Aerial Object Detection And Recognition Based On Fusion Mechanism

Along with t.he rapid deve.lopment of' irma.ging equipment and aerial photography technology,a large number of high-resolution aerial images have been captured.A new detection task is proposed,which is how to efficiently and quickly detect and rec.ognize specific targets in aerial images.The recognition and detection of aerial targets play a vital role in military and civilian applications.However,the characteristics of aerial images,such as large image size,numerous small ta.rgets,diversified target scales,complex backgrounds,extreme aspect ratio and tight target arrangement,make aerial targets difficult to be detected.The deep model based detection algorithms which have excellent pe.rf'ormance on general targets can not work well on aerial targets.In this thesis,by collecting and constructing a variety of aerial image datasets,we conduc.t in-depth rese.arch on the detection of aerial targets and ships.YWe improve the detection model specifically and fuse the multi-cue information of the images.The contents and main contributions of the research are as follows:Firstly,a V:Multi-Cue Fusion based Feature Pyramid VNetwork(VMCF-'-FPN)for aerial object detection and recognition is proposed.For the problem of numerous small targets and diversified target scale.s in aerial iImages,we make full use of the multi-scale feature cues of images,and introduce the feature pyra,mid network to detect the targets of' dif'f'erent scales in the multi-category aerial t.arget dataset DOTA.In order to recduce the impact of complex backgrounds,the semantic information of each feature.layer is enriched through bottle.neck module and deformable convolutiorn module.VMoreover,for the problem of tight target arrangement,a category-sensitive soft-NtMS algorit:hm is proposed.Inspir ed by the idea of image pyramid,we fuse the multi-scale informat.ion of the image by multi-scale testing to further improve the detection accura.cy of' each category.The experimental results on DOTA dataset show that the VMCF-FPN can effectively detect various target.s in DOTA,which is comparable wit.h the current algorithms imp.roved based on Feature Pyramid Network.Secondly,Multi-scale Receptive Field based R-CNN(MRF R-CNN)for aerial object det.ection and recognition is proposed.In order to improve the dete.ction speed of deep model based object detection a.lgorithms on the aerial t.argets,the Light-Head R-CNN network is used as the basic net.work.For the problem of diversified aerial target scale.s,t.he feature maps wit.h different receptive fields are designed and merged as feature cues in the Light-Head R-CNN to improve the scale invariance of the ne t.work.Besides,the multi-scale information of the image is used as a cue to further improve the detection accuracy of different categories in DOTA.The experimenta.l results on DOT A dataset show that,compa.red with the current R-FC\N based algorithms,the proposed algori.thm ha.s the higher detection a.ccuracy,especially for small targets.In the meanwhile,the detection speed of MRF R-CNN is also faster.The MRF R-CNN is generalized on t.he Haggle data.set that c.ointains a large rnumber of small targets(ships),which further valida.t.es the robustness of the model.Thircdly,a DetNet59 embedcded wi t.h feature a.ttention for ship detection is proposed.In order to quickly and e.ff ectively detect warships and c:ivilian ships on the sea,we t.ake the HR SC data.set as the main research cdataset and use DetNe.t.59 network as the basic networ k.By introducing the f'eature att.e.ntion Imechanism in DetNet59,f'eature cues are specifically selectecd and fused.In adcdition,we replace the Rol Pooling Layer with t.he RoI Align Layer to further improve the detection accuracy of t.he model on the ships.The experimental result.s show t.hat.,due to the lightweight structure.,the model processes images in real-time at 50 frames per second.