Application Of Deep Learning In Facial Recognition And Motor Posture Analysis Of Non-Human Primates

It is important to conduct studies on monkey facial recognition and expression analysis in order to facilitate the identification of monkeys and the emotions expressed by their faces by experimenters.The traditional observation recording method relies on the subjective judgment and observation of the researcher,and the large amount of work makes the researcher prone to fatigue and distraction,which reduces the reliability of the study.Face recognition based on deep learning has paved the way for the implementation of facial recognition and expression analysis in monkeys,and the development of human posture analysis has advanced the progress of motion analysis in experimental animals.At present,many commercial software can perform some motion analysis,but most of them cannot meet the current experimental needs,and the 2D behavioral information covered in some recorded videos is not accurate for monitoring data of behavior.The first part of this study is about monkey face recognition and expression recognition.In this study,based on the face recognition,YOLO model algorithm is introduced and the training YOLOv5 artificial neural network method is used to realize the facial information processing of non-human primates.Five rhesus monkeys were subjected to facial analysis in the first stage,and the data samples were increased because the amount of data was insufficient to construct a monkey model.The facial information data of 14 rhesus monkeys were collected and more than 6000 images were collected.Due to the long training time and memory limitation,the total number of model training datasets was reduced,and the average recognition accuracy reached 95% by training the labeled datasets using 1899 images after filtering.Using YOLO migration learning to train 10 rhesus images to form monkey weights,it was found that using monkey weights can accurately identify individual monkeys with faster detection speed and better robustness.For expression recognition,the rhesus monkeys were experimentally stimulated to produce different expression changes to obtain video data.After data preprocessing,the video data are processed to classify the different expressions of monkeys.The data set was trained to generate the corresponding YOLO monkey weights,and because the manual classification of expressions was imprecise and the effective data set was small,the data set needed to be expanded and classified accurately,and the facial information was detected using the GUI program interface built by PYQt5,and the accuracy of expression recognition reached 80%.For motion trajectory analysis,the motion trajectory of monkeys in the research cage is recorded using surveillance cameras,trained with the resnet＿50 network model,marked the required limb points,used fixed length as scale and corresponding proportional relationship to reduce the motion trajectory distance error,and smoothed some coordinate points with low confidence.The monkey pose analysis is similar to the trajectory analysis,but the 3D pose analysis needs to correct the camera video distortion and reduce the pixel error in order to reshape the arm motion pose of the rhesus monkey.The analysis of rhesus monkey motion trajectory can obtain parameters such as moving distance,speed,and residence time in a certain area.The 3D pose reconstruction of rhesus monkeys can obtain the coordinates of the marked pose and clearly delineate the detailed motion forms of the limbs.The study realizes the corresponding analysis of 2D trajectory and posture of rhesus monkeys and 3D trajectory and posture reproduction,which can observe and analyze the movement state of monkeys and help researchers to understand the movement pattern and spatial behavior aspects of monkeys.