| Project/Area Number |
02680033
|
|---|
| Research Category |
Grant-in-Aid for General Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Research Field |
Informatics
|
|---|
| Research Institution | Osaka Electro-Communication University |
|---|
Principal Investigator |
UMEDA Michio Osaka Electro-Communication University, Faculty of Engineering, Professor, 工学部, 教授 (30213490)
|
|---|
| Project Period (FY) |
1990 – 1991
|
| Project Status |
Completed (Fiscal Year 1991)
|
| Budget Amount *help |
¥1,600,000 (Direct Cost: ¥1,600,000)
Fiscal Year 1991: ¥600,000 (Direct Cost: ¥600,000)
Fiscal Year 1990: ¥1,000,000 (Direct Cost: ¥1,000,000)
|
|---|
| Keywords | Human Sensibility / Character Recognition / Neural Network / Back-propagation Learning / Texture Analysis / Scene Understanding / Color Image Processing / Japanese Family Name / 誤差逆伝搬学習 / テクスチャ識別 / Hough変換 / 局所方向寄与度 / 単語理解 / 部首理解 / バックプロパゲ-ション / 文字列抽出 |
|---|
| Research Abstract |
Regarding the recognition of characters information based on human sensibility, we studied from two different view points. One is the elucidation of human flexible processing mechanism and the other is the adaptive application method of abundant linguistic knowledge.
1. To elucidate the behavior and the ability of neural networks which simulate the human flexible processing mechanism, the distinction of textural images was studied for example. The gray level co-occurrence matrix feature and the gray level difference feature are extracted from textual images. The 3-layer neural network was used and the back-propagation learning argolithm was applied. From the experiment for 10 different natural textures, alaest perfect distinction was performed.
2. The distinction between U&-written Kauji and Hiragana characters using neural networks was studied. It was shown that 94 to 96 percents correct rates are obtained for unknown data sets which contain same character categories used in the teaming process.
3. The segmental extraction method of characters owtaiaw in scene images using hue and value information was proposed. The math& can extract 68 percent of characters contained in 20 different scene images.
4. We studied the estimation ability for uncertain characters utilizing Japanese words knowledge. It was shown that 72 percent of words are correctly estimated when the right part of Kanji "Tsukuri" is unknown but the left part "Hen" is clear. Samely, 95 percent of words are correctly estimated when Hen is unknown but Tsukuri is clear.
5. Japanese family names database was constructed in a part of knowledge processing concerned in Japanese linguistics. This database is constructed from 71 thousand family names and 3.8 thousand character categories.
|
