THE SCIENCE OF SIGNATURE DETECTION: UNDERSTANDING THE TECHNIQUES (PART-II)

        Hello Readers, Now I am continuing with the next part of this blog. In the previous blog, we discussed what a fake signature is and how it affects our daily lives. So now in this part, we will discuss the techniques that we can use for signature detection.

        There are various techniques that can be used for fake signature detection using machine learning, including:

• Image processing techniques: Image processing techniques such as edge detection, corner detection, and feature extraction can be used to extract features from the signature image, which can then be used as input to machine learning algorithms.
• Supervised learning: Supervised learning techniques such as Support Vector Machines (SVM), Random Forest, Decision Trees, and Neural Networks can be used to train models on labeled signature data sets, to differentiate between genuine and fake signatures.
• Unsupervised learning: Unsupervised learning techniques such as clustering and anomaly detection can be used to detect patterns and anomalies in signature data, which can be used to identify fake signatures.
• Feature engineering: Feature engineering is the process of selecting and extracting relevant features from the signature data, which can be used as input to machine learning algorithms. Feature engineering can improve the accuracy of the models and reduce computational complexity.
• Hybrid models: Hybrid models, which combine different machine learning techniques such as supervised and unsupervised learning, can be used to improve the accuracy of fake signature detection.
• Time series analysis: Time series analysis can be used to detect patterns in the signature data over time, which can be used to differentiate between genuine and fake signatures.
• Deep learning: Deep learning techniques such as Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) can be used to analyze the signature data and detect patterns, which can be used to identify fake signatures.
• Ensemble learning: Ensemble learning techniques such as bagging, boosting, and stacking can be used to combine the outputs of multiple machine learning models, to improve the accuracy of fake signature detection.

        So now to test made this system I have used CNN(Convolutional Neural Network), with different techniques like Siamese Network, VGG6 Network, etc. 

        Now we will go through each of the techniques. Convolutional Neural Networks are multilayer models consisting of many convolutional layers, with a variety of different kernel sizes joined together by many pooling layers that summarize and down-samples the output of the convolutions before sending them to the next layers. Here a different convolutional kernel is used with sizes from 11x11 to 3x3.

        In the proposed Siamese neural network, CNN, and VGG16 models we have implemented it using different kernel sizes in between the range 3x3 to 11 x 11. There are a variety of different loss functions that can be applied to find out correct weights and bias values. The loss functions are Triplet Loss Contrastive Loss, Binary Cross-Entropy Loss that can be applied. Siamese network is an extension of the Convolutional neural network and it generally has two simultaneously running neural networks. The two CNNs are used with the same configuration and same parameters shared weights. Triplet loss uses three images: 1. Anchor 

                       2. Positive 

                       3. Negative

        When training a Siamese network with triplet loss it has to take 3 inputs to compare at each step. Anchor image is used as a point of comparison for two other classes, a positive image is like the anchor and a negative image is different. At each iteration, it is needed to calculate the distance between positive and negative images. After each iteration, the anchor image has the closest distance to the positive image In this paper contrastive loss defined is used which is famously used in the Siamese neural network it is described as follows, and also this loss function (1) is the same function used for all the two models: loss (d, Y) = 1/2* Y*d^2 + (1-Y)* 1/2* max(0,m-d)^2

                                                        Where d- distance between the output of the encoder.

                                                        Y is the label of model inputs and m is the margin parameters. 

        

        For calculating loss it is necessary to find the Euclidean distance between two output vectors. The Euclidean distances between consecutive images will be far greater if the underlying condition is that the images do not match with each other i.e., it could be rephrased as the images are different, then by the above observation, we can successfully predict that the signature is forged or not. To determine whether the correct signature is given we need to know the threshold value of the loss.

Fig. Siamese Network Model Working

   The given Fig. Siamese Network Model is made with reference from https://www.researchgate.net/figure/Architecture-of-Siamese-network-for-patch-based-feature-extraction-a nd-training and it shows the essence of the working principle of the Siamese Network. As for the VGG16 model it doesn't take 2 inputs at a time. It is required to train 2 models, one of VGG16 and another one user-defined. After training the first model we can transfer the learning of the first model to the second one. The base model is taken and few layers are appended in this model namely sequential. pooling, and dense. A similar process is followed for the second model as well, just with an addition that the inputs will also contain learnings of the previous model.

        Here CNN (Convolutional Neural Network) a deep learning algorithm is used for identifying whether the signature is real or forged. The pattern of convolutional neural networks is much similar to the pattern of neurons in the human brain. It gives different weights and bias values to different parts

the image and helps in identifying one image from another.

        This is all about the techniques, which I have study and hence I transferred to you for your better understanding of new knowledge. I will create one other last blog on this topic, where I will explain all the factors, parameters which are useful for this process.

        Thank You!!!

Triveni Suresh Gaidhane

B.Tech(Electronics and Telecommunication Engineering)

MIT Academy of Engineering, Pune.

Get in Touch