Seminar Topic: Object and character Recognition Tse Ngang Akumawah Lehrstuhl für Praktische Informatik 3
Table of content What's OCR? Areas covered in OCR Procedure Where does clustering come in Neural Networks K-means etc Application Examples Briefing on Object Recognition
Motivation Scanning Payment order Name: John Mustermann KNr: xyzabc Bank: MarxBank OCR-System Output
The different areas covered in character recognition.
The different areas covered in character recognition.
The different areas covered in character recognition. 1. Mechanical character recognition 2. Magnetic character recognition 3. Optical character 1. On-line character recognition 2. Fixed font character recognition 3. Handwritten character recognition 4. Script Recognition => Seminar concentrates on OCR
What s character recognition all about? multiple layer procedure
Preprocessing!!! Scanning / photographing N.B. minimal variation from the original document. Content extraction using horizontal and vertical lines. rotation by necessity Binary conversion 0 s = black spots 1 s = white spots
Character segmentation!!! Example based on 0 for black pixels and 1 for white s Analyse line after line For each line column of 1 s => disjoint characters the lager the number of columns, clearer the demarcation (see fig)
Character segmentation!!!
Character segmentation!!! After identification of distinct character form new matrices with each character empty spaces are filled up with 1 s nominalise and resize each character matrix By necessity smoothen the characters. (see fig)
Character segmentation!!!
Feature extraction!!! Analyse each character separately. 1. Statistical Features 1. number of black pixels in a line(zoning) 2. total mass(nr of pixels) 3. etc 2. Structural Features joining points centroid end points strokes etc Other ALGORITHMS Hough Transformation Fourier Transformation
Statistical feature extraction
Structural feature extraction
Where does clustering come in? Is OCR really unsupervised? Objective of clustering: partitioning the sample set into subsets. OCR systems before the training are unsupervised Problems: 1. optimal partition 2. ideal number of partitions Character of Clustering techniques for OCR systems 1. large amount of sample data. 2. small number of prototypes
Finding the optimal partition for a given number of prototypes NEURAL NETWORKS
Finding the optimal partition for a given number of prototypes NEURAL NETWORKS curve-1 length centre
Finding the optimal partition for a given number of prototypes K-MEANS The K mean technique some kind of like iteratively process 1. for a start, one could randomly pick up such prototypes(cluster centres) 2. For each prototype clusters are build for the remaining spots in the graph 3. When the clusters have been form we then look for the ideal(mean) prototype for each cluster 4. repeat step 1 3 till the changes(variance) in prototype is below a certain range 5. In this way we get the optimal cluster
Finding the optimal partition for a given number of prototypes Others Nearest Neighbour Analysis Mean squared error. (MSE) N.B. Most commercial OCR systems don t use just one clustering technique but a combination of two or more depending on the area of usage.
What's left? So far build up clusters identify characters possible? TRAINING
Real life application of character recognition systems 1. For Data Entry. - e.g. in banks 2. For Text entry. - e.g. in newspapers 3. Process automation. -e.g. in post offices. 4. Readers for the blind
Examples of existing systems Iris Reading System(Visuaide 2000 Inc.) Arkenstone Readers(Arkenstone Inc.) Reading Advant Edge(Xerox Imaging Systems. ) etc.
Summary
What s Object recognition all about? similar to OCR Differences More complex more clusters needed demarcations of an object Input Camera Moving Other Algorithms needed colour histograms gray scale histograms Edge tracing using Graph approach Least cost Trajectory Region growing etc.
Motivation Control Traffic sign recogniton
The End