Research On Oriented Object Detection Based On Corner Positioning

Thanks to the rapid development of computer technology and the increasing maturity of neural networks,the field of computer vision is in full swing,and target detection is the most common but extremely important research area among the many research directions in computer vision.A large number of related studies such as object tracking,behavior recognition,pose estimation,Re ID and other fields are inseparable from the results of target detection,which shows the importance of target detection algorithms.Many real-life application scenarios,such as natural image text detection and remote sensing image target detection,have a large number of images containing certain rotation angles,which would lead to a large amount of redundant information if represented by horizontal detection frames only,and thus the need for rotational target detection arises.Anchor-based algorithms can be divided into Anchor-free algorithms and Anchor-based algorithms,which can significantly improve the detection effect by introducing anchors,but the ensuing problems are:first,the positive and negative samples are unbalanced,because most detection algorithms have thousands of anchors,while the actual number of targets in the image is The second is more hyperparameters,such as the number of anchors,size,aspect ratio,etc.In contrast,the Anchor-free algorithm can achieve excellent detection results without introducing the above problems.Therefore,this paper uses the idea based on key point localization in the Anchor-free algorithm to implement a rotating target detection network,and the main contributions are shown below.First,a deep learning-based keypoint estimation network is implemented.Based on the deep network of Res Net to extract features,the up-sampling,feature correction and channel adjustment modules are added in order to enable Res Net to achieve the key point estimation task,and the deep semantic features are connected to the shallow and finer features through skip connections to achieve the U-Res Net network structure.Subsequently,in order to cope with the problem of small and dense data set targets and large scale disparity,an attention mechanism is introduced,and the network structure of U-Res Net+CBAM is finally realized.Second,a rotating target detection model based on corner point localization is implemented.Inspired by the Corner Net model,the rotation detection frame is firstly defined in the form of a pair of corner points and corresponding boundary vectors,and the features are extracted by a key point estimation network and an adapted Oriented Corner Pooling module is designed.The model shows good results on both the natural image text detection dataset ICDAR and the remote sensing target detection dataset DOTA.