Oriented Object Detection Of Remote Sensing Image Based On Deep Learning

Currently,oriented object detection is one of the key technologies in remote sensing image processing,it is an effective method for humans to utilize space information in various fields such as military defense,traffic planning,and daily life.Oriented object detection aims to analyze remote sensing images,then locate and classify objects in the images.However,objects in remote sensing images generally have orientation,and existing representation methods are difficult to locate objects and extract object features,which hinder the development of oriented object detection applications in remote sensing images.To address these problems,this thesis studies oriented object detection methods based on deep learning.The main work of this thesis includes the following aspects:(1)The current representation methods for oriented objects are rotation and quadrilateral representations.The subtle angle deviation has a significant impact on the performance of oriented object detection in rotation representation;meanwhile,the quadrilateral representation poses a challenge for generating oriented proposals directly in two-stage detectors.To this end,this thesis proposes an oriented object detection method based on short-side excursion(SSEDet).Firstly,we use the corollary of circular angle theorem and the set relationship of rectangles to convert horizontal rectangles to inscribed oriented rectangles.Then,we design a short-side matching rule to map oriented rectangles and their circumscribed rectangles one-to-one,aligning oriented rectangles and representation parameters.Finally,we apply the short-side excursion representation in two-stage methods to generate high-quality oriented proposals directly,and obtain classification and regression through rotation feature alignment.(2)To address the problem of inconsistency between spatial information and features of oriented objects,this thesis proposes an oriented object detection method based on oriented feature refinement network(OFR-Net),which mainly consists of two parts: oriented feature alignment module(OFAM)and oriented feature selection module(OFSM).Specifically,firstly,OFAM takes the original features extracted by the backbone network as input and generates oriented prior boxes with prior knowledge.Secondly,based on the oriented prior boxes,the aligned convolution adjusts the position of the convolution kernel to obtain aligned features that are consistent with the spatial position.Then,OFSM selects aligned features to enrich the spatial information and obtain oriented features.Finally,we use the extracted oriented features with two-stage methods to perform detection tasks and obtain the final prediction results.We have validated the proposed methods on DOTA,HRSC2016 and DIOR-R datasets.The experiments show: compared to the existing one-stage and two-stage methods,the SSEDet and OFR-Net proposed in this thesis have a superior detection accuracy on oriented object detection tasks in remote sensing images.The thesis contains a total of 33 figures,8 tables,and 93 references.