Research On Inshore Ship Detection And Recognition Based On Deep Learning

Ship target play an important role in inland and ocean transportation.Accurate and efficient interpretation of the location and type of ships has extremely high application value and research significance in civil(such as ocean monitoring,fishery management)and military(such as intelligence investigation,long-range strikes)fields.Optical images have the characteristics of convenient data acquisition and high data fidelity.Under proper weather conditions,it can fully capture the colors,shapes and structures of the targets and have become the main data source for the intelligent interpretation of ship targets.Compared with interpretation of ship targets in optical remote sensing images,detection and recognition of inshore ship targets in sequence images(such as surveillance video)can dynamically monitor high-value ships in local areas such as ports and rivers,which further promote the utilization rate of the channels and the management efficiency of the ports in the civil field.Also in the military field,it can provide more abundant and diversified intelligence information for battlefield situation awareness and combat decisions.With the rapid development of deep learning,Convolutional Neural Network(CNN)based intelligent algorithms shown excellent performance in many ship target detection and recognition tasks from remote sensing images.However,due to the particularity of the imaging angle of the monitor,the multi-scale characteristics of inshore ship targets in the surveillance video image are more obvious.Also,the geometric ratios between length and width of the ships are significantly different,and the positional relationship between the targets is more complex(such as occlusion,truncation,crossover,etc.)In addition,complex and diverse backgrounds also can cause great interference to algorithm performance.In order to further improve the algorithms’ intelligent interpretation of inshore ship targets,aiming at the aforementioned problems,the thesis considers the geometric and structural properties of inshore ship targets and the designing of network’s structures tightly and the following work is carried out.1)Aiming at the significant difference in the geometric aspect ratios of inshore ships,the low matches between traditional convolutions and targets as well as the weak representation ability of algorithms for ship with large aspect ratios,an Attention Scaleaware Deformable Network(ASDN)is proposed.The representation abilities of our method for inshore ships with large aspect ratios are improved further by designing an Attention Scale-aware Module(ASM)and introducing Deformable Convolutional Network(DCN).Specifically,The ASM,which is composed of a multi-branch asymmetric convolution module and an attention mechanism,could extract non-local features of inshore ships.The introduced DCN could adjust the locations of convolution adaptively based on distributions of characteristics of ships,which improve the representation abilities for ship targets with unconventional sizes resulting from occlusion,truncation,etc.2)Aiming at the significant multi-scale of inshore ships,the "similarities between classes and differences within classes" of different types of ships,the limit feature extraction abilities of traditional feature pyramid for local and global information of multi-scale ship targets,a Feature Refocusing Network(FRN)is proposed.The key component is Feature Re-focusing Strategy(FRS),which consists of a Multi-level Feature Aggregation Module(MFAM)and an Attention Feature Reconstruction Module(AFRM).Here,The MFAM could fuse multi-dimensional features fully and the AFRM could enhance the representation abilities of our method for significant features of inshore ships,which promotes the detection and recognition performance of our method.3)Aiming at the highly similarities between specific structures,parts of inshore ships and complex and diversified backgrounds,the inaccurate extraction for inshore ship targets,the limit location and recognition abilities of algorithms,an algorithm namely Anchor-guided Attention Refinement Network is proposed.The key components are Attention Feature Filter Module(AFFM)and Anchor-guided Alignment Detection Module(AADM).In AFFM,the multi-scale features of inshore ships are fused and the significant features at low-level feature maps are enhanced by using high-level semantic information.And the refined anchors and the features corresponding to object detection and recognition tasks are aligned,which improves the consistency of features as well as the performance of our method.