Javascript must be enabled for the correct page display

Learning to Learn: An Adaptive Visual Object Recognition Approach for Handwritten Text Recognition

Haak, K.V. (2007) Learning to Learn: An Adaptive Visual Object Recognition Approach for Handwritten Text Recognition. Master's Thesis / Essay, Artificial Intelligence.

AI_Ma_2007_KVHaak.CV.pdf - Published Version

Download (1MB) | Preview


After some years of experience, humans read handwritten texts in a remarkably effortless and swift manner. While during development visual processing streams in the human brain adjust to perform the task of handwritten text recognition fluently, visual processing and recognition strategies become highly specialized. This thesis therefore approaches the problem of automated handwritten text recognition from a developmental perspective, wherein adaptive visual processing mechanisms specialize themselves to perform the task of recognizing writings. Three well justified assumptions form the base of the approach. First, it is assumed that humans are provided with a basic mechanism of visual processing. Accounting for this assumption, the standard model of visual processing of Poggio and Edelman [70] is the cornerstone of the proposed approach and is used to represent images of handwritten texts. Motivated by the first 100-200 milliseconds of primate visual processing, the resulting representations are well suited for particular discrimination tasks. Secondly, it is assumed that the cortical processes underlying handwritten text recognition are merely a specialized form of general visual processing. This assumption is accounted for by specializing the previously mentioned representations of handwritten texts even more, while additionally specializing artificial neural network classifiers for particular discrimination tasks. Thirdly, based on the concept of Neural Darwinism of Edelman [23], it is assumed that specialization of this kind can be modelled with evolutionary algorithms. Incrementally, a handwritten text recognizer is developed within this paradigm, whereas each step if experimentally assessed. First. it is demonstrated that the representations resulting from the use of the standard model of visual processing on handwritten texts are excellent for classification tasks. Secondly, using these representations, artificial neural networks with integrated feature selection abilities evolve into task-specific binary classifiers. After specialization, the task-specific representations are very sparse, while the artificial neural networks are very simple perfrom nearly flawless classification. Lastly, the task specific classifiers are combined to address multiple class recognition problems. With room for improvement, the resulting handwritten text recognizer needs very little characteristics to recognize handwritten texts accurate.

Item Type: Thesis (Master's Thesis / Essay)
Degree programme: Artificial Intelligence
Thesis type: Master's Thesis / Essay
Language: English
Date Deposited: 15 Feb 2018 07:30
Last Modified: 15 Feb 2018 07:30

Actions (login required)

View Item View Item