Yolov3 Object Detection Model Compression And Embedded Deployment Based On ZCU102

In view of the huge consumption of computing and storage resources,it is very difficult to deploy the object detection network model on the embedded devices.This paper selects the yolov3 network model in the object detection,starting from the lightweight network structure design and network pruning.The volume of the model is greatly compressed on the premise that the accuracy loss is acceptable,making it easily to deploy model on embedded devices,and the inference speed of the model can be accelerated.The main work includes:(1)Aiming at the convolution layer of darknet-53 which contains a large number of parameters and calculations,a lightweight network structure design method is designed.The depth separable convolution is used to replace some convolution layers,which greatly reduces the amount of model parameters and calculation on the premise of ensuring the accuracy of model detection.(2)A channel pruning algorithm is designed for yolov3 network model.Taking the γ coefficient of BN(batch normalization)layer as the important measure standard of pruning unit in the model,the network model is trained sparsely first,and then channel pruning is carried out on the basis of γ and the network model is further compressed.(3)After the network is lightweight and pruned,on the premise of ensuring the recognition accuracy of the network,the dnndk development kit of Xilinx is used to deploy the network to zcu102 board to verify the detection accuracy and recognition speed of the model.Experimental results show that the target detection network based on the proposed algorithm performs well on Pascal VOC data set and coco data set.On Pascal VOC dataset,the volume of the model is reduced from 240M to 4.8M,the model is compressed to 2%,the mAP of the model is reduced from 0.79 to 0.78,and the reasoning speed is reduced from 16ms to 6ms;On the coco dataset,the volume of the model is reduced from 240M to 6M,the model is compressed to 2.5%,the model mAP remains unchanged at 0.79,and the reasoning speed is reduced from 16ms to 4.8ms.After the model is compressed,dnndk is deployed on the embedded platform zcu102,and the inference speed of the final model can be maintained within 3.8ms,which is faster than that of PC.