A New Intersection Over Union Loss Function And Its Application In Object Detection

Object detection is one of the basic tasks in computer vision,and is widely used in real life.Improving the accuracy and speed of algorithms are constant targets in object detection research.Bounding box regression is an important step of object detection,which improves the locating accuracy of traditional object detection.The currently widely used bounding box regression method uses a loss function based on IoU(Intersection over Union)to calculate losses and then optimizes them using the gradient descent method.This shows the importance of the loss function in bounding box regression.At present,the common loss functions include IoU loss function,GIoU(Generalized IoU)loss function,DIoU(Distance IoU)loss function,and CIoU(Complete IoU)loss function.Based on the study of existing loss functions,we proposes a new overlapping metric and loss function,and applies them to object detection algorithms.Our main research works and results are listed as follows.(1)An overlapping metric based on IWGIoU(Integrated Weighted GIoU)is proposed.Compared with existing literature,the new overlapping metric selects a weight function that is more suitable for integral calculation to weight GIoU,making the bounding box regression pay more attention to the coincidence of the center part of bounding box and target box.In addition,we found in our experiments that IoU loss function converges faster when two boxes closer to each other.Therefore,we weight GIoU and integrate it with IoU to form a new metric function,which is called IWGIoU.This metric solves the problem that GIoU is not sensitive to the relative positions,and has the characteristics of distinguishing different intersection situations and having fast convergence when the intersection is small.Our numerical experiments show that IWGIoU and its loss function possess the above characteristics.(2)An α-IWGIoU bounding box regression loss function is proposed.The construction idea of this loss function originates from the construction of α-IoU loss functions.Adding the αpower parameter to the IoU based loss functions improves the performance and robustness of the algorithms.In this thesis,the α power parameter is introduced into the IWGIoU loss function.Experiments show that the regression accuracies of the algorithms have been improved.(3)This thesis applies the IWGIoU loss function and the α-IWGIoU loss function to series of YOLO object detection algorithms,and train them on the CCPD vehicle license plate dataset and PASCAL VOC dataset.It is found that the models obtained perform best when α = 2.The experimental results also show that the models trained based on IWGIoU loss function have the highest accuracies in most cases,but occasionally they may not be as good as others.Besides,the models based on the α-IWGIoU loss function have once again improved accuracies on the basis of the former.The main improvement of these two methods is the accuracy under high IoU thresholds,which means that the methods proposed in this thesis are suitable for object detection with high accuracy requirements.From the perspective of detection effect,the YOLO models trained based on the two loss functions proposed in this thesis can better detect the central positions of objects.The center coincidences between their prediction boxes and target boxes are relatively high,which is suitable for the object detection with rich central position informations.In above works,our innovations include the following two points.(1)IWGIoU selects a weight function that is more suitable for integral calculation,and integrates with IoU to create a new overlapping metric.This metric combines the advantages of various methods and is a good metric for measuring the coincidence of center points and distinguishing relative positions.Simulation experiments show that the IWGIoU loss function is the fastest convergent loss function under the Py Torch automatic derivation framework.(2)Inspired by the construction of α-IoU loss function,we introduced a α power parameter into the IWGIoU loss function.Experiments have shown that the appropriate α power parameter improves the accuracy of the model trained based on the α-IWGIoU loss function.The structure of this thesis is arranged as follows.The research background and current status of object detection and bounding box regression are described in Chapter 1.Some common object detection algorithms and their theories are introduced in Chapter 2.The commonly used bounding box regression loss functions and their properties are summarized in Chapter 3.The IWGIoU metric and its loss function are proposed in Chapter 4.In Chapter 5,we introduce the α power parameter into the IWGIoU loss function and propose the α-IWGIoU loss function.Numerical experiments are constructed to verify the effectiveness and practicality of each metric and its loss function in neural network optimization in Chapter 4 and Chapter 5.In Chapter 6,we summarize and analyze the full thesis.