:

• preprocess_input from tensorflow.keras.applications.efficientnet
• random rotation, shift, shear, zoom, flips, brightness and channel jitter
• class oversampling during generator creation to reduce imbalance

Important training settings from the notebook:

• loss: categorical crossentropy
• optimizer: Adam
• initial learning rate: 0.001
• fine-tuning learning rate: 1e-5
• early stopping on validation accuracy
• model checkpointing to save the best .keras model