Automated Identification of Scribes via Neural Networks

1
Automated Identification of Scribes via Neural
Networks

• Vitaly Feldman Matthias Roeder

2
Outline

• Handwriting Identification
• How can computers help?
• Approaches to the solution
• Learning-Based Systems
• Neural Networks
• Experiments
• Results

3
Factors which play a Role in Handwriting
Identification

• Single musical symbols
• Layout of the page
• Placing of symbols
• Size of symbols
• General appearance of the script

4
Relevant symbols for Handwriting Identification

• Clefs
• Time signatures
• Rests
• Key signatures and accidentals
• Noteheads, stems, beams, and flags

• Dynamic marks
• Textual score markings
• Articulation marks
• Braces, repeat signs, double bars

5
Handwriting Identification Overview

• Define standard by
• Analysis of music symbols
• Definition of their characteristics
• Analyze questioned-document by
• Analysis of music symbols
• Definition of their characteristics
• Compare standard with questioned-document

6
Principles of Handwriting Identificationas
outlined by Dexter Edge

• There has to be a distinction between system and
  personal characteristics. System characteristics
  are those belonging to a general system of
  writing used by many people at a particular
  time.
• There has to be a sufficient body of
  authenticated handwriting to serve as a standard
• The degree of variation must be taken into
  account
• There can be no unexplained differences

7
Computers and Handwriting Identification

• Handwriting identification is a complex task
• Computers are far from reaching the precision of
  humans

Why are we still interested in using computers?

• Computers can be used to examine digitized
  collections of manuscripts and sort out similar
  ones
• Computers can greatly reduce the number of
  sources which have to be examined manually

8
Approaches to solve a Pattern Recognition Problem

• Expert system
• Learning algorithm

9
Problem Abstraction

• Input I images of symbols from authenticated
  manuscripts
• Input II images of symbols from unidentified
  scribe
• Output list of scribes whose handwriting is
  similar to that of the unknown

10
A Learning-Based System

• Training For each type of important symbol and
  each scribe create an algorithm that identifies
  the scribe of the a symbol scribe expert
• Query For an unknown symbol run all the scribe
  experts to find out which scribes it might belong
  to

11
Neurons
12
Neural Networks
13
Artificial Multilayered NN
14
Preprocessing of Images
Original Image
Rotated Image
Cleaning
Contrast
Scaling and Alignment
15
Examples

• Johann Sebastian Bach
• Anna Magdalena Bach
• Johann Friedrich Hering

16
Performance Evaluation

• Resolution of images 70x40
• Grayscale vs. black and white
• Network architecture no hidden layers
• Training algorithm back-propagation with
  momentum and flat spot elimination
• Parameters of the training algorithm

17
Performance of Experts
18
FIN