查詢結果分析
來源資料
相關文獻
- 改良式漢明類神經網路應用於光學字元辨認
- Modified Hamming Neural Network and Its Character Recognition System
- Neural-net Decoders for Linear Block Codes
- Character Recognition Using an Assembly of SSO Neural Networks
- An Overview of RNN-Based Mandarin Speech Recognition Approaches
- 快速學習類神經網路手寫英數與中文字辨認系統
- An Overview of Mandarin-Speech Tone Recognition
- 肺部聲音辨認系統之研究
- 將時間延遲類神經網路與動態規劃結合以用於國語數字之辨認--以實體線路觀點剖析
- 類神經網路應用於文字辨認之研究
頁籤選單縮合
題 名 | 改良式漢明類神經網路應用於光學字元辨認 |
---|---|
作 者 | 周義昌; | 書刊名 | 電信研究 |
卷 期 | 22:3 1992.06[民81.06] |
頁 次 | 頁335-351 |
專 輯 | 類神經網路專集 |
分類號 | 312.2 |
關鍵詞 | 光學字元; 改良式; 漢明; 辨認; 類神經網路; |
語 文 | 中文(Chinese) |
中文摘要 | 在本論文中,我們用最小距離分類的觀念,來改良傳統漢明類神經網路(Hamming Neural Network)的學習法則。藉由設定認知(Perceptron)網路不同的臨限值(Threshold),可調整改良式網路對雜訊的容忍度。經實際模擬結果,改良式漢明類神經網路在和傳統網路具有相同的雜訊容忍度下,其收斂速率是優於傳統的漢明類神經網路。此外,我們以改良式漢明類神經網路來發展一光學字元辨認系統。系統能辦認Courier字體之大小寫英文字母及數字,共62字元。每個字元之特徵向量的取得,在於計算一組動量值(Moments),並編碼成二元碼,再由改良式網路來分類。實驗測試以中央動量(Central Moments)為字元之特徵向量時,能得到最好的辦認正確率和最快的收斂速率。因採取之編碼方式使不同字元種類間的漢明距離不夠大,造成系統之收斂速率不如預期理想,我們再提出三個方法以進一步改善收斂速率。由實驗證明,改良式網路不僅辦認正確率上優於傳統網路,其收斂速率亦快於傳統網路。系統之辦認正確率在97%以上,而平均重要次數為5.5次。對於快速與正確辦認認二項要求橋言,改良式漢明類神經網路比傳統網路更適合於完成光學字元辦認系統。 |
英文摘要 | In this paper, a minimum distance classification method based learning rule is proposed to speed up the convergence of the conventional Hamming neural network. In order to compare the modified Hamming neural network with conventional ones, the convergent speed and noise tolerance, consisting of 26 alphabet and 10 digit patterns, have been appropriately arranged to test the two networks. The result of experiment evidence shows that two networks have some noise tolerance, but the convergent speed of modified network is faster than conventional ones above 10 times. Another advantage isn our system is the noise tolerance of modified network can be adjusted by assigning the of different threshold values of PERCEPTRON network. The second task in this paper is to build an optical character recognition system (OCR) which is based on the modified Hamming neural network. The principle of OCR is most like the method of human being to recognize character. Neural networks have the ability of simulation the recognize of human being due to the character of the parallel and distributed processing. Our OCR system can recognize 62 Courier font characters, including upper case, lower case, and digit characters, printed by 24 pin printer. On the other hand, the feature vector extracted from one character is generated by a set of moments of that character. Finally feature vector is encoded into binary vector and input to modified network to be classified . experiment results show that we can obtain a high recognition rate and fast convergent speed when feature vector consists of central moments. Because the Hamming distance of different character class computed by our encode method is not large enough so that the convergent speed of system is not satisfied. We propose three methods (1) to reinforce to effect of Hamming distance, (2) to change the connection weight assignment for MAXNET network, (3) to enlarge the output of processing elements in PERCEPTORN network. The experiment shows that nat only the recognition rate of modified Hamming neural network is better than conventional ones, but also the iteration of convergent of modified network only needs 33.5 times on average. From the convergence speed and correct recognition, modified net work is more adequate to implement an OCR system than conventional network. |
本系統中英文摘要資訊取自各篇刊載內容。