Research On Mouse Fear Recognition And Analysis System Based On Video Deep Learning

In mouse behavior studies,researchers generally have two statistical methods to identify mouse's fear: One uses the traditional inter-frame difference method to count the percentage of gray-value changes in pixels of adjacent frames,but it is not highly correlated with the results from manual statistics;The other one uses deep learning to obtain the coordinates of key points of the mouse body in each frame without manual labeling,and then determine whether the mouse is in a fearful behavior according to the position variation of key points.In order to improve the credibility and accuracy of mouse fear emotion recognition,in this thesis,we optimized the U-Net network model and carried out a lightweight design,and proposed a method to quantify the mouse's fear emotion behavior according to the variation of the mouse contour area in each frame.The major contents of this thesis are as follows:1.A variety of strategies were introduced to improve the segmentation effect of the U-Net model.The SENet channel attention mechanism module and the deep residual shrinkage network module were embedded in the convolutional layer of the UNet networks for the comparison of their segmentation performance.The latter one shown increased indicators by 2?3% and better segmentation effect based on our mouse data set.Among them,the indicator of Intersection over Union(Io U)reached 0.903.The dilated convolution was also introduced to increase the receptive field and the concentrated loss function to solve the problem caused by the imbalance of positive and negative samples.2.Lightweight design based on the improved U-Net model was proposed.Standard convolution in the U-Net network was split into deep separable convolution,grouped convolution,and asymmetric convolution.After the final training,the file size of a single model was reduced from 124.7MB to 27.9MB,and it can process 7-8pictures per second.This thesis,for the first time,also proposed a mouse fear emotion behavior determination algorithm based on adjacent frame recall rate and ergodic thinking.We set the adjacent frame recall rate threshold T1 and the frame number threshold T2 initially.If the number of frames in the continuous interval was greater than T2 and meanwhile the recall rate was greater than T1,it can be considered that the mouse was in a state of fear and rigid emotional behavior during this time period.3.A mouse fear emotion behavior recognition and analysis system were constructed.On the one hand,the system displayed the original video and the real-time segmentation effect video.On the other,it drew the mouse's movement curve in real time and could mark the time period in the curve when the mouse was in fear.By selecting time period parameters,it could calculate the percentage of mouse in fear at different time periods.In the data analysis and verification,50 mouse videos were employed,and the correlation analysis between the system results and the statistical results from experts was carried out,and the person correlation coefficient could reach over 86%.It demonstrated that the presented system in this thesis could complete the analysis of mouse fear emotions,and could improve the data analysis efficiency of mouse behavior studies.Meanwhile,it offers potential applications for behavior analysis of other model animals.