Object Detector Based On Elliptical Proposal Boxes

Object detection is always a hot topic in computer vision field,whose purpose is to locate and recognize objects in images.In recent years,with the rise of deep learning,many object detection methods based on convolutional neural network have emerged.These methods have been successfully applied in face recognition,vehicle detection and remote sensing image detection.However,the detection effect of the deep learning model depends heavily on the training dataset and its annotation form,while the existing datasets is labeled with rectangular boxes so that the object detection model can only detect objects with rectangular boxes,but the rectangular boxes usually cannot well encompass objects with different poses and additional background information is often introduced,which results in inaccurate model positioning.In order to overcome this problem,considering that the oriented elliptical box can adapt to the different orientation of the object to better envelop the object,the dataset with the labeled of the oriented elliptical boxes is constructed and the object detection model used to predict the oriented ellipse box is designed.The main research contents and contributions of this thesis are as follows:(1)An elliptical proposal boxes detection network based on the center of the ellipse is proposed.Existing object detection methods can be divided into two categories: anchor based methods and anchor free methods.The anchor boxes is mainly based on some prior information so that the model can regress from the anchor boxes to the ground truth boxes.Anchor free methods directly predict the position of the object.Considering that the introduction of elliptical anchor boxes would increase the computational complexity,a Ellipse Net is designed based on the basic idea of anchor free methods,which could predict elliptical center,length of long and short axis and oriented angle directly.In addition,in order to improve the detection accuracy of the model,an efficient elliptical intersection ratio loss function is designed.The experimental results show that the model is simple in structure because there is no complex anchor boxes,and the performance is similar to the oriented detection methods.(2)A sparse learnable oriented elliptical boxes object detection methods RSparse RCNN is proposed.Existing object detection methods are dense detection methods and require complex post-processing such as non-maximum suppression in the prediction stage.In order to overcome the above problems,a group of sparse learnable elliptical boxes is preset to participate in network training and the network is learned the preset offset of elliptical boxes.In addition,the network cannot be trained because the loss of the intersection ratio between the two elliptical boxes is indifferentiable.So the Gaussian Weinstein distance function is used to transform the oriented ellipse into the corresponding bivariate Gaussian distribution,and calculate the loss by regressing the Weinstein distance between the Gaussian distributions.In the prediction stage,the detection can be completed by correcting the position of elliptical boxes directly through network output.Since there are no complex post-processing steps during the whole process,the model achieves a trade-off between accuracy and speed.(3)A lightweight object detection method RCenter Net for oriented elliptical boxes is proposed.Existing deep learning models are difficult to deploy to edge devices due to high computational complexity and large memory footprint.By combining the lightweight skeleton network with the idea of anchor free method,a lightweight network which can directly predict elliptical box is designed,and no complex post-processing such as non-maximum suppression is required in the prediction stage.Experimental results show that RCenter Net’s space occupancy is significantly reduced and achieve the real-time requirements.