Research On Neonatal Face Detection And Tracking Based On Deep Learning

Automatic neonatal pain evaluation system is helpful for clinicians to evaluate neonatal pain effectively.The neonatal face detection and tracking is the premise and basis of the automatic pain evaluation system,which is more conducive to filter out interference information and improve the recognition accuracy.However,the facial features of the neonates are quite different from those of adults,most of face detection algorithms designed for adult samples aren't suitable for neonatal face detection.Therefore,the facial features that suitable for neonates are learned adaptively by using deep learning,and the neonatal face detection and tracking is realized.The work is organized as follows:(1)A neonatal image dataset is developed.The dataset is the basis of the research on neonatal face detection and tracking.The dataset containing multi-view,multi-background and multi-expression is established by preprocessing the collected original neonatal videos.(2)Neonatal face detection based on improved multi-scale features network is proposed.According to neonatal facial features,an improved multi-scale features network is proposed,which improves the speed or accuracy of neonatal face detection by using pyramidal hierarchical features.The experimental results show that the ZFNet structure as the extracted feature layer of multi-scale features network improves the detection speed,and the detection accuracy is obtained to 98.2% by selecting the appropriate feature maps according to the basic network.(3)Neonatal face detection based on YOLOv2 is proposed.The best default box size of the neonatal image is obtained by K-means clustering,which improves the detection performance effectively.The experimental results show that YOLOv2 has the best detection performance,compared with multi-scale features network and MTCNN,whose detection accuracy is up to 99.5%.More than those,the detection performance of the neonatal side face,occlusion and expression changes is significantly improved.(4)Research on neonatal face tracking is conducted.The basic principles of KCF and SiamFC are studied and applied to the neonatal face tracking.The experimental results show that SiamFC has better robustness,and the tracking performance of neonatal videos with occlusion,illumination and scale changes is better than that of KCF,while keeps real-time performance.