Model Compression And Optimization Based On Objection Detection Algorithm

In recent years,With the diversification and progress of the object detection field,it provides the technical theoretical foundation for the development of a series of new industries,such as intelligent transportation,intelligent positioning and intelligent factory to some extent development.In today's Internet of everything environment,mobile terminal devices present an explosive growth trend,bringing great convenience to people's work and life,but mobile terminal devices have limited resources,which makes it difficult to deal with object detection tasks that consume a lot of system resources.If the object detection algorithm originally working under a large server can be processed and optimized by a series of model compression,acceleration and optimization methods,so that it can be deployed to mobile terminal devices faster and smaller,it is helpful for the implementation and promotion of the objection detection algorithm.At present,many scholars have studied the lightweight deployment of objection detection and achieved the expected results,but there are still some challenging problems to be solved.By studying the existing model compression optimization methods,this thesis makes some improvements and upgrades to the common objection detection algorithms.Specific research contents are as follows:1.In the third generation fast real-time objection detection(YOLOv3)algorithm,a new improved optimization method is proposed to solve the problems such as poor small-scale objection detection effect,poor accuracy and severe impact of batch processing size on model training.Firstly,bidirectional feature pyramid network(BiFPN)and the filter response normalization(FRN)technology are fused to improve the accuracy of the traditional YOLOv3 algorithm.Secondly,by use lightweight network the second-generation shuffle network(ShuffleNetv2),the model space is compressed and the inference speed is improved.Finally,experiments were carried out on COCO2017 data set,and the simulation results showed that the proposed optimization algorithm can improve the AP average(mAP)significantly,and the space occupied by the model and the total floating point computation amount decreased.In addition,through the comparison experiment of each lightweight network,ShuffleNetv2 is more suitable as a lightweight network model structure for the modified YOLOv3 algorithm2.Aiming at the problems of detection time and model space of Faster-RCNN algorithm,an optimization scheme for model compression and acceleration of traditional fasters-rcnn algorithm was proposed.Firstly,better superparameters were obtained by means of hybrid pruning and ADMM dynamic regularization to complete Faster-RCNN pruning.Secondly,a knowledge distillation scheme based on uniform quantization is considered to further compress the acceleration model,improve the target speed of Faster-RCNN identification and minimize the model;In the end,comparison experiments on PASCAL VOC 07+12 data set show that the proposed optimization scheme can reduce the detection time,improve the model detection accuracy and reduce the model space to some extent.