Research On The Bounding Box Regression Loss Function Based On IoU

In today’s highly developed deep learning technology,the field of computer vision is divided into low-level,middle-level and high-level tasks.The low-level task refers to the operation of image pixels,including filtering,restoration and reconstruction,super resolution and style transfer,etc.The middle-level task is to extract various features on the basis of pixels,while the high-level task is to detect and recognize pictures by simulating the brain.There are many ways to optimize object detection.The loss function is taken as the starting point,and the location loss function is mainly considered,especially the bounding box loss function based on IoU is modified.Through the analysis of the underlying principle of their mathematical formula,it is concluded how to play a role in the process of bounding box regression,and through the analysis and inference of this process,the shortcomings and advantages of the existing functions are summarized,learning from each other,and a new function DDIoU Loss is proposed.After analyzing the advantages and disadvantages of the innovative function,the optimized version of CDIoU Loss is further introduced.In the process of detecting the object,the NMS determines a final rectangular box from the returned candidate boxes to serve as the object’s positioning description,which is a best-of-best way to determine the unique rectangular box.In the research and training,it is found that the original NMS has certain defects,such as missing detection and false detection,which will affect our detection accuracy.Drawing inspiration from the regression loss function,we improved the original NMS and proposed DDIoU-NMS and CDIoU-NMS to fully consider the positioning problem rather than simply using IoU to judge.Our simulation experiment and application experiment show that our proposed loss function is feasible.No matter in one stage or two stage object detection framework,loss function as an independent small module has the characteristics of strong operability,less code,low cost and so on,which can be better integrated into the object detection framework.The training and testing were carried out on PASCAL VOC,MS COCO and other two public large-scale data sets.Compared with the previous regression loss function,the AP value in this paper achieved a better effect.After introducing the improved NMS,compared with the original NMS,the situation of missing detection and wrong detection in some kinds of pictures is improved,and more accurate bounding boxes are given.What is most worth mentioning here is that the combination of CDIoU Loss+CDIoU+NMS proposed has achieved the optimal detection result in either a one stage or two stage framework.