Nose Region Detection for Measurement of Non-Contact Respiration Rate Using Convolutional Neural Network
*Corresponding author: Eui Chul Lee, Department of Intelligent Engineering Informatics for Human, Sangmyung University, Hongjimun 2-Gil 20, Jongno-Gu, Seoul, 03016, Republic of Korea
Vital signs are important clinical measurements that indicate the state of a person's body. Although this technique has been used only in the medical field in the past, owing to the advances in technology, it is now being widely used for collecting information for various purposes by measuring the bio-signals of people.
In this study, we propose a nose region detection method for measuring the respiration rate using thermal imaging. We use the characteristic that the temperature change in the nose region is visible in thermal imaging during breathing. Additionally, we confirmed the accuracy of nose region detection using CNN (Convolutional Neural Network) for continuous and automated respiration rate measurements. Faster-RCNN (Region with CNN features) was used as the CNN model.
In thermal images, we could find a clear temperature change between inspiration and expiration. However, we also found that the nose region in the thermal images did not show distinct features. This means that feature loss problems will occur in the training process. To solve this problem, we used Res Net to prevent feature loss problems in the training and prediction processes. Then, using the fact that the prediction result is output to a rectangular box, the mouth region is detected together with the nose region, and then the final nose region is determined after considering the geometric relationship between the two regions. We used three relationship; 1) distance between the nose and mouth, 2) horizontal length of the nose is shorter than the horizontal length of the mouth, and 3) the centers of the nose and mouth located on a vertical line.
As a result, we confirmed the satisfactory performance of nose region detection in thermal imaging using CNN. In addition, detection accuracy is enhanced through additional verification of the prediction result.
Vitalsign, Respiration rate, Convolutional neural network, Non-contact, Geometric relationship.