Research Of Object Detection Based On Anchor Free

Object detection is a very challenge subject in the field of computer vision and the basis of many applications of computer vision and which focuses on how to capture and recognize objects quickly and accurately from natural scenes.It combines image processing,artificial intelligence,pattern recognition and others.It has a wide range of applications in medical,transportation,military,security and other fields.This paper first analyzes the traditional object detection methods.Traditional methods have many disadvantages,such as large time complexity,poor robustness,rebundant window,and manual design.With the development of artificial intelligence,object detection methods based on deep learning has been proposed.They have been widely studied and applied.All of them use convolution neural network to extract and analyze the features of image.They have high accuracy,fast speed and no manual operation.Object detection methods based on deep learning can be roughly divided into two categories,one is based on the two-stage(eg.rcnn)series of detection models,and the other is based on the one-stage(eg.yolo)series of detection models.The experimental analysis of the two methods in this paper shows that both methods have the disadvantage of artificially constructing a prior proposal or anchor in advance,and rely on the prior knowledge of the prior box too much.In the past two years,scholars have begun to successively study object detection methods that discard prior box,which are known as anchor-free.Although anchor-free methods are noval and improved,they still have plenty of room for improvement.Therefore,our paper will continue to study method based on anchor-free,which does not depend on preset candidate box,enhance robustness and improve the accuracy.The result are very sensitive to the number,size,and aspect ratio,which make it difficult to adjust.The preset anchor size and width are fixed,which limits the generalization ability of the detector.Generally,in order to ensure the detection effect,a lots of anchors are needed,causing the problem of imbalance between positive and negative samples,increasing the calculation amount and memory consumption.It was found that the FCOS method is a good method in terms of accuracy and speed among the existing anchor-free methods.However,the center-ness defined by FCOS is not good enough.So we design Area-ness which can better suppress the generation of low-quality target bounding box,punish the pixels that are far from the center of the target bounding box,and increase the contribution value of pixels that are close to the center.the results show that the area-ness improve center-ness by 0.2 points m AP,and m AP achieves 44.9.In order to improve the accuracy of detection,convolution neural network(CNN),such as residual network(Res Net)mostly increase the depth or width of the network,but they lead to lacking information in the prediction layer.some objects will be misdetect,especially small.Therefore,researchers have proposed Feature Pyramid Network(FPN),which had a top-down architecture with lateral connections developed for building high-level semantic feature maps at all scales.We obsever that it does not make full use of different features between different layers.Different features usually contribute to the fused output feature unequally.In this paper,we introduce a new feature pyramid architecture named Adaptive Weighted Dual-stream FPN(A-DFPN),which can fully fuse the features of the upper and lower layers to achieve the best results.the results show that Adaptive Weighted Dual-stream feature fusion structure improve the existing feature fusion methods by 2.1 points,and m AP achieves 47.0,m AR achieves 36.7.