Research On Ship Detection Technology In Marine Optical Remote Sensing Images

Ship detection on high-resolution maritime optical remote sensing images is an important requirement in marine comprehensive law enforcement,maritime transportation and other fields.However,in ship detection task,there mainly exist the following problems:(1)Lack of specific data;(2)The appearance of the ship is diverse,and is similar to some backgrounds,causing mistakes;(3)The particular location distribution of ships requires higher accuracy for positioning.Further,the refined management of ocean region needs to obtain ship type,but the following problems still restrict the improvement of the model’s fine-grained classification capabilities:(1)Due to unfavorable imaging conditions and small size of ships,image details are of inferior quality;(2)The appearance of samples belonging to the same fine-grained category varies greatly;(3)The number of samples in some categories is small,which easily leads to overfitting.In summary,the main work and research contents of this article are as follows:Firstly,owing to lacking samples,we independently collected and constructed a ship detection dataset based on optical remote sensing images of harbor scenes.The images are collected from Google Earth in many countries around the world,with a resolution of about 1 meter and containing nearly 4,000 ships.The samples are complex and diverse,with large changes in appearance,and large differences in illumination and contrast.In certain harbor scenes,slender ships are often distributed closely,so the arbitrary quadrilateral labeling box is adopted.Although manual labeling is more expensive,it makes the sample positioning contour more accurate.Secondly,a remote sensing image ship detection method is constructed.First,a Channel Weighting Mechanism is devised,which self-learns the weights of different channel to enhance valid features in channel dimension.Second,a Context Enhancement Structure is designed to enhance features in spatial dimension,which extracts more information of ship appearance and excavates potential contextual information.Finally,to promote the detection ability of ships of various scales,a Multi-stage Weighted Fusion Pyramid is applied to optimize the fusion of high-stage features and low-stage features.Experiments show the effectiveness of the above methods in ship detection.Thirdly,an end-to-end remote sensing ship detection and fine-grained classification framework is constructed.First,using the proposed SR-ShuffleMix data augmentation method combined with the meta-learning optimization target,a fine-grained classification model is established,which can effectively focus on the subtle local features,suppress the overfitting phenomenon,and improve the classification performance.Next,use the classification model to build a fine-grained detector.Finally,in order to verify the effectiveness of the algorithm,a synthetic dataset for fine-grained detection of marine ships is established.A number of ship templates were manually obtained,then randomly pasted to the sea background.Meanwhile,annotations with fine-grained classes and positions are generated.Experiments show that the above methods improve the classification ability of the fine-grained classification and detection framework.