Research On One-stage Object Detection Method Based On Full Convolutional Neural Network

Object is one of the challenging tasks in the field of computer vision,and the accuracy of object algorithms based on convolutional neural networks has been gradually improved with the development of relevant theories of deep learning.However,in practical applications,the algorithm model is too complex and large,the computation is large and other shortcomings make the algorithm inefficient.In addition,there is room for improvement in scalability and accuracy.In this thesis,we propose a one-stage object model Efficient FCOS based on full convolutional neural network to address the existing problems of efficiency,accuracy and scalability of the one-stage object algorithm.Secondly,we compare and analyze the current convolutional networks,and use Efficient Net as an alternative to Res Net for the problems of many network parameters and low efficiency.is used to replace Res Net as the backbone network to streamline the network structure of Efficient FCOS and improve the efficiency of the object model.Then,to address the drawback that FPN only has top-down information fusion,BIFPN is used instead of FPN to add bottom-up aggregation network to enhance the network model’s ability to fuse feature map information and improve its detection capability for targets of different sizes.Finally,the geometric factors of real and predicted frames are taken into account in the loss calculation,and the DIo U-based loss function is proposed,which in turn improves the robustness of the model and speeds up the convergence of the model.Experiments on Pascal Voc 2007 dataset show that EF7 achieves 82.1% m AP,which is a2.8% improvement compared to FCOS.The network parameters are about 1/5 of FCOS,the floating-point computation is about 1/3 of it,and the GPU LAT is improved by 12.2% and the CPU LAT is improved by 20.0%.In addition,EF7 achieves an average accuracy of 79.3% and41.9% on the Pascal Voc 2012 and MS COCO datasets,respectively,leading most of the onestage object detection algorithms and some of the two-stage object detection algorithms.