Research On Multi-scale Object Detection Method Based On Anchor Free

With the increasing of people's living standard and life rhythm speeding up,people needs more and more artificial intelligent technology which is comfortable and saves time and effort,and object detection which is based on computer vision is an important part in the field of artificial intelligence,intelligent city,autopilot,electrical inspection and face unlock applications are based on target detection technology.Early models of object detection based on deep learning mostly use the transcendental box(anchor)to reduce the degree of difficulty of learning.Because the parameters of the models is a lot of and detector's generalization ability is limited,in recent years,scholars begin to focus on the research of the model of object detection based on the anchor-free,which abandon the priori box.And not only the network structure of the models are elegant,but also their generalization is stronger.Based on FCOS which is an anchor-free model,this paper focuses on the multi-scale problem in object detection,studies and improves the positive sample allocation of FCOS and the interference by the extreme object in multi-scale testing of object detection models.And we conduct experiments on disease data sets and COCO data sets to verify the feasibility and effectiveness of the improved algorithms.The main innovative contents of this paper are as follows:(1)For the positive sample allocation problem which is non-uniform sampling for slender object in the model of FCOS,we put forward the aspect ratio of adaptive positive sample center sampling algorithm.At the same time for the long sides of the object,centerness target value is falling faster than it's short sides.We put forward the center target value which is adaptive for the the aspect ratio of the object.This algorithm eases the problem the center of slender objects decline rapidly,which leads to inadequate training of the centerness branch of the FCOS.By combining the above two improvement strategy,we improve the robustness of the model for slender objects.(2)When FCOS is training,the sample region of a big object instance is often assigned more because of the large area,and less for a smaller object,which can lead to more attention to the study of big objects.So detection result is bad for small objects.In view of the unbalanced samples for different scale objects,based on the original sample allocation algorithm,we put forward random-k positive sample allocation algorithm,which improve the detection precision of the model as to multi-scale objects.(3)Due to the interference by the objects which is extremely large or small,when object detection model is multi-scale testing,it leads to error check and inspection.In order to ease this problem,in this paper,we put forward the relative attention of multi-scale testing module.Through training two different resolution input images together,the two branches of classification of different resolution times mask weight which is the output of the attention module.And the loss sums together,so that for model that it can learn how to better combined with the predictive results under different scales by itself.Finally the module is combined with FCOS model that is based on per-pixel prediction.Experiments on tunnel disease data set and COCO data set show that the improved network not only improves the overall accuracy,but also improves the robustness of the model for objects of different scales.