Research And Implementation Of Improved Bounding Box Regression Algorithm For Target Detection

Object detection is a core topic in the field of computer vision,object detection technology aims to identify and locate objects from information such as pictures and videos,and is the basis for high-level visual tasks such as semantic segmentation and image description.Nowadays,object detection technology has a wide range of applications in face recognition,automatic driving,industrial inspection,medical image recognition and other fields.For the localization task of object detection,bounding box regression is one of the key factors to determine the accuracy of object localization.The bounding box regression losses can be roughly divided into two categories:the traditional L_n paradigm losses(L₁,L₂,Smooth L₁;IoU-based losses(IoU,GIoU,DIoU,CIoU,etc.;.The traditional L_n paradigm losses has been criticized for calculating the loss of each coordinate parameter independently and ignoring the connection between coordinate parameters,so most existing bounding box regression losses are based on IoU.This thesis analyzes the existing bounding box losses:firstly,it is found that the existing bounding box losses function does not discuss whether IoU has a better form,and they all use IoU as the basic loss item,and they cannot be used in some special cases.Efficiently describe the pros and cons between bounding boxes.Secondly,the imbalance between high and low quality samples in the process of bounding box regression is not handled well,and effective samples are not deeply excavated.Based on the above problems,this thesis researches and improves IoU loss,loss penalty items,and effective sample mining.The main research content and work are as follows:(1;Compared with the gradient of Smooth L₁ loss and IoU loss,it is found that the gradient trends of Smooth L₁ loss and IoU loss are completely opposite,the gradient of Smooth L₁ loss decreases when it is close to the effect of high-quality regression,while the gradient of IoU loss increases when it is close to the effect of high-quality regression.This thesis argues that this is one of the reasons why IoU loss is better than Smooth L₁ loss.Based on this discovery,this thesis improves the representation of IoU on the basis of IoU loss,which is called NIoU loss.Secondly,this thesis analyzes the additional penalty items based on IoU loss,and finds that the additional penalty items are all designed from the perspective of the bounding box returning to the whole,without considering the regression effect of each edge of the bounding box separately.Based on this problem,this thesis proposes a margin loss penalty item,which calculates the distance difference between each side of the target box and the prediction box,and refines the regression for each side.At the same time,considering the unstable fluctuation of bounding box regression that may be caused by adding four relatively independent penalty items.This thesis introduces a penalty term based on the cosine similarity of the bounding box diagonal vector to constrain the regression direction by controlling the angle between the two in the vector space.The above three improvements are combined to propose LIoU loss,which is trained and verified by the MS COCO 2017 dataset,and compared with the existing bounding box regression loss,the best AP is achieved,which proves the effectiveness of the method.Using the bounding box regression loss on models such as Retina Net and PAA,these models have improved AP,which proves the versatility of this method.(2;To address the imbalance between high and low quality samples in bounding box regression,this thesis improves the existing bounding box regression loss weights.This thesis introduces the NIoU proposed in this thesis as a weight control factor,and defines the samples whose average value is lower than the average NIoU value in each training batch as low-quality samples,and those whose average value is higher than the average NIoU value are defined as high-quality samples.The low-quality samples in the bounding box regression process are down weighted,and the high-quality samples are weighted.This thesis proposes a weight coefficient,combines the proposed weight with LIoU loss,and passes the MS COCO 2017 data set training and verification.Compared with the existing methods,the method proposed in this thesis has a higher AP improvement,which proves the effectiveness of the method.And the application of Retina Net,PAA and other models has effective AP improvement,which proves the versatility of the method.