Research And Implementation Of Oriented Object Detection Platform Based On Deep Learning

Oriented object detection is an important part of intelligent visual perception systems and has been used in aerial image detection,highprecision map construction,auto-driving street-view recognition,and other tasks.However,as a special branch of object detection technology,there are still many difficulties and challenges in oriented object detection:(1)The existing oriented object representation methods often have representation ambiguity.Consequently,it leads to the difficulty of model convergence and is hard to achieve high-precision detection.(2)The features of inclined and densely arranged oriented objects are difficult to be accurately captured by the common axial-alignment convolution layer.How to solve the misalignment of features is a key problem to be solved urgently.(3)There are few researches on sample selection and label assignment for oriented object detection,and most of the existing methods directly follow the horizontal object detection.These methods are lack of pertinence to oriented objects,which limits the further improvement of detection accuracy.To solve the above problems,this thesis completes the following research contents:(1)An oriented object detection network named Free3Net based on unambiguous representation and feature aggregation is proposed,and an oriented object representation scheme named FreeGliding is designed to solve the problem of representation ambiguity and multi-task dependency.The orientation invariant feature and multikeypoints feature are extracted to help the model learn the ability of object discriminative features.On average,Free3Net is 2.62%higher than the state-of-the-art method.(2)Starting from the problem of outer sample selection and label assignment for oriented object detection,a Loss-aware Outer Sample Selection algorithm LOSS is proposed,and a Near-ness measurement is designed to measure the spatial alignment of oriented objects.Under the assistance of them,efficient sample selection and label assignment can be ensured.Compared to the baseline method,the proposed method can improve the mAP precision by 4.69%.(3)Based on the researches above,this thesis designs and implements an oriented object detection platform which can verify the effectiveness of the methods proposed in this thesis.The platform provides the functions of oriented object detection,assisted data annotation,and data processing in the form of API service,which can promote the development of oriented object detection technology in academic research and industrial applications.