- 學位論文
GAN-based SSD Segmentation Algorithm 輔助七段顯示數字之字元辨識
GAN-based SSD Segmentation Algorithm to Assist the Character Recognition of Seven-Segment Display Digits
摘要
根據研究顯示患有慢性疾病的人口逐年增加，因此使用血壓計、血糖計及血脂機等醫療器材測量相關生命表徵的需求也越來越高，然而傳統的手動紀錄方式可能造成謄寫錯誤，且目前市面上具有儲存或傳輸功能的器材，價格大多過於昂貴，此外，這些器材大多使用七段顯示數字作為呈現數據的方式，然後這樣的數字中有不連續的狀況發生，與印刷體的數字有較大的差異，容易造成光學字元辨識(Optical Character Recognition, OCR)辨識效果不佳，因此本論文提出利用GAN標記七段顯示數字不連續的區域，並利用GAN-based SSD(Seven Segment Digits) Segmentation Algorithm自動調整影像的方向，接著透過著名的OCR辨識工具辨識標記及校正後的七段顯示數字影像，以此自動紀錄病患的收縮壓、舒張壓及脈搏。在實驗結果的部分，我們計算GAN標記影像之混淆矩陣的相關指標，其中測試階段的準確率為94.5%、精準度為98.4%、敏感度為90.9%以及特異度為98.4%，而辨識數字的準確率則有99%，由上述數據可知，本論文提出之方法可有效提升七段顯示數字的辨識成功率。
並列摘要
According to the research, the population of chronic diseases is increasing year by year. Therefore, the demand for measuring vital signs using sphygmomanometers, blood glucose meters and lipid profile machines is also increasing. However, the traditional manual recording method may cause transcription errors and most of the devices on the market that have storage or transmission functions are too expensive. In addition, these medical devices use seven-segment display digits to display the measurement results. The seven-segment display digit contains the discontinuous field and quite different from the printed numbers that may easily cause poor Optical Character Recognition(OCR). Therefore, we propose using GAN to label the discontinuous fields of seven-segment display digits area and using the GAN-based SSD(Seven Segment Digits) Segmentation algorithm to automatically adjust the direction of the images. Then, use a famous OCR tool to recognize the digit value in the image of seven-segment display digits and automatically record the value to complete the monitoring of the systolic blood pressure, diastolic blood pressure and pulse. The experiment result shows that the relevant performance indicators for evaluating trained/tested model are used. Finally, the accuracy of the test stage is 94.5%, the precision is 98.4%, the sensitivity is 90.9%, and the specificity is 98.4%. The recognition accuracy rate is 99%. According to the above data, the method proposed in this paper can effectively improve the success rate of the seven-segment display digits recognition.