Using computer vision and deep neural networks to analyse recursive data structures

Abstract

Recursive data structures are fundamental to the solution of many problems in computer science. In particular, recursive structures based on graph theory have been successfully applied to a diverse range of problems including search, storage and machine learning. Despite their utility and widespread use in prototyping, design and teaching, little research has been conducted into how hand-drawn representations of graph structures can be automatically detected, parsed and analysed by computers.

This thesis presents research which investigates the feasibility of parsing a hand-drawn undirected labelled graph and translating it into a JSON representation that maintains its isomorphic properties. The JSON representation will include both the text from handwritten labels extracted from nodes and the relationships between nodes present on the graph. Following research of the literature surrounding artificial neural networks, deep learning and computer vision, a software prototype was designed and developed to investigate the feasibility of automated processing of hand-drawn graphs. This thesis presents the design of the prototype application, benchmarks its performance and evaluates its utility as a graph-processing tool.