如今電子支付方式已經開始替代傳統的現金交易模式。但依據調查發現，每年的支票交易金額依舊龐大，且仍有大量支票用於交易。查閲相關支票掃描器發現，現有的支票掃描硬體價格昂貴，且搭配的軟體使用局限性大，并不能辨識手寫區域，且連續掃描性能差。在新光銀行作業中心中，僅僅票據金額正確性檢查每日就需約20位同仁進行比對，無疑耗費大量人力。在明確問題和市場需求後，我們決定開發出一套支票金額識別系統。通過深度學習對支票圖像進行掃描，對其中的手寫金額與數字先進行準確定位，以確認具體的識別區域。然後再對該區域進行字符辨識，並將辨識結果自動轉換為文本，交叉比對，並確認金額是否相同，以提供給銀行專員使用。該系統可以對支票進行有效的管理，減少銀行專員的支票對比時間及錯誤率，以防止造成更大的損失。目前本支票金額識別系統可以很好的解決上述問題，並且功能强大。識別模型針對數字和漢字的單字準確率（Precision）、召回率（Recall）和F-score上的正確率均已達到八成以上。我們與主流的OCR模型（Paddle-OCR、Google-OCR和Tesseract-OCR）進行比較發現，針對支票的手寫金額識別能力已經高於這些既有的OCR模型。

Nowadays, electronic payment has begun to replace the traditional cash transaction mode. However, according to the survey, the annual check transaction amount is still huge, and a large number of checks are still used for transactions. Referring to the relevant check scanners, it is found that the existing check scanning hardware is ex-pensive, and the matched software has great limitations. It cannot recognize the handwritten area, and the continuous scanning performance is poor. In the operation center of Shin Kong Bank, only checking the correctness of the bill amount requires about 20 colleagues to compare every day, which undoubtedly consumes a lot of manpower. After clarifying the problems and market demand, we decided to develop a set of check amount recognition system. Scan the check image through the function of deep learning, and accurately locate the handwritten amount and number in order to confirm the specific recognition area. Then, character recognition is carried out for this area, and the recognition results are automatically converted into text, cross com-pared, and whether the amount is the same is confirmed for use by the banking commissioner. The system can effectively manage the check, reduce the check comparison time and error rate of the banking specialist, and prevent greater losses. At present, this check amount recognition system can well solve the above problems and has powerful functions. The recognition model has achieved more than 80% performance in single word accuracy, recall and F-score for numbers and Chinese characters. Com-pared with the mainstream OCR models (paddle-OCR, Google-OCR and Tesseract-OCR), we found that the handwritten amount recognition ability for cheques has been higher than these existing OCR models.