Research On Techniques Of Ship Detection In Optical Remote Sensing Images Based On Convolutional Neural Network

Modern remote sensing technology emerged in the 1960 s and plays an important role in the fields of agricultural production,land resource utilization monitoring,urban planning,and military investigation.After years of development,optical remote sensing satellites have shown a wide-ranging and high-resolution development trend,and the ability of a single satellite to acquire image data has been continuously enhanced.With the rapid development of satellite manufacturing technology,the manufacturing cost of satellites continues to decrease.More and more optical remote sensing satellites are in orbit and provide data services.Every day,terabyte level image data is transmitted to the ground.How to efficiently and automatically extract useful information from massive image data has become the key to the application of remote sensing information.Ships are the main carriers of international cargo transportation and military power projection.The detection of ship targets through optical satellite remote sensing platforms can meet the needs of maritime traffic control,smuggling inspection and maritime rescue in the civilian field;in the military field,it can monitor the deployment and dynamics of the enemy's forces,improve situational awareness,and provide a basis for decision-making.Therefore,how to detect ship targets from massive optical remote sensing images in an accurately and efficiently way has become an urgent problem to be solved.In recent years,convolutional neural networks(CNN)have been widely used in the field of computer vision.CNN is driven by data,and the high-level semantic features are extracted automatically.CNNs have strong feature description and discrimination capabilities.Ships are moving targets,and scenes containing ships are uncertain.In the offshore scene,the water background is monotonous,and the ship targets are widely and sparsely distributed;in the port area,the ground features are complex,and the ships are densely arranged near the wharf.In the balance of accuracy and efficiency,the detection method designed for a single scene is difficult to adapt to different detection scenarios.Therefore,this dissertation focuses on the ship detection task of visible remote sensing image,applies the mainstream CNN technology,designs different ship detection methods for different detection scenarios,in order to improve the accuracy and timeliness of ship detection methods in multiple scenarios,and promote the construction of intelligent remote sensing information platform.The main contents of this dissertation are as follows:1.In the offshore scene,the color and texture characteristics of the water background are isotropic,ships are sparsely distributed and highlighted in the water background.Based on this feature,a ship detection algorithm for offshore areas is designed,which combines the visual saliency model with the classification convolution neural network.First,the visual saliency model is used to quickly detect abnormal regions in the water body on the down-sampled input image,and extract image slices containing candidate targets.Then,all image slices are input to the classification CNN for discrimination,false alarms are removed,and the detection results are output.Aiming at the problem that he direction of the ship is arbitrary,kernel rotation convolutional layer is used to extract features in multiple directions.The orientation invariant features are encoded by the oriented response pooling operation to improve the invariance of the classification network to the rotation transformation.Compared with the end-to-end convolutional neural network,the visual saliency model has a simple calculation process,relatively low hardware requirements,and can quickly locate candidate targets;compared with manually designed features,lightweight classification convolutional network has strong feature expression ability and can accurately distinguish between real targets and false alarms,and the overall algorithm has higher detection accuracy and faster detection speed.2.In the port scene,the background of the features is complex,and the ships are intensively docked.For this scenario,a regression based oriented ship detection network is designed.The network predicts the rotated bounding boxes,which retain the true size of the target.Aiming at the problem that the Io U value between rotated bounding boxes is sensitive to the angle deviation,which makes it difficult to match the ground truth boxes with anchor boxes,the angle-related Io U is used to calculate the Io U between the inclined rectangular boxes,which increases the number of positive samples and further improves the recall rate;anchor refinement branch and receptive field amplification module are designed and embedded in the network to improve the overall detection accuracy and the positioning accuracy.The ablation study of each improved strategy and the comparison experiment with other detection methods verify the effectiveness of the method.3.Detection methods using anchor mechanism need to preset a large number of anchor boxes with different scales,aspect ratios and rotation angles.The hyperparameters of these anchor boxes need to be set manually,and a large number of anchor boxes also increase the computational and memory consumption.To solve this problem,a rotated ship detection network based on central area prediction is designed.This method takes the pixels near the object center point on the feature map as positive samples,and uses the position of the central point and the stride of the feature map as the reference for the bounding box regression.This method discards the anchor mechanism and has a high detection speed.In addition,for the problem of the misalignment between the rotated ship target and the sampling position of the convolution kernel in the detection head,a rotated ROI convolution is designed to refine the target features to achieve feature alignment,which is applied to the central area prediction network to improve the ability of distinguishing foreground targets and further improve the detection accuracy.Overall,this dissertation focuses on the ship detection task in multiple scenes.Some key technologies,including the extraction of candidate targets,discrimination of real ship targets and the performance optimization of detection networks,are studied in depth.Some results are acquired in the study.The relevant achievements of this dissertation can promote the construction of intelligent remote sensing information platform.