Detection And Fine-grained Recognition Of Inshore Ships On Optical Remote Sensing Images

Detection of inshore ships on optical remote sensing images has an important significance in ocean security and military reconnaissance.Compared with ships' location in remote ocean,the inshore ships are harder to detect due to its' large scale range,various aspect ratios,side-by-side mooring behavior and more complex background.And the fine-grained recognition of inshore ships is also quite helpful to port dynamic monitoring and military strike against warships hidden in civil ships,but it's still very challenging to achieve accurate fine-grained recognition because of the similarity in shape,color and texture between different categories.Although the deep learning based object detection methods have showed superior performance in suppressing cluttered background and identifying different categories,there is currently only a few related research on inshore ships.This paper deeply explores the deep learning based methods of inshore ships' detection and recognition on high-resolution remote sensing images.The main work is as follows:Firstly,we systematically introduce the background and significance of the research in this paper.Then the deep learning based object detection methods are organized from the perspective of strategy in discretizing the search space,the main key difficulties involved are also summarized in detail.Secondly,we propose an instance segmentation based inshore ships detection method.The method firstly predicts both ship regions and ship centers by the framework of semantic segmentation,then separates multiple adjacent inshore ship instances using the center points of ships.The results show that the proposed method not only can achieve pixel-level understanding of inshore ships with promising location accuracy,but also have an advantage in detecting parallel adjacent inshore ships.Thirdly,to better solve the problem of fine-grained inshore ships recognition,we propose a fine-grained inshore ships recognition framework assisted with Generative Adversarial Networks.In order to avoid the adverse effect of co-optimization with location task,we firstly introduce an independent cascaded classification sub-network.Then the generator network is trained to introduce synthetic samples for assisting the exploration of manifold distribution in sample space,thus the fine-grained discriminating power of the classification sub-network will be enhanced.The results show that the method can effectively improve the fine-grained classification accuracy of inshore ships.Finally,since image segmentation based ship detection methods need tedious samples annotation and are also unable to be trained end-to-end,an inclined bounding boxes based end-to-end ships detection and recognition framework is proposed.The method firstly use the proposed anchors angle density strategy to recall more smaller scale ship objects.Then an improved rotation position sensitive region of interest align pooling module is also proposed to extract more accurate local ship features.And to enhance the feature discriminating power of ship regions for fine-grained classification,an attention mechanism based global and local region feature fusion method is adopted.Moreover,a transfer learning based weights remap method is also used to further improve the performance of the model.The results show that the proposed method has promising performance for inshore ships detection and recognition.