201201113 六、發明說明： 【發明所屬之技術領域】 本發明大體上係關於字符輸入。更具體而言，本發明係 關於一種具有改良的輸入效率之手寫辨識方法及對應裝 置，其可辨識由使用者連續輸入的無書寫框（writing_box_ free)字符序列》 【先前技術】 菖刖，手寫辨識模組已被廣泛使用於諸如行動電話的所 有種類之電子裝置中。使用者可便利地與該等電子裝置互 動。利用該等手寫辨識模組，使用者無需學習其他藉由按 鍵盤的字符輸入方法。 非專利文獻1(參見以下）揭示一種設計分割型樣之實體 特徵（斷開筆劃特徵（〇ff_str〇ke features乃以辨識一無書寫框 字f序列的手寫辨識方法。在這個方法中，斷開筆劃資訊 可從前-筆劃的最後取樣點及下一筆劃的第—取樣點獲 取’其在圖4如虛線所表示。該實體資訊進一步包含諸 如：割型樣之寬度/高度及該等對應分割型樣之手寫時間 的育訊。在這個方法中，該實體資訊包含該等分割型樣之 ㈣㈣4f特徵及間隙特徵；筆劃長度；冑開筆劃之 一平均距離；斷開肇書彳 章畫]之一千均時間；斷開筆劃之距離； 该等斷開筆劃之角戶 、土隹+ — 又正弦及餘弦及斷開筆劃間隙。此方 法集中於從前一筆劃士 ^ * τ_ 、'，°束··.占到®前筆劃之開始點的斷開 D處理，且因此辨識手寫輸入。 此手寫辨識方法假設即蚀产丁 Ρ使在不同字符之間發生連接的手 149J39.doc 201201113 寫，字符間之斷開筆劃的距離及時間週期兩者應大於該等 字符内的斷開筆劃之距離及時間週期。此方法亦假設各個 筆劃分佈配合一常態分佈。基於此等假設，此手寫辨識方 法基於„玄4特徵之平均值及變異數藉由使用一機率模型而 十#刀割型樣之似然度。最後，此方法藉由使用動態規劃 (DP)測定一最佳分割路徑（segmentation path)。 上述非專利文獻1中存在的一個問題在於該手寫字符序 列之分割依賴於各個筆劃的手寫時間。斷開筆劃的時間週 期在此方法中係極為重要之特徵6此方法假設在分割型樣 2間之斷開筆劃的時間週期越大，分割精度越高。上述假 設在使用者卩一相對恒定的速度書寫時是合理的。然而， 在使用期間，使用者通常以不同速度書寫，例如快速書寫 一段時間隨後慢速書寫—段時間° SUb ’如果使用者在手 寫處里期間改變書寫速度，非專利文獻^令所揭示的方法 將極難以精確分割該等手寫。 存在於上述非專利文獻丨的$ —個μ題在於此方法僅使 用幾何特徵及時間特徵以測定該分割是否正確。此方法假 設字符間之斷開筆劃的距離大於該等字符内之筆劃之間之 ____ H此—假設並非一直正確。非專利 文獻臺】列出幾個分割錯誤的典型實例，如圖2令所示。從圖 °纟在一些字符之間之斷開筆劃的距離小於字符内之 筆劃的距離。如圖2中的第一實例中所示，「子5 = Π;内，筆劃之間之過大的間隙而被分割過頭。但如 第一及第二貫例中辦+ 木 ΧΚ '、，虽一輸入字符序列之字符之間的 149I39.doc 201201113 距離戲劇化地改變且該等字符之大小明顯不同時，分割錯 誤發生。 引用列表 非專利文獻1 使用斷開筆劃特徵的非限制手寫字串之線上字符分割 方法（Onime Character Segmentation Method for Unc〇_ained201201113 VI. Description of the Invention: TECHNICAL FIELD OF THE INVENTION The present invention generally relates to character input. More specifically, the present invention relates to a handwriting recognition method and corresponding apparatus having improved input efficiency, which can recognize a characterless (writing_box_free) character sequence continuously input by a user. [Prior Art] 手写 Handwriting Identification modules have been widely used in all kinds of electronic devices such as mobile phones. The user can conveniently interact with the electronic devices. With these handwriting recognition modules, the user does not need to learn other character input methods by pressing the keyboard. Non-Patent Document 1 (see below) discloses a physical feature for designing a segmentation pattern (a break hand feature (〇ff_str〇ke features to identify a handwritten recognition method without a frame f sequence. In this method, disconnection) The stroke information can be obtained from the last sampling point of the front-stroke and the first sampling point of the next stroke. It is indicated by a broken line in Fig. 4. The entity information further includes, for example, the width/height of the cut pattern and the corresponding split type. In this method, the entity information includes (4) (4) 4f features and gap features; stroke length; one of the average distances of the strokes; The average time; the distance of the strokes; the corners of the strokes, the band + and the sine and cosine and the gap between the strokes. This method focuses on the previous stroke ^ * τ_, ', ° bundle · · Takes off the D processing of the starting point of the pre-stroke, and thus recognizes the handwriting input. This handwriting recognition method assumes that the eclipse produces a hand that connects between different characters. 149J39.doc 2012011 13 Write, the distance and time period of the disconnected stroke between characters should be greater than the distance and time period of the broken stroke in the characters. This method also assumes that each pen is divided into a normal distribution. Based on these assumptions, this The handwriting recognition method is based on the eigenvalue of the eigen-4 feature and the likelihood of the variation by using a probability model. Finally, the method determines the optimal segmentation path by using dynamic programming (DP). A problem in the above non-patent document 1 is that the division of the handwritten character sequence depends on the handwriting time of each stroke. The time period in which the stroke is broken is an extremely important feature in this method. The larger the time period of the split stroke between the two types, the higher the segmentation accuracy. The above assumption is reasonable when the user writes at a relatively constant speed. However, during use, the user usually writes at different speeds. For example, writing quickly for a while and then writing slowly - for a period of time ° SUb 'If the user changes the writing speed during the handwriting, non The method disclosed in the document will make it extremely difficult to accurately segment the handwriting. The $-μ problem that exists in the above non-patent document uses only geometric features and temporal features in this method to determine whether the segmentation is correct. The distance between the break strokes of the characters is greater than the ____ between the strokes in the characters. This is not always true. The non-patent literature table lists several typical examples of segmentation errors, as shown in Figure 2. The distance from the stroke between the characters is smaller than the distance between the strokes in the characters. As shown in the first example in Figure 2, "sub 5 = Π; inside, the excessive gap between the strokes It is split too much. But as in the first and second examples, the 149I ΧΚ ', while the character of the input character sequence 149I39.doc 201201113 changes dramatically and the size of the characters is significantly different The split error occurred. LIST OF REFERENCES Non-Patent Document 1 On-line character segmentation method for unrestricted handwriting strings using broken stroke features (Onime Character Segmentation Method for Unc〇_ained
Handwriting Strings Using Off-stroke Features), ： a ^ 有限公司’第十屆手寫辨識尖端國際研討會，法國“Handwriting Strings Using Off-stroke Features), : a ^ Ltd. 'The 10th International Conference on Handwriting Recognition, France"
Baule，2006年） 【發明内容】 本發明之技術目的在於提供—種手寫辨識方法及裝置， 其等能辨識由使用者連續輸入的—字符序列而不考慮書寫 根據本發明之—態樣，提議—種辨識由使用者連續輸入 之一無書寫框字符序列的手耷 ^ 厅㈣手寫辨識方法。該方法包括：計 鼻相對於该輸入字符序列中之不同筆劃組合之單一字符辨 識精度的特徵，其係基於不同筆 , A J革蛋j組合及藉由分割該等筆 ^、·且a中的筆劃而形成的子 田 罕' -J 且5之單一子符辨識么士 果；根據藉由分割該等筆書Ρ且人+ 子付辨識.，。 筆畫“,入* 寺㈣組合中之筆劃而形成的該等子 #畫！組合之空間幾何關係 幾何#舛.《 』疋。亥專不同筆劃組合之空間 間幾何特徵測定在不同㈣型樣：4:度的特徵及該等空 筆^^^ 輸人字符序列之分別 章蒼j組合的分割可靠性； 徑，M S # μ + & …亥專刀割可靠性測定分割路 及根據該等經測定的分卹路牺a 刀。』路徑向使用者呈現該等字 149139.doc 201201113 符序列辨識結果。 由mr月之其他態樣’提議—種手寫辨識裝置以辨識 者連續輸人的-無書寫框字符 置包括：—手耷於λ留_ # 于舄辨識裝 鈐入“寫輪入皁疋，其經組態以收集由使用者連續 輸入的字符序列；一 埂瓚 該字符序列中的不… 組態以辨識 -八職 劃組合並獲取單—字符辨識結果； 該二’其經組態以便基於不同筆劃組合及藉由分割 Λ i、·且合中之筆劃而形成的子筆劃組合之單一 而計算相對於該輸入字符序列中之不同筆劃組合之 予符辨識精度的特徵並根據該等子筆劃組合之 何關係測定該尊;；^ PJ 4查 相… 空間幾何特徵，以便基於 、早-子符辨識精度的該等特徵及該等空間幾何特徵 t在不同分割型樣中之輸入字符序列之分別筆劃組合的 t割可靠性，並基於該等分割可靠性測定分割路徑；以及 :顯不控制單元，其經組態以控制一顯示器螢幕以便根據 :等經測定之分割路徑向使用者呈現該等字符序列辨識結 果。 由於铋用無書寫框之方式，使用者可連續輸入—字符序 列以便改善手寫輸人效率。至於需要使用者在各個書寫框 内:曰寫各個字符的輸人方法，手寫字符之間的間歇經常中 斷該使用者之思考從而降低輸人速度。需要各個字符被書 寫於規定之書寫框内的方法(例如，普遍當前之行動電話 中的一框輸入方法要求使用者在兩個書寫框之間頻繁切 換)亦改變使用者之手寫習慣並降低手寫輸入效率。然 l49139.d〇c 201201113 而，根據本發明之一實施例的方法及裝置在不改變手寫習 慣的前提下允許連續的字符序列輸入並允許單獨或總體地 輸出辨識結果。 在計算該字符序列之分割可靠性期間本實施例之方法 及裝置不僅考慮常用的空間幾何特徵，同時亦考慮合併筆 劃組合之單—字符精度以及子筆劃組合之單一字符精度， 其結果係其可在正確分割難以被傳統技術執行的情況下， 例如不同字符中的筆劃在空間上部份 的筆劃間隙過大，達成正確分割。 戍者 此外’本實施例之方法及裝置在執行字符序列分.割時不 依賴於各個筆劃之輸入時間，因此其適用於不同的使用者 輸入習慣。即使-使用者時快時慢地輸入字符，根據本實 施例之方法及裝置，該分割精度將不會下降。 此外，在本實施例之方法及裝置中採用的筆劃組合之空 間幾何特徵係基於字符之估算平均寬度或高度的正常化特 徵，因此本實施例之裝置可適用於具有任何大小的一字符 序列。由於在單一字符辨識單元中採用多範本訓練 (training)及多範本匹配方法，藉由不同使用者之不同書寫 型樣中的字符（例如藉由中文的漢字之簡化字符）可藉由本 實施例之方法及裝置而被精確辨識。此外，本實施例之方 法及裝置使用s吾έ模型及字典匹配因此該裝置具有拼寫檢 查及單字校正的功能。 最後’本實施例之方法及裝置的識別對象可為英文單 字、曰文假名組合、中文句子 '韓文字符組合等。執行手 149139.doc 201201113 寫識別的時機可被任意指定。辨識結果可在使用者輸入該 字符序列時連續更新，或者該等辨識結果可在該使用者完 成整個字符序列輸入後顯示。 【實施方式】 在結合諸所附之圖式而考慮本發明之以下詳細描述後， 本發明之上文及其他目的、特徵與優點將更輕易地瞭解。 較佳實施例將藉由參考所附之圖式而被說明。在該等圖 式中，相同的元件符號將用於指示相同或相似的組件，儘 管其被圖解於不同的圖式中。本發明之不必要部份及功能 將因簡潔性而被省去以避免理解之混亂。 圖3係圖解一種根據本發明之一實施例的手寫辨識裝置 之一結構示意圖。 如圖3中所示，根據本發明之一實施例的手寫辨識裝置 被用於辨識由使用者連續輸入的一無書寫框字符序列。該 手寫辨識裝置由以下組成：一手寫輸入單元11〇，用於收 集孩使用者之筆跡並將其數位化為一輸入筆跡信號；一手 寫筆跡儲存單M 2Q，用於保存由該手寫輸人單產生 之輸入筆跡信號；以及一字符序列辨識單元，用於辨 識該輸入子符序列。該字符序列辨識單元13 〇由三個子單 儿組成.分割單元132、單一字符辨識單元131及後處理單 元 13 3 〇 由於知用無書寫框輸入，該使用者可連續輸入一字符序 歹"便改善手寫輪入效率。一辨識結果將在該侠用者輸入 程序』間即時顯示。或者，該總體辨識結果將在該使用者 149139.doc 201201113 輸入完整的句子之後提供。在需要使用者在書寫框内書寫 字符的傳統輸入方法中，在手寫字符之間的間歇經常中斷 該使用者之思考並降低輸入速度。需要各個字符書寫於規 定書寫框内的方法（例如常用於當前行動電話中的兩框輸 入方法要求使用者在兩個書寫框之間頻繁切換）亦改變使 用者之手寫習慣並降低手寫輸人效率。然而，根據本發明 之一實施例的方法及裝置在不改變手寫習慣的前提下允許 連續的子符4列輸入並允許單獨或總體輸出辨識結果。 該分割單元132從輸入筆跡信號提取該輸入字符序列中 之個別筆劃組合的各種空間幾何特徵、藉由呼叫該單—字 符辨識單it 131而獲取個別之筆劃組合的單—字符識別結 果及單子符减別精度，然後基於一邏輯回歸模型計算 :°〗了彝丨生」並藉由使用一種N最佳（N-best)演算法獲取 最4的N個刀割型樣，该演算法將被詳細描述於稱後 份。 «亥後處理單元1 3 3藉由利用語言模型並匹配字典資料庫 而校正該分割單元132之字符序列辨識結果。 如圖3中所示’根據本發明之一實施例的手寫辨識裝置 進v包含一顯示控制單元1 50及一候選物選擇單元丨4〇 ^ 方面，在該使用者在該手寫輸入單元11〇中輸入筆劃 時，該顯示控制單元15〇控制該系統顯示該等筆跡並在一 顯示器螢幕上向該使用者呈現’另一方面，該顯示控制單 π 15〇在該顯示器螢幕上顯示由該字符序列辨識單元丨3〇產 生的辨識候選物供使用者選擇。該候選物選擇單元刚在 149139.doc 201201113 使用者之操作下從對應的候選物選擇該字符序列或單一字 符並向使用者提供辨識結果或者向例如字典應用的其他應 用提供以說明該等辨識結果。 根據本發明之—實施例，使用於該字符序列辨識單元 ?〇中的邏輯回歸模型之截距及回歸係數藉由該等樣本之 資料訓練而被估算。 圖4係圖解根據本發明之一實施例的手寫辨識裝置之— 訓練處理之一流程圓。 仲根據本發明之—實施例，該.資料訓練中的樣本不僅包含 =字符樣本，而^亦包含該等字符中的各個筆劃以及— 子：内的幾個筆劃之一組合或兩個不同字符内的筆劃之一 組合。上述樣本之各者被界定為_個種類的筆劃組合。 于冩筆跡被收集。在步羯 如圖4中所示’在步驟S10中 川中，被收集之資料被添加至一對應的筆劃組合類別。 然後在步驟S12中進行預處理且筆劃組合特 驟S13中。 在該邏輯回歸模型中樣本訓練之特徵為爪維特徵 (…UM)。該等筆劃組合特徵包含子筆劃組合之邊界框 之間之一間隙、合併子筆劃組合一 見反、子筆劃組合之 :之：向量及距離、合併子筆劃組合之一單—字符辨識精 度、δ併辨識精度及該等子筆劃組合之辨識精度之間之一 差、該合併子筆劃組合之第一候课 弟⑮選物之早—字符精度對 他候選物之單一字符精度之一比率等等。 在步驟S13中的特徵計算之前，一 J 預處理須被執行於步 M9I39.doc 201201113 驟S 12中’其根據輸入字符序列之高度及寬度估算—字符 之平均高度Havg及字符之平均寬度Wavg作為該等筆劃組合 之空間幾何特徵之一正常化準備使得根據本發明之—實施 例的手寫辨識裝置可被應用於具有一任意大小的一字符序 列。 根據本發明之一實施例之子筆劃組合（下文簡稱為「子 筆劃」）之概念將藉由採取一字符序列中從第k個筆劃到第 k+3個筆劃的分割之一實例而被說明。從第让個筆劃起，存 在四個可能的分割型樣，如圖5a、5B、5C及5D中所示。 U —筆劃組合僅包含第k個筆劃且不具有子筆劃。 2)二筆劃組合包含第k個及第k+Ι個子筆劃。 3 )二筆劃組合具有兩個子筆劃分類模式。 模式1 :前一個子筆劃為第k個筆劃且下一個子筆劃為第 k+Ι個筆劃及第k+2個筆劃之筆劃組合。 模式2 :前一個子筆劃為第k個筆劃及第k+1個筆劃之筆 劃組合且下一個子筆劃為第k+2個筆劃。 4)四筆劃組合具有三個子筆劃分類模式。 模式1 :前一個子筆劃為第k個筆劃且下一個子筆劃為第 k+1個、第k+2個及第k+3個筆劃之筆劃組合。 模式2 :該前一個子筆劃為第k個筆劃及第k+1個筆劃之 筆劃組合且下一個子筆劃為第k+2個及第k+3個筆劃之筆劃 組合。 模式3 :前一個子筆劃為第让個、第k+1個及第k+2個筆劃 之筆劃組合且下一個子筆劃為第k+3個筆劃。 149139.doc -12- 201201113 割可看到該子筆劃組合可為藉由循序分 來說，對於―心」中之葦劃而形成的不同組合。舉例 合，盆子筆劃Λ寫順序為「k、W2」的-筆劃組 -子筆劃組合可為藉由在筆 割而產生的「子簦創“， 及k+lj之間分 「k+2 子筆劃類別！」或藉由在筆畫「k+lj及 示。」八割而產生的「子筆劃類別2」，如圖％中所 在根據本發a月夕 ^^ , 特勺〜―貫轭例的裝置中，該筆劃組合之各種 特徵（包含該子筆劃組 裡 幾何特徵)為針於” 辨識精度特徵及空間 曾。兮黧久# 序歹1中的所有可能筆劃組合計 "各種詳細特徵被列出如下： ⑷合併子筆劃之—單一字符辨 合併成—單—字符的可純越大；K叫·其越大， (b)兩個子筆劃之合Baule, 2006) SUMMARY OF THE INVENTION The technical object of the present invention is to provide a handwriting recognition method and apparatus, which are capable of recognizing a sequence of characters continuously input by a user without considering writing in accordance with the present invention. A handwriting recognition method for recognizing a handwriting of a frameless character sequence continuously input by a user. The method includes: characterizing a single character recognition accuracy of a combination of different strokes in the input character sequence, based on different pens, AJ leather j combination, and by segmenting the pens, and a The single sub-character of Zita Han '-J and 5 formed by the stroke recognizes the sorcerer; according to the division of the pens and the recognition of the person + child. Strokes, the strokes formed by the strokes in the combination of the temples (four) are combined with the strokes of the combination of the space geometry geometry #舛. 』疋. The spatial characteristics of the different stroke combinations are measured in different (four) styles: The characteristics of 4:degrees and the segmentation reliability of the combination of the characters of the empty pens ^^^ input characters; the diameter, MS #μ + & ...Hai knife cutting reliability measurement segmentation road and according to these The measured cross-scoring road sacrifices a knife. The path presents the user with the word 149139.doc 201201113. The sequence identification result. From the other aspects of mr month's proposal, the handwriting recognition device is used to identify the continuous input. The no-character characters include: - the handcuffs are λ left _ # 舄 舄 舄 舄 “ “ “ “ “ “ “ 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写 写No... configured to identify - eight job combinations and obtain single-character identification results; the second 'is configured to be based on different stroke combinations and sub-strokes formed by dividing Λ i, · and the strokes in the combination a single combination of calculations relative to the input word The different stroke combinations in the sequence of characters are used to identify the features of the precision and determine the respect according to the relationship of the sub-stroke combinations;; PJ 4 to check the spatial geometric features, so as to be based on the early-child character identification accuracy And other characteristics and the t-cut reliability of the respective stroke sequences of the input character sequences in the different segmentation patterns, and the segmentation path is determined based on the segmentation reliability; and: the display control unit, the group The state controls a display screen to present the character sequence identification results to the user according to the determined segmentation path. Since there is no writing frame, the user can continuously input the character sequence to improve handwriting input efficiency. As for the user's need to write in each writing box: the method of inputting each character, the interval between handwritten characters often interrupts the user's thinking and reduces the speed of input. A method in which each character is required to be written in a prescribed writing frame (for example, a box input method in a current mobile phone requires the user to frequently switch between two writing frames) also changes the user's handwriting habits and reduces handwriting. Input efficiency. However, the method and apparatus according to an embodiment of the present invention allows continuous character sequence input and allows the identification result to be output individually or collectively without changing the handwriting habits. The method and apparatus of the present embodiment not only consider common spatial geometric features during the calculation of the segmentation reliability of the character sequence, but also consider the single-character precision of the combined stroke combination and the single character precision of the sub-stroke combination, and the result is In the case where the correct segmentation is difficult to be performed by the conventional technique, for example, the stroke of the stroke in the different characters is too large in the space portion to achieve the correct segmentation. Further, the method and apparatus of the present embodiment do not depend on the input time of each stroke when performing character sequence division and cutting, and thus are applicable to different user input habits. Even if the user inputs characters slowly and slowly, the segmentation accuracy will not decrease according to the method and apparatus of the present embodiment. Moreover, the spatial geometric features of the stroke combinations employed in the method and apparatus of the present embodiment are based on the normalized characteristics of the estimated average width or height of the characters, and thus the apparatus of the present embodiment can be applied to a sequence of characters of any size. Since multi-template training and multi-template matching methods are used in a single character recognition unit, characters in different writing patterns of different users (for example, simplified characters of Chinese characters by Chinese) can be used in the present embodiment. The method and device are accurately identified. In addition, the method and apparatus of the present embodiment use the sui model and dictionary matching so that the device has the functions of spell check and word correction. Finally, the identification method of the method and apparatus of the present embodiment may be an English word, a pseudonym combination, a Chinese sentence 'Korean character combination, and the like. Execution hand 149139.doc 201201113 The timing of writing recognition can be arbitrarily specified. The recognition result can be continuously updated when the user inputs the character sequence, or the recognition result can be displayed after the user completes the entire character sequence input. The above and other objects, features and advantages of the present invention will become more <RTIgt; The preferred embodiment will be described with reference to the accompanying drawings. In the figures, the same element symbols will be used to refer to the same or similar components, even though they are illustrated in different drawings. Unnecessary parts and functions of the present invention will be omitted for simplicity to avoid confusion. Fig. 3 is a block diagram showing the structure of a handwriting recognition apparatus according to an embodiment of the present invention. As shown in Fig. 3, a handwriting recognition apparatus according to an embodiment of the present invention is used to recognize a sequence of a characterless character input continuously input by a user. The handwriting recognition device is composed of a handwriting input unit 11〇 for collecting the handwriting of the child user and digitizing it into an input handwriting signal; a handwriting handwriting storage list M 2Q for saving the handwriting input by the handwriting a single generated input handwriting signal; and a character sequence identification unit for identifying the input sub-sequence sequence. The character sequence recognizing unit 13 is composed of three sub-singles. The dividing unit 132, the single character recognizing unit 131, and the post-processing unit 13 3 连续 can input a character sequence continuously by using the input box. Improve handwriting rounding efficiency. An identification result will be displayed instantly between the user input program. Alternatively, the overall identification result will be provided after the user has entered the complete sentence at 149139.doc 201201113. In conventional input methods that require the user to write characters in the writing box, the pause between handwritten characters often interrupts the user's thinking and reduces the input speed. A method in which each character is required to be written in a prescribed writing frame (for example, a two-frame input method commonly used in current mobile phones requires the user to frequently switch between two writing frames) also changes the user's handwriting habits and reduces handwriting input efficiency. . However, the method and apparatus in accordance with an embodiment of the present invention allows for continuous sub-character 4 column inputs and allows for separate or overall output of the recognition results without changing handwriting habits. The dividing unit 132 extracts various spatial geometric features of the individual stroke combinations in the input character sequence from the input handwriting signal, and obtains the single-character recognition result and the single-child character of the individual stroke combination by calling the single-character identification list it 131 The accuracy of the subtraction is then calculated based on a logistic regression model: °" and the N-best algorithm is used to obtain the most N N-cut patterns, the algorithm will be The detailed description is referred to as the latter. The post-harvest processing unit 133 corrects the character sequence identification result of the dividing unit 132 by using the language model and matching the dictionary database. As shown in FIG. 3, the handwriting recognition apparatus according to an embodiment of the present invention includes a display control unit 510 and a candidate selection unit ,4〇^, in which the user is at the handwriting input unit 11 When the stroke is input, the display control unit 15 controls the system to display the handwriting and presents the gesture to the user on a display screen. On the other hand, the display control unit π 15〇 displays the character on the display screen. The identification candidate generated by the sequence identification unit 供3〇 is selected by the user. The candidate selection unit just selects the character sequence or single character from the corresponding candidate under the operation of the user 149139.doc 201201113 and provides the recognition result to the user or provides to other applications such as a dictionary application to explain the identification result. . According to an embodiment of the invention, the intercept and regression coefficients of the logistic regression model used in the character sequence identification unit are estimated by training the data of the samples. 4 is a flow chart of a training process of a handwriting recognition apparatus according to an embodiment of the present invention. According to the embodiment of the present invention, the sample in the data training includes not only the = character sample, but also includes each stroke in the characters and a combination of one of several strokes or two different characters. One of the strokes within the combination. Each of the above samples is defined as a combination of strokes of _ types. Yu Yu's handwriting was collected. In the step 羯 as shown in Fig. 4, in step S10, the collected data is added to a corresponding stroke combination category. Preprocessing is then performed in step S12 and the stroke combination is in S13. The sample training feature in this logistic regression model is the claw dimension feature (...UM). The stroke combination feature includes a gap between the bounding boxes of the sub-stroke combination, the combined sub-stroke combination, and the sub-stroke combination: the vector and the distance, the combined sub-stroke combination, the single-character identification accuracy, δ and The difference between the identification accuracy and the recognition accuracy of the sub-stroke combinations, the early-character accuracy of the first waiting classmate 15 of the combined sub-stroke combination, the ratio of the character precision to the single character precision of his candidate, and the like. Before the feature calculation in step S13, a J preprocessing has to be performed in step M9I39.doc 201201113 step S12, which is estimated based on the height and width of the input character sequence - the average height Havg of the character and the average width Wavg of the character. One of the spatial geometric features of the stroke combinations is normalized so that the handwriting recognition apparatus according to the embodiment of the present invention can be applied to a sequence of characters having an arbitrary size. The concept of a sub-stroke combination (hereinafter simply referred to as "sub-stroke") according to an embodiment of the present invention will be explained by taking an example of a division from the k-th stroke to the k+3-th stroke in a sequence of characters. From the first stroke, there are four possible split patterns, as shown in Figures 5a, 5B, 5C and 5D. U—The stroke combination contains only the kth stroke and does not have a substroke. 2) The two stroke combination includes the kth and k+1th substrokes. 3) The two-stroke combination has two sub-pen division modes. Mode 1: The previous substroke is the kth stroke and the next substroke is the stroke combination of the k+th stroke and the k+2 stroke. Mode 2: The previous substroke is the stroke combination of the kth stroke and the k+1th stroke and the next substroke is the k+2 stroke. 4) The four-stroke combination has three sub-pen division modes. Mode 1: The previous substroke is the kth stroke and the next substroke is the stroke combination of the k+1th, k+2th, and k+3th strokes. Mode 2: The previous substroke is the stroke combination of the kth stroke and the k+1th stroke and the next substroke is the stroke combination of the k+2th and k+3th strokes. Mode 3: The previous substroke is the stroke combination of the first, the k+1th, and the k+2 strokes, and the next substroke is the k+3th stroke. 149139.doc -12- 201201113 Cut can see that the sub-stroke combination can be a different combination of the strokes in the "heart" by sequential division. For example, the stroke group-sub stroke combination in which the basin strokes are written in the order of "k, W2" can be divided into "k+2" by "sub-creative" generated by the pen-cutting, and k+lj. "Stroke category!" or "sub-stroke category 2" produced by the stroke "k+lj and s.", as shown in Figure 5%, according to the present issue, a ^^, special spoon ~ yoke example In the device, the various features of the stroke combination (including the geometric features in the sub-stroke group) are for the purpose of "identifying the accuracy characteristics and space." 兮黧久# All possible stroke combinations in the sequence 1 " various detailed features It is listed as follows: (4) Merging sub-strokes—single character recognition into a single-character can be purely larger; K is called, the larger it is, (b) the sum of two sub-strokes
度。及‘之間的一差(2;^，及B =1意未者從兩個筆劃合併成一單一字符的一可能性 ::π㈣分別為單—字符的一可能性。該差越大， σ併成一單一子符的可能性越大； (Cut子筆劃之第一候選物之單-字符辨識精度 寺合併子筆劃之其他候選物之單一字符辨識精 :1=一，代表該單一字符辨識之第Τ個候選 物’且Τ之值可被設定)：如果該比率相對較大…味著 :於该早一字符辨識’該合併筆劃組合及第-候選物之間 的一匹配距離相當近且哕人 ；<间 且°亥合併筆劃組合及其他候選物之間 H9J39.doc -13- 201201113 的匹配距離很遠，其指示合併成一單一字符的可能性相對 較大； (d)子筆劃之兩個邊界框之間之一間隙gap/w^ (或 gap/Havg):該等子筆劃之間隙越小，在合併後形成一單— 字符的可能性越大。如果該間隙為一負值，在合併後形成 一單一字符的可能性更高； ⑷一合併子筆劃寬度Wmerge/Wavg (或Wmerge/Havg):該 合併寬度越小’形成一單一字符的可能性越大； (f) 前一個子筆劃之結束取樣點及下一個子筆劃之開始 取樣點之間之一向量Vs2.ei/Wavg (或Vs2.el/Havg); (g) 前一個子筆劃之結束取樣點及下一個子筆劃之開始 取樣點之間之一距離ds2_el/wavg (或ds2_ei/Havg); (h) 該前一個子筆劃之開始取樣點及下一個子筆劃之開 始取樣點之間之一距離ds2_sl/Wavg (或ds2_s丨/Havg)。 在上述特徵中，「/」代表一除法記號，且Wavg及代 表在預處理程序期間的估算字符平均寬度及字符平均高 度。（d)-(h)之空間幾何特徵參考圖6A至6D且該等圖式中之 點代表各個筆劃之一開始點。 對於上述特徵（a)、（b)及（c) ’該等合併子筆劃之單—字 符辨識精度Cmerge及其他候選物精度CmergeT以及兩個子筆書 之單一字符辨識精度Cstrl及Cstr2藉由在步驟S14中呼叫該單 一字符辨識單元而獲取。 根據本發明之一實施例的單一字符辨識單元採用一種範 本匹配方法以辨識該單一字符。該單一字符辨識精度藉由 149139.doc • 14- 201201113 該範=匹配之距離而測定。該距離越小，精度越大。在該 單-子符辨識之樣本訓練中，採用機器學習演算法（例如 GLVQ)以產生特徵範本。該單一字符特徵向量包含「筆劃 方向刀佈特徵」、「網格筆劃（㈣strQke)特徵」及「周邊 :向特徵」。在特徵提取之前進行預處理，其包含諸如 、Λ '月」开/〜正常化」及「非線性正常化」的操作 以便調節該等樣本之特徵。在範本匹配中，—「多級串接 匹配」方法被採用以便逐級過遽出候選物以便改善匹配速 ^上述單一字符辨識方法被揭示於中國專利申請公開案 第CN 101354749Α號中，且，士由咬安Λ “ T J'此申清案中之所有内容被併入 本發明中以供參考。 在貫際書寫程序期間，不同使 刑祥+办 用者通*可以不同的書寫 1樣書寫相同的字符。舉例來 士、穴乂予母1 A」可具 有複數個書寫型樣，如圖7中所示。 一-曰文漢字「機」可具有三個書寫型樣，如圖8中所 不，其中後兩個書寫型樣為簡化字符。 因此，為改善該手寫辨識之穩固性，—種「多 1相=被採用於根據本發明之一實施例的裝置中以便對 本同子符之不同書寫型樣執行個別的訓練使得該「多範 本匹配」方法可被心職各種書⑼^ 竹該「多範本訓練」，所收隼&M4^ 為執 查哲 集的樣本首先根據其等之不n 書寫型樣分類。舉例來說，對於上述时 採用圖9A、9B及9C中所§頁亍的_彳、 」 發明 的二個樣本格式以m““ 训練期間形成該多範本訓練。 更在樣本 149l39.doc •15- 201201113 如圖4中所示，在步驟S 15中，該邏輯回歸模型之係數被 什算。實現手寫字符序列之識別的關鍵為正確分割該字符 序歹]本發明之一實施例的裝置及方法根據輸入字符序列 之各種特徵叶算在各種種類分割型樣中之輸入字符序列之 分別筆劃組合的分割可靠性。本實施例之一分割可靠性公 式採用邏輯回歸模型（LRM)，該模型為： nr) 1 + (1) 圖10中顯示該邏輯回歸模型之一函數曲線圖。當¥ 在-0°〜的一範圍内變化時，f(Y)之一值之範圍為從〇到 1，其意味著該分割可靠性之範圍為從0。/。到100。/^當γ=〇 時，f(Y)=0.5，其指示該分割可靠性為5〇%。 在上述邏輯回歸模型中， r = gm = 〜+M”2X2+...”mXm ……⑺。 ^(^人）為該邏輯回歸模型之一風險因數^當本實 把例;之裝置及方法計算該等分割可靠性時，心 ，表該筆劃組合之—m維特徵。（mu)代表該邏輯回 歸模型之一截距及回歸係數。 在計算該字符序列中之所有 另j此的筆劃組合之m維特徵 後，本實施例之裝置及 ^ 方法知用—種最大似然估算方法 (或諸如最小平方估算方法的 #哲八… ▲ ，、他參數估鼻方法）以便對於 忒專刀割可罪性估算該邏輯 mR β、 科杈型之截距β。及回歸係數 Φΐ，Ρ2，···，Ρηι)。 假設存在η個筆劃組合樣本且 料^ ^ 彳_ 觀 149139.doc -16 - 201201113 察值為Yi。N個回歸關係可被表達為·· K爲+爲‘Y! 1 +爲不，+ · · · +^ 峨X㈣X於，·；又、 η β〇 βχλ (ί1 + β2Χη2 +... +βη}Χ^ ......(3)。 在樣本訓練期間，對於兹 卞於第1個筆劃組合，如果該筆劃组人 可靠，令 fi=f(Y)=__L_ ° u 1+一—1，奶）>〇.5，即，Yi>0 ... (4); 如果該筆劃組合不可靠r l + e-Yi —〇 ’ f(Yi)<0.5，即，γγ〇 罪（即此筆劃組合型樣不正確），令 _(5) 中，則獲得f(Y) 將Y=g(x)=MPiX|+M2+兔代人該邏輯回歸模型公式 1 + e~Y ~Γ^:8(χ)' = π(χ) ......(6) 將 Ρ，Ρ(ί；=Ι|：^)設定為 fi=1 的 率為。因此 p(fi)=Pif,(i-Pi)(1—fi)。 一機率’則fi=o的一條件機 一個觀察值之一機率為 由於分別的觀察係獨立的，苴黧 J 兵寺之聯合分佈可被 表示·為分別之邊際分佈之—乘積其為 上述等式被稱為η個觀察之一似然函數。其目的為估算 最大化此函數值的參數。因&，此最大似然估算之關鍵為 估算最大化上述似然函數的最合適參數叭，p,，I，，k卜對 上述似然函數取對數，則獲得一對數似然函數。然後計算 該對數似然函數之一導數以獲得111+1個似然等式。最後， NeWton-Raphson方法可被應用以反覆計算這些爪+1個似然 149139.doc •17· 201201113 等式且因此該邏輯回歸模型中的係數（P〇A，P2，...，Pm)可被獲取 並可被保存於本實施例之裝置中以便用於辨識程序中。 根據本發明之另一實施例，輸入字符序列在分別之分割 型樣中的分割可靠性亦可利用一常態分佈模型計算。 圖Π係圖解根據本發明之一實施例的一手寫辨識程序之 一流程圖。如圖11中所示，在步驟S20中，該使用者輸入 手寫且該字符序列之筆劃被收集於該手寫輸入單元i ι〇 中。然後在步驟S2 1中’被收集之筆跡被保存於該手寫筆 跡儲存單元120中並在步驟S22中被該顯示控制單元i 5〇顯 示於使用者介面中。 然後’對於保存於該筆跡儲存單元中的該等筆劃，該字 符序列辨識單元130分別在步驟S23、S24、S25、S26、 S27、及S28中執行「預處理」、「筆劃組合特徵計算」、「單 一字符辨識」、「分割可靠性計算」、「分割最佳路徑選擇」 及「辨識後處理」的操作。 詳細而言，在步驟823、S24及S25中的執行程序相似於 藉由樣本訓練之上述邏輯回歸模型係數估算中的該等步 驟。在步驟S23中，執行一預處理以便根據該字符序列之 高度及寬度而估算該字符之平均高度字符之平均寬 度Wavg作為該筆劃組合之空間幾何特徵之一正常化準備使 付根據本發明之-實施例的手寫辨識裝置可被應用於具有 任意大小的字符序列。 在步驟S24中’對於該字符序列中所有可能的筆劃組合 計算該筆劃組合之各種特徵，包含該子筆劃組合之單—字 149139.doc 201201113 符辨識精度特徵及空間幾何特徵。 ，在步細中’彳叫該單一字符辨識單元以獲取該等合 併子筆劃之單-字符辨識精度及其他候選物精度 cmergeT以及兩個子筆劃之單—字符辨識精度Cs…及c心。 在步驟S26中，藉由利用該邏輯回歸模型之上述公式⑴ 及⑺’根據本實施例之方法基於該輪入字符序列之分別 特徵及在該樣本訓練中獲取的係數 (/UWWU而計算在各種分割型樣中之輸入字符序列之 为別筆劃組合的分割可靠性f(Y)。 在步驟S27中，根據本實施例之方法利用該N最佳方法 計算最可能的N個分割路徑。各個筆劃之一開始點被界定 為一元素節點且由該元素節點或-元素節點組合組成的一 路:為-對應筆劃組合。各個部份路徑之—成本函數為 COOUOO,換言之，該分割可純越高，該部份路徑之 成本函數之值越小。該N畏枯古、上、Α 路徑，其等使所有已通= :::= 選擇最佳的_ 小、㈡、、…第Ν小，的成本函數之值的和為最 抑最佳方法可藉由各種方式實施，舉例來說，多候選 物可藉由將動態規劃_方法及堆疊演算法組合而產生。、 在本實施例中，該崎佳方法包含兩個步驟：向前 向後搜尋。，向前搜尋採用一改良的維特比（viterM)演算法 (維特比决鼻法為—種搜尋最可能之隱含狀態序列 規劃方法）以便記錄傳送至各個元素節點之最佳N料 徑之狀態(即已通過路徑之成本函數值之-和)且第k個二 I49139.doc -19· 201201113 節點之狀態僅相對於第k+1個元 尋為一 個70素即點之狀態。該向後搜 哥為一種基於A*演算法的 ? A 隹豐肩鼻法。各個節點k之一試 铋函數為一「路徑成本函數 「 侗π如 °式探估鼻函數」之兩 ;、數之和，該「路徑成本函數 个山數」代表從該開始點到第k 個郎點之最短路徑的成本函數 数值之和，且「試探估算函 数」代表從第k個節點到目標節駄 _ 你即點之路徑成本之估算。在 该向後搜尋中’該堆疊中之_ : 吟奴刀數為一完全路徑分數 且最佳路經永遠位於該堆疊頂部。因&，此演算法為一全 域最佳演算法。 假設該使用者已輸入如圖6A中所示的一手寫字符序列 「define」，圖丨2A圖解根據本發明之一實施例的該手寫字 符序列之-分割結果。圖12八、咖及12(：中分別圖解藉由 N最佳方法之三個最可能的分割型樣。在第一分割型樣中 之各個字符之單一字符辨識結果的第一候選物為 「define(即正確答案）」’在第二分割型樣中的第一候選物 為「ccefine」且在第三分割型樣中的第一候選物為 厂 deftine」° 在步驟S28中，本實施例之方法最終執行後處理並藉由 與子典（英文單字字典）匹配或使用語言模型（例如雙字母組 模型）而對於該等辨識結果校正錯誤（例如該英文單字之拼 寫錯誤）。 在步驟S29中，該顯示控制單元1 50控制該顯示器螢幕以 向使用者呈現該等手寫辨識結果及該等相對候選物使得使 用者可在該候選物選擇單元140中選擇或確認經顯示的辨 149139.doc -20- 201201113 識結果（内定辨識結果為在第一分割型樣中之各個字符之 單一字符辨識的第一候選物）。該使用者可從該字符序列 之候選物分割型樣選擇正確的分割型樣或可從分別字符之 候選物選擇正確的辨識結果以便手動校正該字符序列中之 辨識結果之一部份，例如點選一單一字符或—片語以便從 其等之對應候選物選擇辨識結果。圖15係圖解根據本發明 之-實施例肖使用纟提供以便選擇&校正的點選單一字符 之候選物之一示意圖。 步驟S30偵測該使用者是否已確認或選擇某__候選物。 如果該使用者繼續書寫而未破認或選擇任何候選物，該程 序轉移到步驟S20並繼續上述辨識處理。如果其已偵測到 某―候選物被選擇’步驟S31從該等候選物中選擇該辨識 結果並顯示該辨識結果或提供給其他應用。同時，該手寫 輸入之辨識結果在步驟S32中被更新。 … 在計算該字符序列之分割可靠性期間，本實施例之方法 及裝置不僅考慮常用的㈣幾何特徵，亦考慮合併筆劃組 合之單-字符辨識精度以及子筆劃組合之單一字符辨識精 度’其結果係在正確分割難以被傳統技術執行的情況下， 例，不同字符中之筆劃在空間上部份重疊，或者一字符中 之筆劃間隙過大，其可達成正確分割及辨識結果。 此外’本㈣狀方法及裝置在執行字符序列分割時不 依賴於各個筆劃之輸入時間’因此其可適應於使用者的不 同輸入習慣。即使一使用者時快時慢地輸入字符，根據本 實施例之方法及裝置，分割精度將不會下降。 149l39.doc -21· 201201113 此外’採用於本實施例之方法及裝置中的筆劃組合之空 間幾何特徵係基於字符之估算平均寬度或高度的正常化特 徵，因此本實施例之裝置可適用於具有任何大小的一字符 序列。由於該多範本訓練及多範本匹配方法被採用於該單 一字符辨識中，藉由不同使用者之不同書寫型樣中的字符 (例如藉由中文的漢字之簡化字符）可藉由本實施例之方法 及裝置而被精確辨識。此外，本實施例之方法及裝置利用 語言模型及字典匹配使得該裝置具有拼寫檢查及單字 的功能。 最後，本實施例之方法及裝置的辨識對象可為英文單 字、日文假名組合、中文句？、韓文字符組合等。執行手 寫辨識的時機可任意減。該辨識結果可在該使用者輸入 該字符序列時連續更新，或者該等辨識結果可在該使用者 完成整個字符序列輸入後被顯示。 圖心邮、13(：及13D係圖解根據本發明之一實 的手寫辨識裝置之手寫辨識結果之示意圖。在辨識處理期 間，不僅考慮該筆劃組合之空間幾何特徵亦考慮單: 辨識精度，其結果伟太眚始 n 呆係本貫轭例之方法可在例如不同字符中 之筆劃在空間上部份重疊或字符之間之距離小於—字符中 =劃之間的距離或者字型大小在手寫輸人 統技術難以執㈣分割之情況 = 說’如_中所示，「d」.及、」之筆劃：「:」及她 tr及在「空：上部份重疊。如圖13咖中所示二 」 」之間的間隙小於「人h内的筆劃間距離且 149139.doc *22- 201201113 曰」及「本」之間的間隙小於「語」内的筆劃間距離。 如圖加及13D中所示，在「力m m、九」及「define」 中的字4之字型大小彼此不同β根據本發明之實施例的方 法可在上述情況中執行正確辨識。 圖14圖解一種根據本發明之一實施例的電子字典。如圖 14中所不’一系列英文手寫字符被辨識且該等辨識結果被 顓不輸入手寫之曰文翻譯藉由在一英曰字典中查找經辨 汽之奂文單子而向使用者呈現。如圖15中所示，當使用者 從該㈣結果點選某—單一字符時’將向使用者提供此單 一子符之候選物以便校正。 簡單而5本貫施例可允許使用者對於整個字符序列之辨 識、纟α果執行總體校正，且亦可允許使用者校正任何單一字 符辨識結果。 根據本發明之另一個實施例’該顯示區域及手寫輸入區 域可被組態於不同平面或相同平面上’如圖16八及16Β中 所示。舉例來說，膝上型電腦之手寫區域可被組態於該鍵 盤所在之平面上。 如上述，本發明之方法及裝置可被應用於或合併於任何 可採用手寫作為輸入或控制方式的終端產品，例如具有大 型觸控螢幕的個人電腦、膝上型電腦、pDA、電子字典、 MFP、行動電話、手寫裝置等等。 此描述及圖式僅圖解本發明之原理。應注意技術熟練者 可達成不同結構’儘管這些不同結構未被明顯描述及指 示，但這些結構體現本發明之原理且應#包含於本發明之 149139.doc •23· 201201113 精神及範圍内。在以上招述中，多個實例針對分別的步驟 2被描述。雖然發明人盡力說明相對實例，但這並不意味 著根據代表性符號這些實例應具有對應關係。只要在限於 斤l實例中之條件之間無矛盾，具有非對應代表性符號的 實例可構成一技術解決方案且此技術解決方·案應被視為在 本發明之範圍内。 應理解該申請專利範圍不限於以上圖解之精確組態及組 件。在本文描述之系統、方法及裝置之配置、操作及細節 中可做出各種修改、改變及變動而不脫離該申請 之範疇。 【圖式簡單說明】 圖1圖解-種基於斷開筆劃特徵的習知字符辨識方法； 圖2圖解先前技術中在基於該等斷開筆劃特 符時發生的問題； 飞子 圖3係圖解一種根據本發明之—實施例的手寫辨識裝置 之一結構示意圖； 圖顿圖解根據本發明之-實施例的手寫辨識裝置之一 樣本訓練處理之一流程圖； 圖5A係圖解在㈣本發明之—實施例之手寫辨識農 的筆劃組合及其等之子筆劃組合之一示意圖； 圖5B係圖解在根據本發明之—實施例之手寫辨識裳置 的筆劃組合及其等之子筆劃組合之一示意圖； 圖5C係圖解在根據本發明之一實施例之手寫辨識襄置 的筆劃組合及其等之子筆劃組合之一示意圖； 149139.doc -24- 201201113 圖5D係圖解在根據本發一 的筆-合及其等之子筆劃組合之= 寫辨識裝置中 圖6 A係說明在根據本發明之一實 沾仙以★ 例之手寫辨識裝置中 的筆劃組合之空間幾何特徵之一示意圖； 圖6B係說明在根據本發明一實 貫施例之手寫辨識裝置中 的筆劃組合之空間幾何特徵之一示意圖； 圖6C係說明在根據本發明之一實施例之手寫辨識裝置中 的筆劃組合之空間幾何特徵之一示意圖； 嶋說明在根據本發明之一實施例之手寫辨識裝置中 的筆劃組合之空間幾何特徵之一示意圖； 圖7係圖解根據本發明之一實施例的相同字符之不同書 寫型樣之一示意圖； 胃 圖8係圖解根據本發明之一實施例的相同字符之不同書 寫型樣之另一個示意圖； 圖9Α係圖解根據本發明之一實施例的多範本訓練及多範 本匹配之一示意圖； 圖9Β係圖解根據本發明之一實施例的多範本訓練及多範 本匹配之一示意圖； 圖9C係圖解根據本發明之一實施例的多範本訓練及多範 本匹配之一示意圖； 圖10係圖解根據本發明之一實施例的一邏輯回歸模型之 一函數曲線圖； 圖11係圖解根據本發明之一實施例的一手寫辨識程序之 一流程圖； 149139.doc -25- 201201113 圖12A係圖解根據本發明之一實施例的經由不同分割路 徑的分割之一示意圖； 圖12B係圖解根據本發明之一實施例的經由不同分割路 徑的分割之一示意圖； 圖12C係圖解根據本發明之一實施例的經由不同分割路 徑的分割之一示意圖； ° 圖13Α係圖解根據本發明之一實施例的手寫辨識裴置之 手寫辨識結果之一示意圖； 圖13Β係圖解根據本發明之一實施例的手寫辨識裝置之 手寫辨識結果之一示意圖； 圖13C係圖解根據本發明之一實施例的手寫辨識裝置之 手寫辨識結果之一示意圖； 圖13D係圖解根據本發明之一實施例的手寫辨識裝置之 手寫辨識結果之一示意圖； 圖14係圖解根據本發明之一實施例的手寫辨識方法在一 電子字典上的一應用之一示意圖； 圖15係圖解根據本發明之一實施例向使用者提供以用於 選擇及錯誤校正的辨識結果之至少一部份之候選物之一示 意圖； 圖16Α係圖解根據本發明之一實施例的手寫辨識方法在 一筆記型電腦上的應用之一示意圖；及 圖16Β係圖解根據本發明之一實施例的手寫辨識方法在 一行動電话上的應用之一示意圖。 【主要元件符號說明】 149139.doc •26· 201201113 110 120 130 131 132 133 140 150 手寫輸入單元 手寫筆跡儲存單元 字符序列辨識單元 單一字符辨識單元 分割單元 後處理單元 候選物選擇單元 顯示控制單元 149139.doc -27·degree. And a difference between (2; ^, and B = 1 means that one of the two strokes is merged into a single character: π (four) is a single-character possibility. The larger the difference, σ The more likely it is to form a single sub-character; (the single-character identification accuracy of the first candidate of the Cut sub-stroke), the single character recognition of other candidates of the merging sub-stroke: 1 = one, representing the single character recognition The second candidate 'and the value of Τ can be set): if the ratio is relatively large... taste: at the early one character recognition 'the matching distance between the combined stroke combination and the first candidate is quite close and The matching distance between H9J39.doc -13- 201201113 is very long, and the probability of merging into a single character is relatively large; (d) Sub-stroke A gap gap/w^ (or gap/Havg) between two bounding boxes: the smaller the gap between the substrokes, the greater the possibility of forming a single-character after merging. If the gap is a negative value , the possibility of forming a single character after merging is higher; (4) a merging sub-stroke width Wmerge/Wavg (or Wmerge/Havg): The smaller the merge width, the more likely it is to form a single character; (f) one of the end of the previous substroke and the beginning of the next substroke Vector Vs2.ei/Wavg (or Vs2.el/Havg); (g) one of the distance between the end of the previous substroke and the beginning of the next substroke, ds2_el/wavg (or ds2_ei/Havg); (h) one of the distances ds2_sl/Wavg (or ds2_s丨/Havg) between the start of the previous sub-stroke and the start of the next sub-stroke. In the above feature, "/" represents a division mark. And Wavg and represents the estimated character average width and the character average height during the pre-processing procedure. The spatial geometric features of (d)-(h) refer to Figures 6A to 6D and the points in the patterns represent one of the starting points of each stroke. For the above features (a), (b) and (c) 'the single-character identification accuracy Cmerge and other candidate precisions CmergeT of the combined sub-strokes and the single character recognition precisions Cstrl and Cstr2 of the two sub-pens are used by It is acquired by calling the single character recognition unit in step S14. A single character recognition unit according to an embodiment of the present invention adopts a template matching method to identify the single character. The single character recognition accuracy is determined by the distance of the matching value of 149139.doc • 14-201201113. The smaller the distance The greater the precision, the machine learning algorithm (such as GLVQ) is used to generate the feature template in the sample training of the single-child symbol identification. The single character feature vector includes "stroke direction knife feature" and "grid stroke ( (4) strQke) Features and Peripheral: Directional Features. Pre-processing is performed prior to feature extraction, which includes operations such as Λ 'Monthly Open/~Normalized' and 'Nonlinear Normalization' to adjust the characteristics of the samples. In the template matching, the "multi-level tandem matching" method is adopted to extract the candidate step by step in order to improve the matching speed. The above single character recognition method is disclosed in the Chinese Patent Application Publication No. CN 101354749, and士 Λ Λ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ “ Write the same characters. For example, a taxi, a hole, and a mother 1 A" may have a plurality of writing patterns, as shown in FIG. The 曰 汉 Chinese character "machine" can have three writing styles, as shown in Fig. 8, wherein the latter two writing patterns are simplified characters. Therefore, in order to improve the stability of the handwriting recognition, a "multiple phase = is employed in a device according to an embodiment of the present invention to perform individual training on different writing patterns of the same sub-character such that the "multi-modal" The "matching" method can be classified into a variety of books (9) ^ bamboo "multi-model training", and the samples collected by the &M4^ are first classified according to their writing styles. For example, the multi-sample training is formed during the training period for the two sample formats of the _彳, ” ” ” ” ” ” ” ” ” ”. Further in the sample 149l39.doc •15-201201113 As shown in Fig. 4, in step S15, the coefficients of the logistic regression model are calculated. The key to realize the recognition of the handwritten character sequence is to correctly segment the character sequence. The apparatus and method of an embodiment of the present invention calculate the respective stroke combination of the input character sequence in various types of segmentation patterns according to various characteristics of the input character sequence. Segmentation reliability. One of the segmentation reliability formulas of this embodiment adopts a logistic regression model (LRM), which is: nr) 1 + (1) A function curve diagram of the logistic regression model is shown in FIG. When ¥ is varied within a range of -0°~, one of f(Y) values ranges from 〇 to 1, which means that the segmentation reliability ranges from zero. /. To 100. /^ When γ = ,, f(Y) = 0.5, which indicates that the segmentation reliability is 5〇%. In the above logistic regression model, r = gm = 〜+M"2X2+..."mXm (7). ^(^人) is one of the risk factors of the logistic regression model. When the device and method calculate the reliability of the segmentation, the heart and the table are combined with the m-dimensional feature. (mu) represents one of the intercept and regression coefficients of the logical regression model. After calculating the m-dimensional features of all the stroke combinations in the character sequence, the apparatus and method of the present embodiment know the maximum likelihood estimation method (or #哲八... ▲ such as the least squares estimation method) , he parameter estimates the nasal method) in order to estimate the criminal mR β, the intercept of the scientific type of the interception β for the guilty. And the regression coefficient Φΐ,Ρ2,···,Ρηι). Suppose there are η stroke combination samples and the value is ^ ^ 彳 _ 149139.doc -16 - 201201113 The value is Yi. N regression relations can be expressed as ···K is + is 'Y! 1 + is not, + · · · +^ 峨X(4)X is, ·;; η β〇βχλ (ί1 + β2Χη2 +... +βη }Χ^ ......(3) During the training period, for the first stroke combination, if the stroke group is reliable, let fi=f(Y)=__L_ ° u 1+ one— 1, milk) > 〇.5, that is, Yi>0 (4); If the stroke combination is unreliable rl + e-Yi - 〇 ' f (Yi) < 0.5, that is, γγ〇 ( That is, the stroke combination type is incorrect. Let _(5) get f(Y). Y=g(x)=MPiX|+M2+ rabbit generation. The logistic regression model formula 1 + e~Y ~Γ ^:8(χ)' = π(χ) ......(6) Set Ρ,Ρ(ί;=Ι|:^) to the rate of fi=1. Therefore p(fi) = Pif, (i-Pi) (1 - fi). A probability of a conditional machine with a probability of 'fi=o' is one of the observed values. Since the separate observation systems are independent, the joint distribution of the 苴黧J bingsi can be expressed as the marginal distribution of the respective - the product is the above The equation is called n observation one likelihood function. Its purpose is to estimate the parameters that maximize the value of this function. Because &, the key to this maximum likelihood estimation is to estimate the most suitable parameter to maximize the above-mentioned likelihood function, p, I, and k. For the logarithm of the likelihood function, a pair-like likelihood function is obtained. A derivative of the log likelihood function is then calculated to obtain 111+1 likelihood equations. Finally, the NeWton-Raphson method can be applied to repeatedly calculate the coefficients of the claws +1 likelihood 149139.doc •17· 201201113 and thus the coefficients in the logistic regression model (P〇A, P2,..., Pm) It can be acquired and saved in the device of the embodiment for use in the identification process. According to another embodiment of the present invention, the segmentation reliability of the input character sequence in the respective segmentation patterns can also be calculated using a normal distribution model. BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 is a flow chart illustrating a handwriting recognition program in accordance with an embodiment of the present invention. As shown in Fig. 11, in step S20, the user inputs a handwriting and a stroke of the character sequence is collected in the handwriting input unit i 〇 . Then, the handwriting collected in step S2 1 is stored in the handwriting writing storage unit 120 and displayed in the user interface by the display control unit i 5 in step S22. Then, for the strokes stored in the handwriting storage unit, the character sequence identification unit 130 performs "preprocessing" and "stroke combination feature calculation" in steps S23, S24, S25, S26, S27, and S28, respectively. The operation of "single character recognition", "segment reliability calculation", "segment optimal path selection" and "identification post-processing". In detail, the execution procedures in steps 823, S24, and S25 are similar to those in the above-described logistic regression model coefficient estimation by sample training. In step S23, a pre-processing is performed to estimate the average width Wavg of the average height character of the character according to the height and width of the character sequence as one of the spatial geometric features of the stroke combination is normalized to be prepared according to the present invention - The handwriting recognition apparatus of the embodiment can be applied to a sequence of characters having an arbitrary size. In step S24, various features of the stroke combination are calculated for all possible stroke combinations in the character sequence, including the single-word 149139.doc 201201113 character identification accuracy feature and spatial geometric feature of the sub-stroke combination. In the step detail, the single character recognition unit is called to obtain the single-character recognition precision of the combined sub-strokes and other candidate precisions cmergeT and the single-character identification precision Cs... and c-heart of the two sub-strokes. In step S26, the above formulas (1) and (7)' using the logistic regression model are calculated based on the respective features of the rounded character sequence and the coefficients acquired in the sample training (/UWWU). The input character sequence in the split pattern is the split reliability f(Y) of the stroke combination. In step S27, the N best method is used to calculate the most likely N split paths according to the method of the present embodiment. One of the starting points is defined as an element node and consists of a combination of the element node or the - element node: a - corresponding stroke combination. The cost function of each partial path is COOUOO, in other words, the segmentation can be purely higher, The smaller the value of the cost function of the partial path, the N is obsessed with the ancient, the upper, the Α path, etc., so that all the passed = :::= select the best _ small, (two), ..., the third small, The sum of the values of the cost functions is the most optimal method can be implemented by various means. For example, multiple candidates can be generated by combining the dynamic programming method and the stacking algorithm. In this embodiment, Saki method contains Steps: Search backwards and forwards. Forward search uses a modified ViterM algorithm (Viterbi's method to search for the most likely implicit state sequence planning method) for record transfer to individual element nodes. The state of the best N-diameter (that is, the sum of the cost function values of the passed path) and the state of the kth two I49139.doc -19· 201201113 node is only 70% relative to the k+1th element. The state of the point. The backward search brother is a kind of A* algorithm based on the A 隹 肩 shoulder and shoulder method. One of the test functions of each node k is a “path cost function” 侗π如°式测鼻鼻” The sum of the numbers, the "path cost function mountain number" represents the sum of the cost function values of the shortest path from the starting point to the kth lang point, and the "exploratory estimation function" represents the kth node To the target thrift _ you estimate the path cost of the point. In this backward search, the _ in the stack: the number of slaves is a full path score and the best path is always at the top of the stack. Because & This algorithm is a global best algorithm. It is assumed that the user has input a sequence of handwritten characters "define" as shown in FIG. 6A, and FIG. 2A illustrates the result of segmentation of the sequence of handwritten characters according to an embodiment of the present invention. FIG. 12, coffee and 12 ( : The three most likely segmentation patterns by the N-optimal method are respectively illustrated. The first candidate for the single character recognition result of each character in the first segmentation pattern is "define" The first candidate in the second segmentation pattern is "ccefine" and the first candidate in the third segmentation pattern is factory deftine". In step S28, the method of the present embodiment finally performs post-processing and borrows Correcting errors for such identification results (eg, spelling errors in the English word) by matching with a sub-code (English word dictionary) or using a language model (eg, a two-letter model). In step S29, the display control unit 150 controls the display screen to present the handwritten recognition results and the relative candidates to the user so that the user can select or confirm the displayed identification in the candidate selection unit 140. 149139.doc -20- 201201113 The result of the recognition (the default identification result is the first candidate for single character recognition of each character in the first segmentation pattern). The user can select the correct segmentation pattern from the candidate segmentation pattern of the character sequence or can select the correct recognition result from the candidate of the character to manually correct a part of the recognition result in the character sequence, for example, A single character or phrase is selected to select the recognition result from its corresponding candidate. Figure 15 is a diagram illustrating one of the candidates for selecting a single character for selection & correction in accordance with an embodiment of the present invention. Step S30 detects whether the user has confirmed or selected a certain __candidate. If the user continues writing without unresolving or selecting any candidate, the process moves to step S20 and continues the above identification process. If it has detected that a "candidate is selected" step S31 selects the recognition result from the candidates and displays the identification result or provides it to other applications. At the same time, the recognition result of the handwriting input is updated in step S32. During the calculation of the segmentation reliability of the character sequence, the method and apparatus of the present embodiment consider not only the commonly used (four) geometric features, but also the single-character recognition accuracy of the combined stroke combination and the single character recognition precision of the sub-stroke combination. In the case where the correct segmentation is difficult to be performed by conventional techniques, for example, the strokes in different characters partially overlap in space, or the stroke interval in one character is too large, which can achieve correct segmentation and recognition results. Further, the 'fourth-order method and apparatus do not depend on the input time of each stroke when performing character sequence division' so that it can be adapted to different input habits of the user. Even if a user inputs characters quickly and slowly, the segmentation accuracy will not decrease according to the method and apparatus of the present embodiment. 149l39.doc -21· 201201113 Furthermore, the spatial geometric features of the stroke combination employed in the method and apparatus of the present embodiment are based on the normalized characteristics of the estimated average width or height of the characters, and thus the apparatus of the present embodiment can be adapted to have A sequence of characters of any size. Since the multi-template training and the multi-template matching method are used in the single character recognition, the characters in different writing patterns of different users (for example, simplified characters of Chinese characters by Chinese) can be obtained by the method of the embodiment. And the device is accurately identified. In addition, the method and apparatus of the present embodiment utilize a language model and dictionary matching to enable the device to have spell checking and word function. Finally, the identification method of the method and device of the embodiment can be an English word, a Japanese kana combination, a Chinese sentence? , Korean character combinations, etc. The timing of performing handwriting recognition can be arbitrarily reduced. The identification result may be continuously updated when the user inputs the character sequence, or the recognition result may be displayed after the user completes the entire character sequence input. Figure 13: and 13D are diagrams illustrating the handwriting recognition result of the handwriting recognition device according to one embodiment of the present invention. During the identification process, not only the spatial geometric features of the stroke combination but also the single: identification accuracy is considered. The result is that the method of staying in the yoke is such that, for example, the strokes of different characters are partially overlapped in space or the distance between characters is less than - the distance between characters or the size of the font is handwritten. The loss of the system is difficult to implement (4) the division of the situation = say 'as shown in _, "d". And, the stroke: ":" and her tr and in the "empty: the upper part of the overlap. As shown in Figure 13 The gap between the two shown is less than the distance between the strokes in the person h and the gap between 149139.doc *22-201201113 曰 and "this" is less than the distance between the strokes in the "language". As shown in Fig. 13, the font sizes of the words 4 in "force mm, nine" and "define" are different from each other. The method according to an embodiment of the present invention can perform correct recognition in the above case. Fig. 14 illustrates a An electronic dictionary of an embodiment of the invention. In the 14th, a series of English handwritten characters are recognized and the recognition results are not input into the handwritten text. The translation is presented to the user by searching the list of the recognized steam in a dictionary. As shown in Figure 15, when the user selects a certain single character from the (4) result, the user will be provided with the candidate for the single sub-character for correction. Simple and 5 embodiments can allow the user to use the entire character sequence. The identification, 纟α fruit performs overall correction, and may also allow the user to correct any single character recognition result. According to another embodiment of the present invention, the display area and the handwriting input area may be configured on different planes or the same plane. 'As shown in Figures 16 and 16'. For example, the handwritten area of the laptop can be configured on the plane of the keyboard. As described above, the method and apparatus of the present invention can be applied or incorporated in Any end product that can be handwritten as an input or control method, such as a personal computer with a large touch screen, laptop, pDA, electronic dictionary, MFP, mobile phone, hand The present description and the drawings are merely illustrative of the principles of the invention. It is to be understood that those skilled in the art can In the above description, a plurality of examples are described for the respective step 2. Although the inventors tried to explain the relative examples, this does not mean that the representative symbols are based on the representative symbols. These examples should have a corresponding relationship. As long as there is no contradiction between the conditions in the example, the examples with non-corresponding representative symbols may constitute a technical solution and this technical solution should be considered as being within the scope of the present invention. It should be understood that the scope of the patent application is not limited to the precise configuration and components illustrated in the above. Various modifications, changes and variations can be made in the configuration, operation and details of the system, method and apparatus described herein without departing from the scope of the application. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 illustrates a conventional character recognition method based on a broken stroke feature; FIG. 2 illustrates a problem occurring in the prior art based on such disconnected stroke symbols; A schematic diagram of one of the handwriting recognition devices according to the embodiment of the present invention; Figure 1 is a flow chart of one of the sample training processes of the handwriting recognition device according to the embodiment of the present invention; FIG. 5A is a diagram of (4) the present invention. FIG. 5B is a schematic diagram showing one of a stroke combination of a handwriting recognition skirt and its sub-stroke combination according to an embodiment of the present invention; FIG. 5C is a schematic diagram showing a stroke combination of a handwriting recognition device and a sub-stroke combination thereof according to an embodiment of the present invention; 149139.doc -24- 201201113 FIG. 5D is a diagram of a pen-and combination according to the present invention. FIG. 6A shows the space combination of the stroke combination in the handwriting recognition device according to one of the embodiments of the present invention. FIG. 6B is a schematic diagram showing spatial geometric features of a stroke combination in a handwriting recognition apparatus according to a practical embodiment of the present invention; FIG. 6C is a diagram illustrating a handwriting recognition apparatus according to an embodiment of the present invention. Schematic diagram of one of the spatial geometric features of the stroke combination; a schematic diagram illustrating one of the spatial geometric features of the stroke combination in the handwriting recognition apparatus according to an embodiment of the present invention; FIG. 7 is a diagram illustrating an embodiment of the present invention. A schematic diagram of one of the different writing patterns of the same character; FIG. 8 is another schematic diagram illustrating different writing patterns of the same character according to an embodiment of the present invention; FIG. 9 is a diagram showing a multi-template according to an embodiment of the present invention. Schematic diagram of one of training and multi-model matching; FIG. 9 is a schematic diagram showing one of multi-model training and multi-template matching according to an embodiment of the present invention; FIG. 9C is a diagram illustrating multi-model training and multi-modality according to an embodiment of the present invention. Matching one of the schematic diagrams; FIG. 10 is a diagram showing a function curve of a logistic regression model in accordance with an embodiment of the present invention Figure 11 is a flow chart illustrating a handwriting recognition program in accordance with an embodiment of the present invention; 149139.doc -25-201201113 Figure 12A illustrates one of segmentation via different segmentation paths in accordance with an embodiment of the present invention. Figure 12B is a schematic diagram showing one of segmentation via different segmentation paths in accordance with an embodiment of the present invention; Figure 12C is a schematic diagram showing segmentation via different segmentation paths in accordance with an embodiment of the present invention; 1 is a schematic diagram showing a handwriting recognition result of a handwriting recognition device according to an embodiment of the present invention; FIG. 13 is a schematic diagram showing a handwriting recognition result of a handwriting recognition device according to an embodiment of the present invention; 1 is a schematic diagram of a handwriting recognition result of a handwriting recognition device according to an embodiment of the present invention; FIG. 13D is a schematic diagram showing a handwriting recognition result of a handwriting recognition device according to an embodiment of the present invention; FIG. 14 is a diagram illustrating implementation according to one embodiment of the present invention. A schematic diagram of an example of a handwriting recognition method in an electronic dictionary; FIG. 15 is a diagram illustrating One embodiment of the present invention provides a schematic diagram of a candidate for at least a portion of the identification result for selection and error correction; FIG. 16 is a diagram illustrating a handwriting recognition method in a notebook computer according to an embodiment of the present invention. A schematic diagram of one of the applications; and FIG. 16 is a schematic diagram showing one of the applications of the handwriting recognition method on a mobile phone according to an embodiment of the present invention. [Main component symbol description] 149139.doc •26· 201201113 110 120 130 131 132 133 140 150 handwriting input unit handwriting handwriting storage unit character sequence identification unit single character recognition unit division unit post-processing unit candidate selection unit display control unit 149139. Doc -27·