Research On Compression And Deployment Of Deep Learning Model For Object Detection

In recent years,with the rapid development of deep learning in target detection task,the deep learning model deployed in UAV becomes more and more mature.In order to obtain better detection performance,the depth of the network in previous studies has been continuously increased,resulting in extremely large number of parameters and computation of the network,which poses a challenge for UAVs with limited resources.Therefore,when applying deep learning networks in practice,it is necessary to reduce the number of parameters and computation while maintaining the detection performance,which is an urgent problem to be solved in the current research.Therefore,this paper focuses on the compression and deployment of the deep learning model,explores how to optimize the existing target detection network to achieve better performance on the self-made data set,and on this basis,studies how to efficiently compress the parameters and computation amount of the model,and designs a reasonable architecture to deploy the compressed network model for application.The research content of this paper includes the following three aspects:(1)Basic network optimization: In this paper,the deep learning model is compressed and deployed for unmanned aerial vehicles(UAVs)used in military reconnaissance.Due to the different types of measured targets,large size span and complex image background.In order to resolve the measured objects with high precision and fast,this paper selects YOLOv3 as the basic network to detect the above objects.On the basis of collecting and analyzing the image features of the self-made data set,clustering algorithm and genetic algorithm are used to recluster the anchor frame and the classifier is replaced with Softmax function.In addition,due to the large size span of the target to be measured in the image,SPP module is introduced after the Dark Net-53 backbone network in order to realize the fusion of multiple sensitivity fields.The experimental results show that compared with the basic detection network,the optimized network in this paper has improved m AP by 3% and reduced loss function by 18%,which has better detection efficiency.(2)Model compression: In order to solve the contradiction between the huge number of parameters and computing amount of the YOLOv3 network model and the limited computing and storage resources of the UAV,the model must be compressed.In this paper,the deep separable convolution is introduced to reduce the calculation of the model.Then,the PR regularization term was used as the regularization term of sparse training to better distinguish scaling factors.Then,according to the scaling factor,the model was combined with channel pruning and layer pruning,in order to reduce the number of model parameters and the amount of calculation.Finally,since the data generated by the training platform is a 32-bit floating point number,Do Re Fa-Net quantization method is used to quantify the model,so as to compress the storage capacity of the model.The experimental results show that the proposed compression scheme can effectively reduce the number of parameters of the model by 5.89 M and the calculation amount by 12.51 GFLOPS,while maintaining the original detection performance.(3)Model deployment: In order to verify the application of the optimized model on the corresponding hardware platform,based on the analysis of the advantages and disadvantages of the existing hardware deployment platform,a typical airborne computing platform ZYNQ is selected and an architecture mode of "ARM+FPGA" is designed to deploy the YOLOv3 network model.In addition,combined with the reasoning process of the network model and the operation characteristics of the deployment platform,the data transmission unit and CNN acceleration unit of the model are optimized by means of Ping-Pong data cache,line cache and parallel pipelining.Finally,the ZYNQ development tool is used to verify that,compared with other relevant research designs,the performance is guaranteed while the resource proportion is less and the power consumption is lower,which can meet the conditions for the deployment of UAVs.