Research On Compression Algorithm Of Deep Learning Model And Its Application In Target Detection

The development of artificial intelligence has shown great application value in the economic field or social life.Among them,deep learning has achieved excellent results in image processing,natural language processing,and computer vision.The deep network model can achieve such outstanding effect thanks to its large number of parameters and complex network structure,but it also brings two problems: one is the huge amount of calculation,which results in the reduction of model speed;Second,a large number of network parameters lead to a sudden increase in the size of the model,resulting in the need for more space for model storage.These problems make the deployment and operation of deep network model in some environments or devices with limited resources such as storage and operation more difficult.Therefore,how to simplify the parameters of the network of target detection that the speed of model reasoning is improved and the accuracy is acceptable has great research significance.Firstly,this paper analyzes the algorithm of one-stage target detection,analyzes the characteristics of YOLOv3 model based on different backbone networks,and selects the appropriate model as the research object.Combined with the relevant literature,this paper proposes a combination strategy of model compression algorithm suitable for the target detection model YOLOv3,according to the current real-time task landing requirements,the model deployment based on the Jetson TX2 embedded platform is completed.The experimental results show that the compression method in this paper can achieve 3.9times model acceleration and 4.19 times model volume compression in the embedded platform,and the accuracy is improved by 1.1%.