Multi-scene Optical Remote Sensing Ship Image Object Detection Based On Deep Neural Network

Ship detection is of great significance in national defense construction,port management,cargo transportation,maritime rescue and combating illegal ships.Optical remote sensing ship images have become the main source of ship detection due to their strong anti-interference ability,rich ship texture information and adaptability to a wide range of scenes.The optical remote sensing ship detection based on deep neural networks has developed rapidly.In this paper,aiming at the characteristics of ship in optical remote sensing images,a method based on deep neural networks is used for ship detection,mainly including the following:(1)In order to solve the problem of the single ship scene in the public optical remote sensing ship data sets HRSC2016 and DOTA-Ship,this paper focuses on the characteristics of ships in remote sensing images,collects images of ships in various scenes in practical applications,and creates a private multi-scene ship data set SHIP.(2)A ship detection method based on improved Faster R-CNN is proposed.By adding a feature pyramid structure,we can extract richer multi-scale ship features and improve the detection effect of ship multi-scale and small object features;Add the SE module to emphasize the feature map channel related to the ship,suppress non-object channels,and get a feature map highlighting the ship’s features.The experimental results on the three data sets respectively prove that compared with the traditional method,this method is helpful to extract the deep semantic features of the ship,and improve the detection accuracy of the ship’s multi-scale,small object and complex background regions.(3)A ship detection method based on rotating frame is proposed.Extract richer ship features through the dense connection of dense feature pyramids,increasing multiscale feature reuse;Add anchor that conform to the shape of the ship in the RPN,use multi-scale ROI Align to solve the problem of feature misalignment,reduce the redundant area in the detection frame,and make the bounding box fit to indicate the positioning of the ship.The experimental results on the three data sets respectively prove that the method improves the detection effect of dense areas of ships.(4)A ship detection method based on stacked bidirectional RNN with selfattention is proposed.Integrate the ship’s multi-scale features by constructing a crosslevel fusion multi-level feature network;In the rotating proposal network,improve the size of the anchor and the ROI pooling layer,suppress the side effects of the NMS algorithm,and generate a proposal that conforms to the shape of the ship;Add a stacked bidirectional RNN fused with a self-attention to form an NMS re-scoring network,and adjust the confidence of the proposal region to obtain a more accurate proposal region.The comparison and ablation experiments on the three data sets respectively prove that the method is effective in detecting multi-scale,small object,dense and complex background features of ships,and improves the accuracy of ship detection in multiple scenes.