Contactless blood oxygen estimation from face videos: A multi-model fusion method based on deep learning

Contactless Blood Oxygen Saturation Estimation from Facial Videos Using Deep Learning

“…A and B are experimentally evaluated coefficients that are determined by identifying the line of best fit between the ratios of the red and blue channels and the SpO 2 estimated by a ground truth device. Following Equation ( 2 ), remote SpO 2 measurement with an RGB camera was further validated in [ 21 , 22 , 23 , 48 , 50 ]. However, only two methods used deep learning and were tested on a public benchmark dataset [ 48 , 49 ].…”

“…Following Equation ( 2 ), remote SpO 2 measurement with an RGB camera was further validated in [ 21 , 22 , 23 , 48 , 50 ]. However, only two methods used deep learning and were tested on a public benchmark dataset [ 48 , 49 ].…”

“…Deep learning techniques have achieved state-of-the-art performance for the remote measurement of physiological signs such as HR [ 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 ] and RR [ 36 , 39 , 40 , 41 , 42 , 43 , 44 , 45 , 46 , 47 ]. However, remote SpO 2 measurement is still in its infancy, with only a few papers using convolutional neural networks (CNNs) to predict SpO 2 from RGB facial videos [ 48 , 49 , 50 ]. Additionally, most existing methods are evaluated on private self-collected datasets, preventing a fair comparison of algorithmic performance [ 51 ].…”

Oxygen saturation estimation from near-infrared multispectral video data using 3D convolutional residual networks

“…To overcome these limitations, recently many studies have demonstrated that various vital signs can be measured using cameras, [9][10][11] among which a growing number of studies have investigated camera-based oxygen measurement techniques. [12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27] Bui et al 12 and Ding et al 13 require participants cover a finger on the smartphone built-in camera and flash lamp to obtain the SpO 2 from captured reemitted light, which can be strictly assumed as a contact-weak rather than contactless methods. Main stream studies on contactless measurements usually use RGB cameras to capture ROI on hands, 14,15 forearms 16 or faces [17][18][19] with ambient light, and extract weak physiologically pulsatile information through different analytical filtering techniques [20][21][22] or neural networks 23,24 based on remote PPG (rPPG) signals to calculate SpO 2 .…”

“…The spatial information of the captured skin regions has been proved by Wieringa et al 25 and Rosa et al 26 to contain blood oxygenation information, which implies that remote SpO 2 measurement relies not only on rPPG signals but also on other spatial encoded patterns. Hu et al 27 proposed a study based on 2D-residual cascade and coordinate attention mechanism to focus on the correlation between feature channels and the spatial information of feature space. They extracted the spatial color information from images using a neural network, and then the extracted spatial features of multiple frames were concatenated for estimation.…”