Research On Optimization Of Object Detection Model Based On FPN

In recent years,the Artificial Intelligence(AI)based on Convolutional Neural Network(CNN)has promoted the development and practical application of computer vision.Object detection has made great progress due to the successful application of CNN model.Various advanced object detectors emerge one after another,which promotes the practical application of computer vision technology based on object detection.The backbone network of the object detector based on the CNN model not only influences the final detection,but also affects the performance of the detection head.On the basic of the image classification model,the researchers proposed and improved a series of backbone networks for object detection.The backbone network represented by the Feature Pyramid Networks(FPN)has become an indispensable part of the object detection model because of its good effect on small objects and its adaptability to multi-scale changes of objects.However,the existing detectors based on FPN and its improvement faces many problems and challenges.Firstly,there are disadvantages on feature fusion for the existing detectors based on FPN and its improvement.Secondly,there is a contradiction between feature extraction performance and the depth of backbone network on the existing detectors based on FPN.In view of the above two problems,this paper conducts the following research:First of all,this paper analyzes the disadvantages on feature fusion for the existing detectors based on FPN and its improvement,and proposes Twin Feature Pyramid Networks(TPFN)for object detection.When ensuring the advantages of FPN model in small object detection,this paper proposes S Pyramid and E Module based on the original FPN network(F Pyramid).Among them,S Pyramid is a small feature pyramid networks added on the basis of FPN,which effectively improves the detection effect of medium object with further fusion of context information.The E Module can effectively improve the detection performance of the model on the large object by effectively integrating the shallow features of the F pyramid and the features of the S pyramid.With these three parts,we design an object detection network based on TFPN.A large number of experiments have proved that compared with the previous object detector based on FPN and its improvement,the object detection network based on TFPN can simultaneously improve the detection performance of small,medium and large objects.What’s more,this paper analyzes the contradiction between feature extraction performance and the depth of backbone network on the existing detectors based on FPN,and proposes an object detection model optimization method,named Cascade FPN.This method has the following advantages: first,it builds a lightweight FPN based on a Conv Block with less parameters and computation,which effectively reduces the redundancy caused by the deepening of the horizontal expansion of the network.Second,the context information fusion mechanism is built on the basis of feature fusion module FFM,which fully integrates context information while extending lightweight FPN in horizontal cascade.The experimental results show that while deepening the backbone network lengthwise,the lightweight network in the horizontal direction can effectively improve the result of object detection with less redundancy.