Research On Detection And Classification Of Ship Targets In SAR Images

Synthetic Aperture Radar(SAR),thanks to its noticeable advantages of working in all-day and various weather conditions and observing the broad areas,has become an irreplaceable device for continuous,real time and long term monitoring over wide sea areas.With the advent and application of various advanced SAR systems,the available SAR data set,especially the satellite SAR data,has been increasing rapidly.Therefore,how to amply interpret the SAR data from the littoral and open sea areas is a core research issue for the SAR remote sensing community.This dissertation strives to make an in-depth research on the detection of weak ship targets in various sea conditions from Polarimetric SAR(Pol SAR)images and the classification/recognition of different ship types in Medium Resolution(MR)SAR images,so as to develop the effective and operational automatic recognition systems of ship targets in SAR imagery.The main contents of the dissertation will be summarized as follows:Exploring the scattering dissimilarity between the ship targets and sea clutters in Pol SAR images,a novel feature extraction method for Pol SAR image is developed and an adaptive Pol SAR ship detection algorithm is constructed.Generally speaking,there are usually small ship targets with weak backscattering in complex sea background and the complex sea clutters often have strong backscattering,therefore the targets in the acquired SAR images present small Signal-to-Clutter Ratio(SCR).Moreover,the weak targets are not so discriminative from the sea clutters and the traditional energy-based detection methods are no longer effective.This dissertation proposes to measure the difference of the double-bounce scattering and volume scattering between the ship targets and sea clutters in a local area,and thus the Local Scattering Mechanism Difference based on Regression Kernel(LSMDRK)feature for Pol SAR image is derived.Combining the LSMDRK and a self-resemblance based saliency detection method,a saliency map is derived in which the contrast between the ships and the sea clutters is further improved,and then the ship targets are detected via an adaptive threshold segmentation method.The experimental results using the real satellite Pol SAR images show that the proposed method can acquire a better detection result on weak targets with a low false alarm rate.To solve the problem that the traditional pixel-level detection method is severely impacted by the speckle noise and is sensitive to the complex sea conditions,three regional dissimilarity measurements are proposed,and then a novel automatic Pol SAR ship detection method is developed in a supervised way.It is first to perform superpixel segmentation on the Pol SAR image,and then the statistical dissimilarity among the coherency matrices in different superpixels,the distance dissimilarity on the Riemannian manifold of the Equivalent Superpixel Coherency(ESC)matrices and the scattering power distribution difference of different ESC matrices under each scattering mechanism are employed to measure the difference between the ship targets and the sea clutters.The three proposed superpixel-level dissimilarity measurements can fully improve the ship-sea contrast in different backgrounds and are nearly free from the backscattering intensity.By combining the dissimilarity measurements of the multiscale superpixel segmentation,the robustness of the dissimilarity measurements are improved for different ship types and various sea clutters with different backgrounds.Finally,the kernel Fisher discriminant analysis and the linear SVM classifier are used to implement the automatic Pol SAR ship detection.The experiments on the synthetic and real RADARSAT-2 data validate the effectiveness and robustness of the proposed detection method for different SCRs.To solve the problem that the traditional feature extraction is less powerful to perform ship classification in the medium and low resolution SAR images,this dissertation explores the Deep Convolutional Neural Networks(DCNNs)for the merchant ship classification in MR SAR images.And a novel densely connected CNN architecture is proposed to strengthen the feature extraction in MR SAR images.Compared to the classical CNN models,the newly proposed densely connected CNN architecture has a high classification accuracy of the tanker,container ship and bulk carrier for the ship classification data set collected from the MR Sentinel-1 SAR images.The proposed densely connected CNN model can generalize robustly to classify the tanker,container ship,bulk carrier,cargo ships and general cargo ships and behaves well when performing the classification task with less training data.To mitigate the large within-class diversity and between-class similarity for MR SAR ship classification,a Multi-Task Learning(MTL)classification framework is proposed.The MTL framework jointly optimizes the softmax classification log-loss and the triplet loss derived from the Deep Metric Learning(DML)scheme such that the deep representations of the same class are pulled much closer to each other and those from different classes are pushed as farther apart as possible,and thus the ship classification performance is improved.The DML method utilizes the densely connected CNN model as the backbone network of the triplet networks,and then the triplet images are processed to extract deep embeddings for triplet loss computation.However,the normal DML utilizes the triplets in a training batch independently which will lead to slow convergence and overfitting.Therefore,a novel regularization term inspired by the Fisher discrimination is proposed to explore the global information of the holistic triplets in a training batch such that the training stability and generalization of the network are improved.The experimental results using the MR SAR ship data set demonstrate that the MTL classification framework has better classification performance compared with the state-of-the-art CNN models and can generalize well for MR SAR ship classification with more types.