| 摘要 |
PURPOSE:To obtain a high recognition accuracy irrespective of the number of learning samples by correcting the threshold of a category obtained from the learning sample in accordance with the number of learning samples. CONSTITUTION:An observation part 1 converts an input pattern A into a signal B, and an extraction part 2 preprocesses, normalizes and extracts the feature of the signal B to generate a feature pattern C, which is stored in a learning sample memory 3, and then read out. A calculation part 4 calculates the standard pattern and threshold of the category C1 from the learning samples U1-Un, which are stored in a dictionary memory 5. Then the standard pattern and threshold are read out of the memory 5, corrected in a correction part 6 in accordance with the number of learning samples using the range of the category C1, and stored in a dictionary memory 7. Said processing is executed at every category, thereby obtaining a desired dictionary in the memory 7.
|
