SigNet is a Siamese Convolutional Neural Network modeled to verify original and forged signatures offline. It takes just one genuine signature of a person and then all other signatures, whether genuine or fraudulent, can be verified by it. This paper can be referred for better understanding.
- The model is trained on CEDAR dataset which can be downloaded from here.
- Extract the files and then you will get the following file structure :
|-- signatures
| |-- full_org
| | |-- original_1_1.jpg
| | |-- original_1_2.jpg
| | |-- ...(all original 24 signs of 24 signers i.e. 24x24 = 576 images)
| |-- full_forg
| | |-- forgeries_1_1.jpg
| | |-- forgeries_1_2.jpg
| | |-- ...(all forged 24 signs of 24 signers i.e. 24x24 = 576 images)
- The generate_dataset.py creates a pandas dataframe and the train and validation datasets are splitted here.
The model is trained on colab using both keras and tensorflow which can be found here signet_keras.ipynb and signet_tf.ipynb respectively. The train accuracy of the model so far is 81.42%. Better results can be achieved by augmenting the dataset with more examples.
Contrastive loss was used for the training purpose alongside RMSprop optimizer.
def contrastive_loss(y_true, y_pred):
margin = 1
sqaure_pred = K.square(y_pred)
margin_square = K.square(K.maximum(margin - y_pred, 0))
return K.mean(y_true * sqaure_pred + (1 - y_true)* margin_square)
Note: For training, the label for similar signatures is '1' and for dissimilar images it is '0'.
| Original | Genuine | Forged |
 |
 |
 |
| Distance (compared to original) as output by model : | 0.12116826 | 1.43014560 |
| Predicted Label: | 1 (similar) | 0 (forged) |