Research On Ship Recognition Algorithm In Optical Remote Sensing Image

Since the twenty-first century,the ocean has received more and more attention from all countries.Whether it is the development and utilization of marine resources or the protection of territorial sea from infringement,ships are indispensable participants in maritime missions.Therefore,how to achieve fast and accurate positioning and recognition of ships is an urgent need for the development of the times.It is a potential and practical research method that using deep learning technology for ship recognition in remote sensing images.Ship recognition tasks not only need to achieve the usual ship detection,but also need to achieve fine-grained classification of ship types.Therefore,ship recognition includes positioning and classification of ship types.The positioning of ships requires the network to extract and use the feature information.In the deep learning algorithm,feature pyramid network is a key module to improve the utilization of feature information.However,the traditional feature pyramid still has the problem of insufficient utilization of feature information.The classification of ship types is that the classifier predicts the object category according to the probability of each category.The realization of proper competition in the fine-grained model classification is the key to ensure the classification accuracy.However,the ship type data set presents the characteristics of long tail distribution,which is not conducive to the deep learning network to balance the weights of various categories.In view of the above problems,this paper conducts the following research to improve the detection ability and fine-grained classification ability of ship recognition algorithm.To improve the problem of insufficient utilization of feature information in traditional feature pyramid,a feature enhancement architecture based on attention mechanism is proposed.On the one hand,the attention mechanism is added to the top-level feature layer to ensure that sufficient top-level feature information is retained,providing rich context information for ship recognition;on the other hand,through the adaptive enhancement of ROI feature information,it is ensured that each ROI adaptively fuses the feature information of the four feature layers and is no longer limited by the size of ROI.In order to improve the problem of data imbalance and head category and tail category unfair competition in long tail data,a long tail balance framework based on instance quantity mechanism is proposed.On the one hand,from the data level,dynamic augmentation is designed.More data augmentation is applied to the tail category to ensure that the network improves the learning of the tail category features and increases the importance of the tail category.On the other hand,from the algorithm level,multi-branch classifier is designed to achieve the category instance quantity gap in the same classification branch is not big.There is fair comprtiton between categories.The effectiveness of this method is verified by multiple sets of experiments.This paper realizes more use of feature information in ship recognition algorithm based on deep learning,and proposes a balanced framework for fine-grained recognition of long-tail ships.The proposed two frameworks can be simply plug and play for the two-stage object detection network.This paper lays the foundation for the subsequent ship recognition based on deep learning.