Deep Model Pruning Method For Embedded Scene Recognition

Exoskeleton robots can enhance the capabilities of the human body and have a wide range of applications in civilian,medical and military fields.In view of the problem of changing terrain and strong instantaneity during the descending process,it is necessary to design a fast scene recognition module for the exoskeleton robot to recognize the type of the descending scene in real time.Scene recognition can be done through deep learning models.However,the large amount of parameters and calculations of the deep scene recognition model make it difficult to deploy on embedded devices with high real-time requirements.Model compression technology can reduce the amount of memory and calculations during model inference,and improve the efficiency of model inference,thereby alleviating the difficulty of model deployment in deep learning.Considering that the model obtained by the structured pruning method is easier to deploy to embedded devices,this paper mainly studies the model compression method based on filter pruning.Inspired by feature selection and principal component analysis methods,this paper designs a filter pruning algorithm based on the variance of feature maps.The variance of the feature map can be used to evaluate the information content of the contour in the feature map,and the contour is an important feature in the classification task.The algorithm uses the variance of the feature map as the basis for pruning.Each layer cuts out the filters whose variance is lower than the threshold.After iteratively pruning the original model many times,the original model is compressed.Experimental results prove that the algorithm can effectively detect useless feature maps,thereby removing the corresponding filters,and can effectively improve the accuracy of retraining after network pruning.On the other hand,most of the current filter/channel pruning methods are separated structures and do not consider the relationship between two adjacent layers in the network.Based on this,this paper proposes a pruning algorithm that weighs the geometric median of adjacent layers.The algorithm adopts a geometric median pruning strategy,and on this basis,a weighting factor is set to weigh the influence of the filter and convolution kernel group corresponding to the adjacent layer on the pruning effect.The experimental results prove that the corresponding filters and convolution kernel groups between adjacent layers in the neural network jointly affect the performance of the model pruning,but the filter of the current layer has a greater impact on the pruning effect than the convolution kernel group of the next layer.Adding a weighting factor to associate adjacent layers can overcome the shortcomings of the separated structure and make the accuracy of retraining after the network pruning higher.In addition,in a jumping scene,it is necessary to design a fast scene recognition module for the exoskeleton robot to recognize the terrain of the jumping scene in real time.This paper designs and implements an embedded hardware system based on NVIDIA Jetson TX2.The system collects external image data through a camera and uses a battery as the main power source to design an embedded hardware system with acquisition,portability,storage,processing and transmission capabilities.The scene recognition model is pruned by the pruning algorithm proposed in this paper,and then deployed on the embedded platform of the design.The test results show that the inference speed of the pruned Res Net-50 is faster than the original model,and the image processing speed can reach 37 frames per second,which can realize the rapid scene recognition of the system.The research work in this paper shows that the pruning strategy based on the variance of the feature map can effectively remove the useless filters in the network,and adding a weighting factor to the filter pruning to associate two adjacent layers can overcome the shortcomings of the separated structure.Both can improve the performance of pruning.In addition,the research work in this paper provides a solution for the rapid scene recognition module for exoskeleton robots in a jumping scene.