請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/78288
標題: | 長短期記憶網路實現面部光體積描記法特徵提取之研發 Research and Development of Long Short-term Memory Network for Feature Extraction of Facial Photoplethysmogram |
作者: | Yu-Jie Hsu 徐郁捷 |
指導教授: | 李世光(Chih-Kung Lee) |
關鍵字: | 長短期記憶網路,光體積描記法,血氧分佈, Long Short-term Memory Network,imaging-Photoplethysmogray,Oxygen saturation distribution, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 本研究使用了長短期記憶網路來實現非接觸式光體積描記法(imaging- Photoplethysmographic, iPPG)的特徵提取,同時得到血氧飽和度分佈圖,其應用可 提供心血管疾病和動脈硬化程度、血管分佈等相關訊息。過去大多數傳統的 PPG 量測系統都是藉由接觸手指頭表面來進行量測,但對於末梢血液循環不好的長者 或是手指缺陷的患者,無法進行準確的量測,居家照護方面也無法進行長時間監控, 此外目前使用非接觸式量測的研究對於量測到的血氧波形並不如接觸式量測提供 的資訊多,原因是臉部波形本身的特徵點較不明顯,及非接觸式量測到的波形有時 具有缺陷,因此在特徵點擷取上有一定的困難。 本實驗開發了新的架構及演算法量測非接觸式光體積描記法,實驗使用兩種 特定波段的主動光源,配合選定波長的濾片來改善訊號雜訊不穩定狀況,並使用高 幀率來描繪更細緻的波形。架構方面以同軸光架構進行量測,獲得區域性的血氧分 佈,突破了過去的單點量測;演算法方面除了傳統影像分析,也加入了類神經網路 的 LSTM 架構模型。選擇臉部當作量側目標,從解剖學觀點尋找臉部接近表皮的 血管位置當作感興趣區域(Region of interest, ROI),同時使用商用血氧儀量測生理 訊號作為基參考訊號,最後訓練出以耳朵為參考訊號的 LSTM-E 模型,以及手指 為參考訊號的 LSTM-F 模型。透過兩種訓練模型分別得到臉部 iPPG 訊號及特徵提 取 iPPG 訊號,臉部 iPPG 訊號在上升時間、心律及特徵點的時間準確度都與參考 訊號有高度相關且穩定;特徵提取 iPPG 訊號更能從看出完整特徵點,並且保留了 受試者本身的波形特徵,代表建構此模型的方法在特定使用情境及前提下均合理。 結果顯示,使用類神經網路來處理傳統影像分析後的 iPPG 波形,有利於未來 研究特徵點擷取,進一步得到生理訊號指標。若為來有機會增加訓練資料數量及多 樣性,有機會提供更普遍的模型適用度。由兩道不同波長得到的血氧分佈,將其量 測結果與臉部血管分佈圖做對照,發現有很大的相關性,未來可以應用在監控斷肢 重新接合的復原等狀況。在探討血氧分佈受環境光影響行後,證實本架構可在環境 光下操作。 This thesis developed a new design to realize the feature extraction of imaging- Photoplethysmogray (iPPG) with the using of Long Short-term Memory Network (LSTM). This study also uses this design to measure the oxygen saturation distribution map of the objects. The applications of the distribution map such as: diagnostic information about cardiovascular condition, arteriosclerosis degree, and blood vessel distribution are available. In the past, most traditional PPG measurement systems were measured by the contact of the finger, but it was inaccessible for elderly people with poor peripheral blood circulation nor the patients with finger defects. It was a challenge for the accurate measurements and home care. In addition, previous non-contact measurements provided less information than the contact measurements, because the characteristic points of the facial waveform were not obvious. Therefore, the feature point extraction was hard to measure. We developed a novel device and algorithm to measure the iPPG signal. There are two selected active light sources with narrow-band filters in the device to improve the signal-to-noise ratio, and it uses high frame rate to get more detailed from waveforms. The coaxial optical architecture is used in the device settings to obtain regional blood oxygen distribution, and it is an improvement in comparison of the previous single-point measurement. We also used a neural network, LSTM model, in the image analysis. We used the device to measure face area, finding the position of the blood vessel as the region of interest (ROI). A commercial wireless physiological signal measurement device (BioRadio) was used simultaneously during the experiment to provide the ground truth PPG signals. The LSTM-E model was trained from the ear reference signal, and the LSTM-F model was trained from the finger reference signal. The facial iPPG signal and the feature extraction iPPG signal are observed from the LSTM-E and the LSTM-F, respectively. The facial iPPG signal's crest time, heart rhythm, and time accuracy of the feature points are highly correlated and stable with the reference signal. Feature extracted iPPG signal provides obvious feature points and the waveform of the object. It shows that the device is reasonable under the specific situation. The results show that the use of LSTM to process the iPPG waveform is conducive to future research feature point extraction and a further step to obtain physiological signal indicators. If we could increase the number of data base, we could provide more general and suitable model for every situation. In the oxygen saturation distribution map which was obtained from the two different wavelengths, we found that it has a high correlation between the blood vessel distribution map of the face. In the future, it can be used to monitor the recovery of amputated limbs. Result also shows that the blood oxygen distribution is not affected by ambient light, confirming that this device is suitable for operation under ambient light. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/78288 |
DOI: | 10.6342/NTU202002732 |
全文授權: | 有償授權 |
顯示於系所單位: | 工程科學及海洋工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-0908202021411600.pdf 目前未授權公開取用 | 23.69 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。