Research On Accurate Locating And Balanced Learning Algorithms In Object Detection

Object detection is a fundamental and challenging task in the field of computer vision,its research results can be widely used in automatic driving,video monitoring,robot navigation and so on.The object detector is the core of object detection research,which directly determines the application level and scope of object detection.However,the existing object detectors have not yet reached the pass line on the MS COCO dataset,the low locating accuracy and unbalanced learning are two main restrictions.Therefore,it is of great significance and value to conduct the related research.Aiming at the above two restrictions,the thesis conducts specific research from three perspectives:loss function,feature fusion,and hard example mining.The main work and contributions are as follows:(1)Research has shown that there is no inevitable relation between the ln loss commonly used by object detectors and the locating metric IoU,so a loss function directly related to IoU needs to be designed.Aiming at the problems of slow convergence speed,large training error and poor training result in the existing loss functions based on IoU,the causes of the problems are analyzed through the simulation experiment of bounding box regression,and a loss function called LIoU is designed.Related experimental results show that LIoU overcomes the problems of similar loss functions and can further improve the locating accuracy of the object detectors.(2)Aiming at the problem of imbalance features between high and low layers in Convolutional Neural Networks,the result of the Neural Architecture Search is analyzed,and a neural network longitudinal cross-layer feature fusion algorithm called PaFPN is designed.PaFPN makes up for the shortcomings of the existing feature fusion algorithms and retains all the advantages.Comparative experimental results show that PaFPN reduces the complexity of the algorithm while improving the object detection performance.(3)Aiming at the imbalance between easy samples and hard samples,the general form of the loss function based on IoU is improved by analyzing the gradient of the loss function,making it have the characteristics of hard example mining.All loss functions based on IoU can benefit from the improved general form,and an improved version of LIoU called Hard-Mining LIoU is proposed based on this form.The experimental results show that the convergence of Hard-Mining LIoU is more stable,and the ability of hard example mining is better than the existing algorithms dedicated to hard example mining.(4)Based on the above work,a object detector called CompleteDet is designed.Compared with existing object detectors,CompleteDet greatly reduces the algorithm complexity when the object detection performance is equivalent;when the algorithm complexity is equivalent,it improves the object detection performance.Based on CompleteDet,The performance improvements of every work is compared in ablation experiment.Through error diagnosis,the deeper restrictions hindering the performance improvement are analyzed,which points out the direction of further research.The above research results can further improve the performance of existing object detectors and provide some theoretical and methodological support for the future development of object detectors.