Related Experiment Videos
Emergent componential coding of a handwritten image database by neural self-organisation
1Defence Evaluation and Research Agency, Malvern, UK.
Network (Bristol, England)
|April 30, 1999
Summary
This study shows unsupervised component discovery in large handwriting images. The method identifies pen-stroke segments, advancing image analysis and pattern recognition.
Area of Science:
- Computer Vision
- Machine Learning
- Pattern Recognition
Background:
- Unsupervised component discovery traditionally used small image fragments.
- Previous methods struggled with large, homogeneous datasets like handwriting.
Purpose of the Study:
- To demonstrate unsupervised discovery of localized components in large, homogeneous image data.
- To identify specific structural elements within handwriting images.
Main Methods:
- Utilized an objective function that encodes and reconstructs data via a Markov process.
- Employed density modeling techniques for component extraction.
- Applied the method to large handwriting image datasets.
Main Results:
- Successfully discovered localized components corresponding to pen-made line- and curve-segments.
- Components identified were specific to the size-scale and structure of the handwriting.
- Avoided the extraction of general wavelets, focusing on specific structural elements.
Conclusions:
- The method enables unsupervised discovery of meaningful components in large, homogeneous image datasets.
- The approach has implications for analyzing structured data beyond natural scenes.
- The underlying neural model also relates to speech processing and symmetry breaking discoveries.