| Object detection aims to locate and classify regions of interest in images and is widely used in many fields like autonomous driving,intelligent security,and auxiliary medical care.In situations with a regular object distribution,existing general object detection methods that use horizontal bounding boxes to locate objects already have high detection accuracy.However,the horizontal bounding box cannot effectively characterize the arbitrary-oriented object in scenes like aerial images,remote sensing images,and scene texts,which leads to poor performance of general object detection.To this end,researchers conduct research on rotated object detection,using oriented bounding boxes with angle information to locate the objects,which is of great significance in solving the problem of multi-directional object detection in practical scenes.At the same time,rotated object detection is a challenging task with the following problems:(1)The standard convolution of rules struggles to characterize the geometric transformation of the object,and the simple method of feature fusion struggles to achieve the full interaction of semantic and detailed information.(2)The generation strategies of rotated proposals based on dense rotated anchors and horizontal proposals are redundant.(3)The stacking of a large number of techniques dedicated to accuracy improvement considerably slows down inference speed.To address the above issues,this thesis performs the following work:(1)In terms of feature extraction,aiming at the problem of insufficient feature representation ability,this thesis proposes a multi-scale object representation method based on the bi-directional feature pyramid.Specifically,this thesis introduces deformable convolution to broaden the receptive field and adaptively adapts to the geometric transformation of the object.It also proposes a convolutional attention module that combines channel and spatial attention,and then develops a multi-level feature bidirectional fusion strategy based on this module,thus achieve full interaction of semantic and detailed information and extraction of more discriminative multi-scale feature extraction.Compared with baseline,the proposed method gains 4.62% m AP on the BBox-SSDD dataset.(2)In terms of head design,aiming at the problem of large redundancy in existing generation strategies of rotated proposals,this thesis proposes the solution of learning rotated proposals from horizontal anchors.A fully convolutional rotated region proposal network is constructed to generate high-quality rotated proposals directly.Additionally,a cascade prediction head is built in a progressive learning manner to improve the precision of detection step by step.Compared with baseline,the proposed method gains 5.30%m AP on the DOTA dataset.(3)In terms of precision and speed balancing,aiming at the problem that strategy stacking decreases inference speed,this thesis proposes the solution of abandoning the refinement process of bounding boxes.A single-stage rotated detection model is built,and the elliptic energy loss based on curve similarity measurement is designed.This novel loss accurately fits Skew Io U,which not only maintains inference speed but also improves detection accuracy.Compared with the multi-stage model on the DOTA dataset,it maintains an accuracy of 91.06% and improves the FPS from 7.8 to 12.0. |
