Browse Prior Art Database

A character recognition system using neural sub-networks for similarly shaped characters Disclosure Number: IPCOM000015401D
Original Publication Date: 2002-Oct-10
Included in the Prior Art Database: 2003-Jun-20
Document File: 5 page(s) / 77K

Publishing Venue




This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 54% of the total text.

Page 1 of 5

  A character recognition system using neural sub-networks for similarly shaped characters


This disclosure describes a character recognition system that adapted two kinds of neural networks. After the primary network outputs the results, the system judges whether similarly shaped character sets exist in candidate characters. If they are detected, the system starts the secondary neural network processing for performing detail discrimination, and obtains the whole recognition result.

Detail of invention:

There are about eighty categories for Japanese Katakana set. Thus, it is very difficult to get high recognition ratio by single neural network processing. The reason is that many similarly shaped characters like " (KO)", " (YU)", and " (YE)" exist in Katakana character set. Nevertheless, the probability that correct answer is included in top three candidates is enough high.

However, if the number of categories is small enough, such as numerals (e.g. ten categories), it is possible to discriminate similarly shaped characters by single neural network with high accuracy.

This disclosure describes the system that equips two kinds of neural networks --- primary network and sub-network. If the primary network outputs similarly shaped characters in its candidates, the system performs the secondary neural network (sub-network) processing, and gets the detailed discrimination results. By using two layered neural networks, the recognition rate is improved.

Figure 1 shows the whole processing of this recognition system. Image data for a character is segmented, and passed into character recognition engine. The feature vector is extracted from the image data by feature extraction process. Recognition is performed by primary neural network using the extracted feature vector. The candidate categories are output by primary neural network.

Additional recognition process using sub-network is performed in Figure 2. It depends on the results of primary network processing. If the results contain the similarly shaped characters among third candidates, this process will be done. One sub-network is selected from Table 1 by fitting candidate distribution.

Figure 2 is an example of sub-network discrimination process.

(1) Pick up the first and the second candidates category from the primary...