Research On Dcnn Based Ship Detection Technology In High-Resoluton Remote Sensing Image

With the development of human society,human activities in the ocean are becoming more and more frequent.Thus,it is becoming more important to monitor and control human activities in the ocean.As a critical node of human-ocean contact,ship targets play an important role in both military and civilian fields,especially in the fields of marine security,maritime traffic,border control,fisheries management and maritime transport.The positioning and behavior information of ship targets constitute an important cornerstone of consciousness in the marine field.In recent years,deep convolutional neural network(DCNN)as a new technology is widely used in remote sensing image processing,especially in remote sensing image target detection tasks.Based on the actual requirements of the target detection mission in remote sensing image,this thesis makes in-depth research and experiments on the problems such as common environmental interference and poor detection performance of different ship targets caused by factors such as incomplete samples in the process of target detection.A framework is proposed in this thesis to improve the ship detection performance,which contains an application system to test the validity of the ship target detection model and improves the iterative efficiency,named as broad area target search systems.In addition,in terms of detection model,some experiments and improvements of the algorithm model are executed to solve the problem of weakening of ship target characteristics in remote sensing images caused by interference factors.Furthermore,the detection of relevant information such as ship target category and direction are inserted in model to improve the detection accuracy.Finally,in this framework,we propose a ship target detection model called Saliency-Faster R-CNN,which is used to solve the problem of poor detection of large-scale vessel target due to the imbalance of multiscale vessel target samples,and to alleviate the leakage and frame misconduct caused by the insistency of large-scale vessel target and training sample scale in the actual scene.