Research On Lightweight Method Of Target Detection Based On Improved FCOS

Target detection is one of the mainstream directions of computer vision.With the rise of deep learning,it is widely used in smart cities and new infrastructure,security,finance,medical and health,e-commerce and physical retail,unmanned driving and other scenarios.Now,target detection algorithms based on deep learning have developed into two technical routes: Anchor Based method and Anchor Free method,due to the large number of overparameters and redundant frames required by Anchor Based method,serious imbalance of positive and negative samples results in high design requirements,while Anchor Free method has larger and more flexible solution space.It reduces the amount of calculation caused by the use of Anchor so as to further improve the real-time performance and accuracy of target detection.Therefore,the Anchor Free target detection algorithm FCOS(full convolution single-stage target detection)based on deep learning is improved in this paper to improve its detection accuracy,and lightweight replacement of the network is carried out to reduce the model size and give consideration to the accuracy,which can make it more widely applied to various scenarios.The specific research work of this paper is as follows:(1)Aiming at the problem of insufficient detection ability of FCOS,a new model structure IM-FCOS is designed.First,Backbone network remains unchanged.Then,in the Neck network structure,the improved BiFPN weighted bidirectional feature pyramid was introduced,and repeatd blocks were not used repeatedly,which ensured that the multi-scale feature information of different layers could be fully extracted,and the network was also reduced to reduce the computation and parameter number of the model.Finally separately in the Head(testing)detection in order to detect the characteristics of the different size chart input,use classification branch QFL focus(Quality Focal Loss)to predict the classification and improve the quality of frame,further eliminate the differences in training and forecasting confidence level,the branch of regression using DFL(Distribution Focus Loss)to improve the border position distribution.At the same time,GIo U Loss is added to assist convergence to improve the accuracy of coordinate regression.Compared with FCOS(Base Resnet-50)algorithm,the m AP of detection accuracy is improved by 3.6% through COCO2017 data set test,and the detection accuracy is improved.(2)In view of the improved FCOS method proposed above,its model occupies large storage space and has a large number of parameters,which limits the real-time performance and deployment problems of the model.Therefore,a lightweight network structure named LIM-FCOS is proposed in this paper.In Backbone,ShuffleNetV2 is used to replace Resnet-50,and the improved BiFPN method proposed above is continued to be used in Neck structure.DSC(Depthwise Seperable Convolution)is used to replace ordinary convolution in convolution layer,and the number of characteristic channels is reduced from 256 to 128 to reduce the number of parameters.The Head method remains unchanged.Through the test on COCO2017 data set,compared with FCOS(Base resnet-50),its computation power is reduced by 10%,and its accuracy reaches 27.5%.Meanwhile,compared with The Anchor Free lightweight model Yolox-Nano,it has 1.54 m more parameters and 0.43 GFLOPs,and the accuracy has been improved by 2.2%.It shows that Network lightweight is effective.