Research On The Model Lightweight And Pruning Algorithm Of Object Detection Network

Society’s demand for intelligent vision technology is becoming stronger and stronger,which also promotes the rapid development of computer vision technology.The highspeed development of visual technology is also feeding society.The social development of the new era cannot be separated from computer vision technology.Object detection is a branch of computer vision technology,which has been used in various fields such as industry,medicine,transportation,security,military,and so on,and has achieved good results.It can make social life more convenient,can help all walks of life to deal with a variety of visual tasks,cost-saving.Current research in the field of object detection focuses on improving the detection accuracy,detection speed,and easier deployment of object detection technology.However,most of the object detection network models have poor generalization ability and low accuracy in the recognition of general objects.Meanwhile,most of the network models are difficult to deploy on mobile terminals and embedded devices.Focusing on the challenges faced by lightweight object detection networks,this thesis starts from improving the accuracy of the object detection model and a lighter object detection model and carries out research on lightweight object detection technology based on deep learning.The research contents include Anchor parameter generation algorithm based on clustering,lightweight network model based on multiscale fusion,structured network prune method based on L1 regular.Specific research contents and contributions are as follows:1.Aiming at the problem of insufficient accuracy of current Anchor hyperparameter settings and excessive dependence on manual settings,a new cluster-based Anchor generation algorithm is proposed.This method generates a more suitable data set by processing the object frame in the real data set.The size of the Anchor box.The experimental results prove that this method can improve the detection accuracy of the object detection network.At the same time,the method has good portability and can be used in multiple Anchor-based object detection network models.2.Due to the high complexity of most object detection network models and too many parameters,it is difficult to deploy problems to mobile terminals or embedded devices in practical applications from the perspective of lightweight design models and network models.The problem puts forward a method of combining lightweight object detection with the anchoring algorithm proposed in this paper.This method fully combines the image context information,and at the same time,replaces the deeper network structure with a wider network structure.Experimental results show that this method has advantages in the lightweight object detection network model,and can reach an accuracy of 20.9% on the MS-COCO2014 data set.3.For large-scale network models(high-precision,but complex and slow),it is difficult to deploy them to actual application devices.A structured network pruning method based on the L1 norm paradigm is proposed.This method can prune the trained single-layer or multi-layer network and delete the extra layers.Through experiments,it is found that this method can basically maintain the accuracy of the object detection network model,and greatly reduce the complexity of the network model.The pruning rate of the method proposed in this paper is 47.33% for the network model and 28.61%for the image classification network VGG-16,which proves its application value.