Research On Hand Fine Motor Recognition Algorithm Based On SE-CNN Network

With the advent of China’s aging society,the number of patients with limb paralysis caused by stroke is increasing year by year,bringing heavy pressure and economic burden to society and families.The reconstruction system for paralyzed limb motor function designed by our research group can effectively help stroke patients recover lost limb motor function to the greatest extent and improve their self-care ability.Gesture recognition is a key part of the development of a prototype system for hand fine motor function reconstruction,the purpose is to identify the collected multi-channel EMG signals,determine the movement intention of the patient’s unaffected limb,and realize the control of the functional movements performed by the patient’s affected limb.At present,the research on gesture recognition algorithms based on deep learning is still in the preliminary stage,and the accuracy and number of model parameters of deep neural network model gesture recognition pose great challenges to the realization of deep learningbased gesture recognition on embedded systems.The purpose of this paper is to study the convolutional neural network algorithm that can perform gesture recognition,and to propose an improved algorithm to further improve the gesture recognition rate,as follows:（1）With the high development of modern science and technology,there are many action recognition models based on deep learning,in order to improve the accuracy of their label classification,the number of stacked network layers is often used to improve the complexity of the model.However,this increases the reaction time of the system and increases the delay between system recognition and rehabilitation training,which is not conducive to the patient’s real-time training rehabilitation exercise.（2）The sharp increase in data volume will increase the training time of the model,which is not conducive to cost savings,however,the improvement of model recognition accuracy also depends on large-scale datasets.（3）The use of a large number of pooling layers will inevitably lead to the loss of fine information,which makes it difficult for the model to identify fine movements.But at the same time,in order to reduce the amount of parameters,the pooling layer is an indispensable existence,and the fine information is retained.In view of the above three problems,firstly,this paper builds a SE-CNN network model,which uses a large number of pooling layers to reduce the amount of parameters of the input connection layer,improves the recognition accuracy in terms of sliding window sampling length and optimizer selection,and effectively reduces the number of parameters and reduces the reaction time of the model by accumulating dropout parameters.Secondly,in order to preserve fine movements,the model uses the SE module to compensate for the shortcomings of the pooling layer.Finally,the training effect of the model on NinaPro and SIA_delsys₁6_movement datasets is compared,and the appropriate configuration is saved to improve the training effect of the network model,thereby reducing the corresponding time of the system and improving the recognition accuracy of the network model,and the average recognition accuracy can reach 93.84%.Ablation experiments show that the proposed algorithm obtains higher recognition accuracy than the existing algorithms,and has better recognition effects on different types of datasets.