Research On Deep Learning Object Detection Method Based On The Distribution Of Prediction Boxes

At present,task conflict has gradually become one of the bottlenecks restricting the development of object detection networks.The reason is that the mainstream object detection networks need to optimize the classification task and regression task in the same backbone network.On the one hand,the two tasks can not effectively assist each other during training,but compete for the optimization direction of the network instead.On the other hand,the mainstream object detector adopts the classification confidence to measure both classification quality and regression quality of prediction boxes,but a large number of studies have indicated that the regression quality of the prediction boxes cannot be well measured by the classification confidence in some cases.In this study,it is found that the prediction boxes present different distribution when in the target area and non-target area,which explicitly indicates the location of the target.This paper proposes the concept of density based on the distribution characteristics,and improves the Faster R-CNN network with the density,trying to alleviate above problem by coupling the classification and regression tasks through the density.The main work is summarized as follows:(1)In the process of regression,it is found that the cluster degree of the prediction boxes in the target area is much greater than that in the non-target area.Therefore,this paper defines the degree of overlap between each prediction box and other prediction boxes as density.The density can reflect the degree of deviation of the prediction boxes from the foreground object,and further can evaluate the regression quality of the prediction boxes.This paper proposes two methods of computing density for different usage scenarios,which can describe the density of a single prediction box and the overall density of the image respectively.(2)In order to improve the ability of region proposal network to classify the foreground and background,this paper proposes density-region proposal network(D-RPN).D-RPN firstly calculates the density heatmap using the anchor boxes after regression,and then enhances the original feature map to obtain a new feature map with the heatmap.Lastly,the new feature map is used to predict the foreground confidence.The proposed method makes the regression results efficiently assist the classification task with the density heatmap,reducing the impact of task conflict on the classification task.Experimental results show that the density heatmap in D-RPN can effectively assist the classification task,and improves the ability of RPN to distinguish between foreground and background without affecting the regression task.(3)To tackle the problem that the classification confidence cannot effectively measure the regression quality of the prediction box,this paper proposes a general confidence metric combined with the density.ROI head,based on the general confidence,firstly utilizes Ro I pooling to get the tensor of the corresponding position of the candidate box from the density heatmap as its density,and then,the general confidence is obtained through the comprehensive learning of the fully connected network,which takes the classification confidence and the density as inputs.Experiments show that the general confidence can remedy the defect of classification confidence through the density of prediction boxes,which can more completely evaluate the quality of the prediction box and improve the detection ability of the network.