Munkong, Rungsun (rmunkong@stanford.edu)
 Mason, Chris (mason@stanford.edu)
 Mathur, Somit (somit@stanford.edu)
 
 
 Numeral Recognition Proposal
 
 
 The objective of our project is to apply three different recognition
 techniques to identify handwritten numerals.  The three techniques we
 propose to investigate and implement include the following:
 
 - Wavelet Transform (using 2-D Haar wavelets)
 - Feature Recognition (Structural Analysis)
 - Eigenimage Recognition
 
 -Wavelet Transform
 Images and patterns such as handwritten numerals usually contain many
 transients or localized elements.  Such localization in spatial domain
 cannot be captured efficiently by the family of Fourier transforms.
 Wavelet transform, on the other hand, can capture the localization in 
 time and frequency quite efficiently.  It thus can be used as an
 efficient alternative for the analysis of image and patterns including 
 handwritten numerals.  Our project will utilize the multiresolution
 features extracted by 2-D Haar wavelet of the numerals (based on  [1]:
 Lee et al., 1996).
 
 The use of Haar wavelet is sufficient for local detection of line 
 segments and global detection of line structures.  Character images at
 different resolution and the wavelets coefficients can be analyzed in
 the recognition of unconstrained handwritten numerals.
 
 -Feature Recognition
 Feature recognition entails normalizing the numeral images and
 extracting the features that characterize each individual numeral. 
 Feature recognition will be aimed at essentially capturing the skeleton
 or features of the individual characters. These properties include the
 arcs, junctions, endpoints, loops, strokes, etc. which will be used in
 subsequent detection schemes. 
 
 -Eigenimage Recognition
 After gathering the training data, we generate the eigenimages through
 the autocorrelation matrix and use this set of eigenimages to detect
 the appropriate numeral. Numerous preprocessing and decision rules will
 be tested to optimize the eigenimage method. 
 
 Each group member will be responsible for the research, design, and
 implementation of their algorithm.  The other two members will offer
 feedback and suggestions to help each other improve their algorithm, 
 and overcome challenging problems.
 
 We will also investigate the performance of a majority vote based on 
 our decoding schemes.  For example, if two algorithms decode 8, and the
 
 other algorithm decodes 9, the overall output would be an 8.  If all 
 algorithms differ, the output is rejected.  By using a majority voting
 system, we hope to improve our recognition percentage.
 
 The scope of the project does not include scaling, rotating, or 
 separation of numbers.  We plan to use a fixed size, black and white
 image for testing our algorithms.
 
 Thank you for considering our proposal for our EE368A project.  Please
 feel free to offer any feedback you feel will help improve our project.
 
 
 

 
 [1] Lee, S. W., Kim, C. H., Ma, H., and Tang Y. Y. (1996).
 Multiresolution recognition of unconstrained Handwritten Numerals with
 Wavelet Transform and multilayer cluster neural network.  Pattern
 Recognition , 29:195-1961.
 
 [2] Computer Recognition of Unconstrained Handwritten Numerals.
 Ching Y.Suen et al, Proceedings of the IEEE, Vol.80, No.7, 
 pp.1162-1180, July, 1992.