Ship Detection And Recognition In Optical Remote Sensing Images Based On Deep Neural Networks

Since optical remote sensing images are widely used in the military and civilian domain,as one of the research directions in the field of remote sensing,object detection in optical remote sensing images has attracted more and more attentions.Ship is the key transportation and combat tool in water,so conducting accurate and rapid ship detection shares great research significance and practical value.In this paper,ships in high-resolution optical remote sensing images have the characteristics of arbitrary orientations,large aspect ratio,large scale difference and so on.Taking into account the efficiency of practical applications,we have conducted much in-depth research on the basis of general object detectors for natural images.Then,a series of ship detection and recognition methods in optical remote sensing images based on deep neural networks are proposed.The main research contents are as follows:(1)A ship detection and recognition method with rotated bounding boxes based on improved CenterNet is proposed.In view of the difficulty in designing suitable anchors for an Anchorbased detector,we adopt the Anchor-free detector CenterNet based on key-points as the basic detection framework,and introduce deformable convolution to extract better features of geometrically deformed ships.This method needs to learn and predict the rotation angle,therefore,the regression branch for rotated bounding box is added to the detection head.At the same time,L1 loss function is modified to solve the loss discontinuity problem caused by the critical case of rotation angle regression.After that,we add an additional crosscategory NMS for rotated boxes to filter the redundant detection results.Finally,experiments are carried out on the ship detection and recognition dataset,which is collected and labeled by rotated bounding boxes.It is proved that our method is more flexible and effective than the Anchor-based method.(2)A ship detection and recognition method with rotated bounding boxes based on multiscale and specified feature fusion is proposed.To better solve the problem of small and multiscale object detection,we adopt Feature Pyramid Networks(FPN)as the basic architecture to realize ship detection on multi-scale feature maps.Then,a specified feature fusion module is introduced to optimize the fusion of high-level and low-level features.This method can classify the object and regress rotated bounding box by pixel points,which improves the performance of multi-scale object detection through increasing the number of positive samples.At the same time,in order to adapt to more resource limited application scenarios,we improve the model with lightweight Mobile Net V2 to reduce the amount of parameters.Finally,we propose the R-Io U quality evaluation strategy to score the sampling points,which can quickly suppress low quality false detection boxes caused by dense prediction.(3)A ship detection and recognition method with rotated bounding boxes in wide remote sensing images based on sea-land segmentation is proposed.Considering the challenge of ship detection and recognition because of complex land scenes in wide remote sensing images,we introduce a preliminary sea-land segmentation based on U-Net++.Then,the morphological operation is combined to improve the segmentation result and realize the land retraction,which ensures the high detection rate of inshore ships and densely arranged ships.We make full use of the prior knowledge of sea-land segmentation to extract waters where ships may exist as regions of interest before detection,and filter the results heuristically after detection.In this way,the overall speed and accuracy of the algorithm can be improved effectively.