A Multi-task-Learning Study Of Facial Expression Recognition And Parkinson’s Disease Diagnosis Based On Deep Learning

It is generally acknowledged that convolutional neural network has been widely used in deep learning studies.As a high research hotspot,computer vision includes image recognition,object detection,semantic segmentation and other modules,the research of these modules has also made great progress.at the same time,the application of artificial intelligence technology in the medical field is also booming.In order to apply these well-known deep learning models and show the powerful role of artificial intelligence in medical field,this paper proposes a facial expression recognition model based on convolution neural network and A multi-task learning model for Parkinson’s disease diagnosis based on convolution neural network and facial expression recognition technology.In this way,it can not only get the results of facial expression recognition,but also judge whether they have Parkinson’s disease,which provides a reference for clinical diagnosis of Parkinson’s disease.This paper mainly studies the field of image recognition,and its contents are as follows.(1)Facial expression recognition is to get the current personal emotional state through the given image and video data in order to achieve the purpose of humancomputer interaction.In this paper,six types of facial expressions,including anger,disgust,fear,happy,sad and surprise,were selected for experimental study.Through cooperation with local hospitals,the expression dataset of Parkinson’s disease patients(called PDface)was collected,while the normal people used the open expression dataset Oulu CASIA.(2)Parkinson’s disease is a neurodegenerative disease.The main cause of Parkinson’s disease is the degeneration and death of dopaminergic neurons in substantia nigra,which leads to the significant decrease of DA content in striatum.The data show that the prevalence rate of the elderly over 65 years old is 1.7%.Therefore,it is also of great significance for the early screening of this disease.Therefore,the purpose of this paper is to establish a multi-task learning network model,which can recognize the expression and predict whether suffering from Parkinson’s disease,and improve the accuracy.(3)For facial expression recognition,the neural network model in this paper mainly adopts Dense Net and improved Goog Le Net structure.The improved structure is the Inception module,and the original 5×5 convolution kernel is replaced by two 3×3convolution kernels,at the same time,a 3×3 convolution kernel is added after the original 3×3 convolution kernel;The two structures were combined to get the recognition results of facial expression.By using the improved expression recognition model for experiments and comparing with other classical CNN models,the effectiveness of the improved model is verified,and the accuracy of facial expression recognition is improved.In the experiment,two ways of facial expression recognition are designed.The recognition rate of the improved model is 69.73% in the first way and 70.22% in the second way.(4)For Parkinson’s disease prediction task,because of the difference between some expressions(such as fear and anger)of Parkinson’s patients and those of non Parkinson’s patients,it is possible to recognize Parkinson’s patients driven by convolutional neural network based on facial expression recognition.Therefore,a novel multi-tasking learning network is proposed to recognize Parkinson’s patients,called DMSNet model.The model continues to use the above facial expression recognition structure.The specific methods are: firstly,the original data are input into the two networks,and the potential features are extracted by residual learning and multiscale convolution,and then the two kinds of potential feature vectors are added to the full connection layer to obtain expression recognition results;then,the combination of each person’s six types of expression vectors is input into the next full connectivity layer to obtain the prediction results of the most severe Parkinson’s patients.Through a 5-fold crossover experiment on Oulu CASIA dataset and PDface dataset and a horizontal quantitative comparison with other classical deep learning models,the effectiveness of DMSNet model in the diagnosis of Parkinson’s disease has been verified,the average recognition rate of DMSNet model for the prediction of Parkinson’s disease is nearly 100%,the superiority of DMSNet model using multi-task learning mechanism is also verified by comparing with other single task learning models which only predict Parkinson’s disease,the recognition rate of improved structure in single task learning is 99.89%.In the experiment,the optimization target of the multi-task learning model is facial expression recognition and Parkinson’s disease diagnosis,and the total loss of the experiment is the sum of the two kinds of prediction tasks.In addition,the Oulu CASIA dataset and PDface dataset were combined,and a 5-fold crossover experiment was conducted.Finally,the experimental results show that the multi-task learning of Parkinson’s disease diagnosis combined with facial expression recognition technology is effective.