Université de Fribourg, Schweiz
Graph-Based Methods for Handwriting Analysis and Recognition
Lehrstuhl Technische Informatik und Eingebettete Systeme
Mustererkennung in Eingebetteten Systemen
Prof. Dr.-Ing. Gernot A. Fink
In statistical pattern recognition, objects are represented with a fixed number of real-valued features, for example by means of a fixed-size image that is provided as input to a convolutional neural network. In structural pattern recognition, objects are represented as parts and their relations, using strings, trees, or, in the most general case, graphs. When using graph-based representations, both the parts (nodes) and their relations (edges) can be labelled with real-valued feature vectors, leading to a powerful and flexible representation formalism. In this talk, we will demonstrate how this formalism can be used to model the rich structure of handwriting. First, for the production process, where the kinematic theory of rapid human movements allows us to decompose complex pen movements into sequences (strings) of neuromuscular strokes. And secondly, for the reading process, where inkball models (trees) and graph-based representations can be used to model the global structure of the handwriting for tasks such as keyword spotting and signature verification. The focus of the talk will be on graph representation and graph matching, i.e. the direct comparison between two graphs. As an outlook, recent advances in geometric deep learning will be addressed that aim to make graphs amenable to machine learning using graph neural networks.