Technical Report Number
Most handwritten character recognizers use either graphical (static) or first-order dynamic data. Our research speculates that the mental signal to write a digit might be partially encoded as an energy profile. We used artificial neural networks (ANN) to analyze energy-related features (first and second time derivatives) of handwritten digits of 20 subjects and later 40 subjects. An experimenal environment was developed on a NeXTstation with a real-time link to a pen-based GO computer.
Although such an experiment cannot confirm an energy profile encoded in the writer, it did indicate the usefulness of energy-related features by recognizing 94.5% of the 600 test patterns after 29,000 random presentations of 800 training digits. This three-layer ANN had 54 input units (representing 29 trinary features), 4 hidden units (0, 2, 3, 4, 8, 10, 12, 15, 20, and 25 hidden units were tested), and 14 output units. Another ANN recognized 91.7% of the test digits after only 6,000 training presentations. Later testing with 40 subjects (including very erratic writing) resulted in 91% recognition.
This same feature abstraction was tested in a Supervised Competitive learning (SCL) implementation which was free to create as many or few digit prototypes as was needed to recognize characters. Using the former data set (20 subjects), it created from 19 to 34 prototypes (depending on control parameters) and achieved 94.8% recognition.
Fuller, Thomas H. Jr., "Energy-related Feature Abstraction for Handwritten Digit Recognition" Report Number: WUCS-92-17 (1992). All Computer Science and Engineering Research.