GRAPHICAL-DATA-PROCESSING RESEARCH STUDY AND EXPERIMENTAL INVESTIGATION.
Quarterly rept. no. 3, 1 Aug-31 Oct 66,
STANFORD RESEARCH INST MENLO PARK CALIF
Pagination or Media Count:
The results of experiments on piecewise-linear learning machines with handprinted characters are reported. The data, as before, consisted of three 46-character FORTRAN alphabets from each of 16 writers. The results of Experiment 3, a nine-view experiment, were compared with seven, five, three, and one-view results on the same training and testing set. The test error rate on characters not included in the training set fell steadily from 42 percent with the single view to 24 percent when the category of the character was determined from nine views. Results are reported on the program, which includes both preprocessing and classification. It is based on contour following and the determination of features including connected subfigures, strokes, concavities, and enclosures. The particular problem treated is the separation of 6s for Gs. A method of using the 1024-image optical preprocessor for the proprocessing of speech waveforms is described and some of the considerations that govern its practical implementation are discussed. The method is based on the determination of the amplitude coefficients of a Fourier series. Author
- Computer Hardware