| Heart Rate(HR)is an important physiological signal reflecting a person’s physical and mental state,which has great guiding significance for people’s healthy life and the prevention and treatment of cardiovascular diseases.Traditional contact HR measurement technology requires the testee to contact with the measuring instrument,which may bring discomfort to the testee and can’t cope with some special situations such as skin burns.With the development of science and technology,it has been found that the changes of blood volume and oxygen saturation in blood vessels caused by heartbeat can lead to slight color changes of skin,which can be captured by RGB sensors and used for HR measurement.Based on this,more and more accurate and efficient non-contact heart rate measurement methods have emerged one after another.These methods have become a research hotspot in recent years because of their advantages of comfort,convenience,simple operation,high efficiency and flexibility.The method based on Deep Learning(DL)is to use the powerful modeling ability of Deep Neural Network(DNN)to model the relationship between heart rate measurements.However,with the increase of network depth,the training becomes more and more difficult,and the Residual Network(Res Net)solves the training difficulty caused by network depth.Based on Res Net,this paper improves and mainly completes the following work.(1)Integrate the Long Short Term Memory(LSTM)network module.LSTM can learn relevant and dependent information for a longer time and the design of the network is also simpler.In terms of information processing,it can also save useful information that needs long-term memory while discarding irrelevant and useless information,so it is especially suitable for processing longer.Sequence,so incorporating the LSTM module can make the network have a better performance when dealing with long video sequences.(2)Add Convolutional Block Attention Module(CBAM).Through CBAM,more weights can be assigned to the relevant features with rich heart rate information in the extracted face features,so that the weight distribution is reasonable,the noise caused by useless information is avoided as much as possible,and the prediction accuracy is improved.At the same time,CBAM is also a lightweight module,which can be added to the network with little extra cost,with good stability and applicability across multiple databases and tasks.(3)Using Ada Mod optimization algorithm.This paper selects the Ada Mod optimization algorithm in the back-propagation stage of training,which sets a dynamic upper bound for the adaptive learning rate,prevents the calculated learning rate from rising too fast,controls the variance of the adaptive learning rate,and eliminates the adaptive learning rate Unexpected fluctuations in rates.The algorithm can effectively handle the extreme learning rates that occur during training,thereby solving the non-convergence problem.Based on the above work,a new prediction network Face-HRNet for heart rate estimation was constructed.Face-HRNet was used to estimate heart rate based on face video and achieved good results.The final experiment proves that this method is superior to other methods in the accuracy of heart rate prediction and the result is more intuitive for the output is the input video with the predicted heart rate value. |
