| 摘要 |
<p>PURPOSE:To extract a character part from plural candidate character images with a high recognition rate even when the strokes of characters are overlapped on character frames by providing the plural candidate character images while using a neural circuit model. CONSTITUTION:An input image A is inputted and an associated image B1 is provided by using the pattern association storage ability of a neocognitron as a processor utilizing the neural circuit model. The part of a dotted line shows a part not to be expressed as an image. Concerning this associated image B1, the associated image is changed from an associated image B2 to an associated image B4 by switching attention. Next, frame removed images C1-C4 are prepared by executing the frame removing processing of associated images B1-B4. On the other hand, a frame removed image E is prepared from the input image A by executing frame removing processing, and an overlap rate is decided for each of the frame removed image E and the frame removed images C1-C4. Recognizing processing is executed to adopt the associated image B3 as a character candidate for which this overlap rate is set within a certain fixed range.</p> |
