Convolutional Neural Network for Off-Line Writer Identification based on Line Segments of Handwritten Documents

Puspoki, Eniko (2022) Convolutional Neural Network for Off-Line Writer Identification based on Line Segments of Handwritten Documents. Bachelor's Thesis, Artificial Intelligence.

Preview	Text bAI_2022_PuspokiE.pdf Download (4MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (121kB)

Abstract

Identifying individuals based on their handwriting can be challenging, but it is an essential step in document verification tasks. This is because everyone has their unique style of writing, which means that there are no two people who have exactly the same handwriting. However, when analyzing handwritten documents, the differences between writers can be hard to notice, and the higher the number of writers, the more difficult this task is. For this reason, utilizing computer technology in this field can be beneficial. In this thesis work, a low-cost convolutional neural network is presented, which is trained on two different datasets containing images of handwritten documents to identify writers with the classification of previously unseen data. The two datasets used are the IAM English Handwriting Database and the Firemaker image collection. The input to the model consists of patches of 113 x 113 pixels, cropped from the text’s line segments. The model achieves an accuracy of 88.9% on the IAM images, and the highest accuracy obtained on the Firemaker images is 58%. In the Firemaker image collection, four different writing conditions were implemented when gathering the data, and the results suggest that within-writer variability might play a role in writer identification.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Dhali, M.A.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	02 Aug 2022 06:53
Last Modified:	02 Aug 2022 06:53
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/28225

Actions (login required)

View Item