Research On CNN Based Inshore Ship Detection And Recognition In Remote Sensing Images

Inshore ship detection and recognition in remote sensing images is a key technology of maritime surveillance and reconnaissance interpretation.The features of the harbor background are similar to those of ships;densely arranged ship targets are arbitrary-oriented and have a large aspect ratio;the overall shape,color,and texture of different types of ships are similar.Those problems make ship target detection and recognition difficult.To solve the above problems,this paper focuses on inshore ship detection and recognition in remote sensing images based on convolutional neural network.This research has not only significant scientific value for enriching the theory of target detection in the complex background and rotating target detection but also has significant practical value for improving the accuracy of inshore ship detection and recognition.The main contributions of this paper are as follows:Aiming at the lack of a fine-grained labeled ship dataset,Fine-Grained Ships in Aerial Images Dataset(FGSAID)is presented.FGSAID has 1690 images and 5410 ship targets,which are divided into 45 classes,and most of them are military ships.All of the targets have both bounding boxes and rotated bounding boxes information.This dataset provides strong support for the research of the detection of ships in remote sensing images.The features of the harbor background are similar to those of ships,resulting in false alarms.To solve this problem,a docked ship detection method based on multi-stage hard example mining is proposed.In this method,a background suppression module based on multi-receptive field feature fusion is added to the feature extraction network,and the attention mechanism is used to fuse the features of different levels to obtain more distinguishing features of the proposals.In processing the proposals,a multi-stage detection network with a complex background example mining strategy is proposed.By using the output results of the former detector to resample the input of the later detector,the background and ship samples that are easy to misclassify are mined,which promotes the network to learn to distinguish these samples and improves the detection accuracy.Extensive experiments show that the proposed methods improve the detection performance of docked ships.Densely arranged ship targets are arbitrary-oriented and have a large aspect ratio,and locating accuracy is influenced by angle regression greatly.To solve this problem,an anchor-free rotation ship detector based on axis detection is proposed.The axis bounding rectangle is designed to detect the axis and get the target direction,which locates targets with a large aspect ratio more accurately compared with predicting angle directly.It also avoids boundary problems caused by predicting the axis endpoint directly.Orientation center-ness is designed to screen positive samples,which is more suitable for targets with large aspect ratios and uncertain directions and can suppress the low-quality results generated by the feature points away from the target more effectively.Experiments show that the proposed methods improve detection accuracy for ships in the harbor while maintaining high detection speed.The overall shape,color,and texture of different ship types are similar,resulting in the difficulty of fine-grained ship detection and recognition.To solve this problem,a ship fine-grained detection and recognition method based on key sub-region features is proposed.The recognition network extracts the overall features and sub-region features from the proposals generated by the detection network.The weight of the sub-region is calculated based on the discriminant significance and the key sub-region features are fused with the overall features,which helps to improve the accuracy of fine-grained recognition of ship targets.Meanwhile,the angle relation loss function is designed to train the detection network,which better reflects the influence of angle on locating,thus improving the locating accuracy.Experiments show that the proposed methods improve the performance of fine-grained ship detection and recognition in remote sensing images.