Signature Recognition

Signature Recognition

The post is written about an AI model which can predict the forgery and genuineness of a Signature sample.

Problem Statement

In organizations like banks where the Authenticity of the signature of a person is very important, there still doesn’t exist any autonomous system which can predict forgery of a signature.

Artificial Intelligence is the field of computer science that emphasizes on creating intelligent systems and through which a machine can learn to make decisions, I have been working in this field for a while, so I decided to solve this problem with the help of some knowledge that I have been able to gain in past few years.

The Idea

In organizations like banks, they basically have the database of every customer hence I made my model accordingly. So let’s dive into the technicalities now.

Dataset —

I used various Online datasets available for the model. The idea was to have a number of samples of every person/customer including genuine as well as forged signatures to make a person dependent system. Luckily I found various online datasets of the same manner online. I used Dutch as well as English signatures so as to increase the dataset

Preprocessing –

The Right image is the genuine signature and the left one is the forged one.

the first problem was to remove the effects of different inks from the image as it can be of any form. hence for that, I applied thinning to the image by using morphological functions of sklearn.

here is the image after thinning it.

The image was also converted to grayscale from RGB and was also normalized and smoothed to negate factors which could affect the model, every image was also resized to same size i.e [50,100].

left is the grayscaled image and right is the smoothed image

Image Preprocessing is one of the important tasks in deep learning especially in a task like signature Recognition where small variations can truly affect the model.

Features Extraction

As this task is not just recognizing dogs or cats or any other classification task, I had to extract some hand-picked features from the image, here I took help from some research papers I read beforehand.

Density of the signature

The Features tells us the ratio of the number of pixels of the signature to the total number of image pixels, for this the image was binarily segmented by open cv after converting it to grayscale.

here is the binary image through which the ratio was calculated

Compute the number of spatial symbols within the signature Image.

Every person in their signature uses some spatial symbols, such as they use some x marks (cross marks), star marks or other symbols. The total number of spatial symbols of a person’s signature is unique.

For calculating the total number of spatial symbols in a signature image we have to preprocess the image up to thinning. Then If we find that one pixel having more than two neighbors each of which get the values 1 then those pixels will form a Spatial symbol

Height/width ratio of the signature

I read in a research paper that a person signature’s height is to width ratio always remains the same, hence I computed this feature again open cv helped a lot, I had to make Contours to basically make a bounding rectangle around the signature.

Other Features

Feature such as skewness , kurtosis and standard deviation were also used which have great statistical relevance in image processing. I won’t go into much details for them, but here is some information regarding this.

Architecture

The features which I am using are kinda specific to a person that is, they could identify the forgery of a specific person on compared to his/her genuine signature but they couldn’t do this job in general hence I ended up with two models.

First Model

This Model is a Convolution neural network , it’s task is to classify the sample signature into its true class, here all the people (customers in case of banks) are treated as different classes, the CNN’s task is to identify the class of the signature regardless of it being genuine or forged . For example, a bank has 2000 customers then there will be 2000 classes. In simple words CNNs task is to identify the person to which the signature belongs regardless of it being genuine or forged.

Second Model

After the class of the Signature is classified I trained another network which uses the handpicked features to predict the desired output by using KNN classifier , this classifier was trained only on the data of the particular person(or the class which was classified by the first model). For this, I made a number of data frames that is if there are 100 customers then 100 data frames each containing genuine as well forged as samples of that particular person.

Conclusion

follow this link for the GitHub repo of the project
The network gave about an accuracy of 85 % on the test set, which is quite considerable I guess. This whole work was done for an online competition though I didn’t get through but it gave a lot of experience and I hereby apologize for all the spelling mistakes and grammatical errors I did throughout this blog, thanks for bearing.