Recognition of Handwritten Numbers by eYV9Eq


									Recognition of Handwritten Numbers Using Artificial Neural Networks
Alexander Gutierrez
Mentor: Amar Raheja
California State Polytechnic University, Pomona
Computers are widely used for different applications ranging from communication to the modeling
of supernovae. However, despite the capability and speed of modern computers and even
supercomputers, the human brain has yet to be surpassed in efficiency in tasks such as speech and
visual recognition. While many tasks are automated in the modern world, trivial tasks for humans
such as reading a license plate cannot be done reliably with even the most advanced computers.
Artificial neural networks are an important and influential approach to attempting to bridge this gap
in technology. In this experiment, several configurations of multilayer perceptrons, a type of artificial
neural network, are tested using images of seed germination trials with numbers written for batch
identification. The purpose of this study is to analyze efficiency ratings of different architectures of
multilayer perceptrons in recognizing handwritten numbers and develop an application that can be
used to read numbers associated with seed batch specimens. A number of configurations were
trained and tested using the MNIST database of digits where number of layers, nodes in each layer,
and training set size were varied. From the survey, a two-layer network with 300 hidden nodes was
chosen for further training and testing. This network was trained with the conjugate gradient
algorithm and achieved a digit accuracy of over 95% with a larger training set. Image preprocessing
was added in order to create an application to fully automate the identification of the seed
germination photographs.

To top