Research On Ship Recognition Based On Deep Learning

Ships are important military and civilian carriers at sea,and the rapid and accurate identification of ship targets has important military and civilian significance.With the development of optical remote sensing technology,the amount of visible ship image data is also increasing.At present,a series of researches on ship target recognition have been carried out,but there are still some key issues to be solved:(1)Some types of ship images in optical remote sensing ship images are difficult to obtain,resulting in imbalance of these types in ship image dataset.Existing classifiers may obtain higher overall recognition rates,but ignore the recognition rate of minority categories.(2)The optical remote sensing ship image data grows exponentially.It is extremely difficult to label all the images manually for the learning of the classifier.However,it is difficult for the existing classifier to use a small number of labeled samples for effective classification.(3)Classification of existing ship classification and recognition is relatively coarse-grained.Generally,it only involves the classification and recognition of several major categories,such as aircraft carriers,frigates,destroyers,containers,cargo ships and oil tankers.There is very few research of fine-grained ship classification,which is not conducive to the accurate recognition of ship targets.(4)Ship number identification can further provide the identity information of the ship.However,the identification of the ship number is often difficult to identify due to factors such as background,illumination and angle.At present,the rapid development of deep learning technology provides a new way to solve the above problems.This dissertation applies deep learning technology to the key issues of ship target recognition.The main contents of this dissertation include:(1)Aiming at the problem of lacking of some special ship high-resolution remote sensing images lead to the imbalance of data sets,affecting the precise identification of a few types of ship targets.The process of probability calculation and backward propagation learning of leaf nodes in deep neural decision forest is studied.A weighted deep neural decision forest method is proposed.In the data processing stage,the method preprocesses the unbalanced data sets to obtain multiple balanced data sets.The idea of class weighting is introduced in the leaf node prediction results,and the deep neural decision-making forest method is extended to achieve good classification results on imbalanced ship image dataset.(2)Aiming at the problem of lacking labels of optical remote sensing ship images in the training and classification of deep belief network.The training process and classification mechanism of deep belief network are studied,and an active deep belief network method based on Bv SB(Best-versus-Second Best)mechanism is proposed.The samples classification probabilities are obtained from the unsupervised learning process of the deep belief network.A small number of high-value samples are selected by the Bv SB mechanism for labeling,and then the deep belief network is supervised and trained to effectively reduce the demand for labeled samples in the deep belief network.At the same time,the classification performance of the deep belief network on the ship's target is maintained.(3)Aiming at the problem of the coarse-grained classification of ships in the current research can not fully meet the need of accurate recognition of ship targets.The different levels/depths convolutional neural network features are studied,and a fine-grained classification method based on multi-feature fusion is proposed.Two types of convolutional neural network are used to extract global high-level features and fine features of ship images.Class activation map is used to identify the regions in ship images which are meaningful for classification and obtain the fine features of ship targets.In the whole training process,no bouding box and part locations are needed.Then,multi-feature fusion method is used to discover the correlation among these features,which improves the fine-grained recognition rate of ship targets.(4)Aiming at the recognition of ship number in natural scenes,an end-to-end network based ship number recognition method is designed.In the stage of ship number detection,the rotation region proposal network is used to generate rotated proposal with ship number orientation and angle information.In the stage of ship number recognition,a bidirectional Long-Short Term Memory Network is designed for ship number recognition.And the task of ship number detection and ship number recognition is unified into a network for end-to-end training,which can improve the performance of both tasks at the same time and obtain higher ship identification and detection rate.Aiming at the four key issues of ship target recognition proposed in this dissertation,algorithms and models based on deep learning are proposed,and theoretical analysis are carried out.Experiments are carried out on ship image datasets and other public image datasets to verify the proposed methods and models.It shows that the proposed method and model have higher application value for ship target recognition based on visible image,and have reference significance for other image recognition.