Ship Detection Of SAR Image Based On Deep Learning

For the ship detection problem in SAR image data, due to the advantages of deep learning model, we propose a ship detection method in SAR images based on deep learning. Deep learning is seen as a feature-learning tool, and it can automatically learn the feature to express objectives from the image data. It learns multiple nonlinear relationship by building a learning model with multiple hidden layers, which other non-depth learning algorithms cannot reach. So they can learn more useful features from ship training data, and enhance the accuracy of ship target classification.This paper analyzes the deep learning of the basic models and methods, and has made the experiment in the related image data sets. Firstly, we study the Restricted Boltzmann Machine(RBM) in SAR Ship Target Detection and introduce the theory knowledge of RBM structure. Then combine such theory with ship detection problem, we present the implementation of specific algorithm. In view of the matter of too much training parameters and long training time, we proposed that based on Convolutional Neural Network(CNN) for SAR images of ship detection. CNN network structure is formed by convolution alternating layer and the lower sampling, through the local receptive field and the weights of Shared characteristics, not only solved the previous problem, also greatly improve the detection performance.Compared with the traditional shallow learning, the different of deep learning is that:(1) emphasizes the depth of the model structure, characterized by containing multiple hidden layer;(2)Highlight the superiority of the learning characteristics, through the characteristics of the original space transform step by step to a new space, facilitates the classification tests.Compared with the manual to extract features, use big data to study characteristics, can describe the nature of the data information more.Deep learning advantage for ship model expression ability strong, can highlight the ship targets in around the target feature, at the same time, this method also has certain biological basis.Based on the existing test data sets, prove the effectiveness of the proposed method.