Conv Net presents an efficient tool for SAR data interpretation.In this thesis,we use augmented Open SAR dataset for training and comparing several existing deep learning models for the purpose of ship classification.All the deep learning algorithms reveal the difficulty of classifying between Cargo and Tanker ships in the dataset,Consequently CNN models designed with sparse connections between convolution layers are found to considerably improve the classification performance against the existi... 

【文章來源】：上海交通大學(xué)上海市 211工程院校 985工程院校 教育部直屬院校

【文章頁(yè)數(shù)】：99 頁(yè)

【學(xué)位級(jí)別】：碩士

【文章目錄】：
abstract
Chapter 1 Introduction to Synthetic Aperture Radar Imagery
    1.1 Introduction
        1.1.1 Present Oceanic Surveillance Services
    1.2 SAR Imagery
        1.2.1 Spaceborne SAR
        1.2.2 SAR Polarimetry and Bandwidth
        1.2.3 Wave Scattering
        1.2.4 SAR and Optical Image Comparison
    1.3 Ship Detection using CFAR
        1.3.1 CFAR
    1.4 Sentinel-1
    1.5 Sentinel-1 Image Dataset
        1.5.1 High Resolution Dataset
    1.6 OpenSARShip Dataset
        1.6.1 Lower Resolution Dataset
        1.6.2 Train, Validation, and Test Dataset
    1.7 Thesis Structure
Chapter 2 Convolutional Neural Networks Literature
    2.1 Neural Network
        2.1.1 Neural Network Structure
        2.1.2 Convolutional Layer
        2.1.3 Pooling Layer
        2.1.4 Non-Linearities
        2.1.5 ReLU and Leaky ReLU
        2.1.6 SoftMax Layer
        2.1.7 Loss Function
    2.2 Back Propagation
        2.2.1 Gradient Descent
        2.2.2 Learning Rate
        2.2.3 Optimizer
        2.2.4 Adam Optimizer
    2.3 Regularization and Overfitting
        2.3.1 Dropout
        2.3.2 Batch Normalization
        2.3.3 Transfer Learning
    2.4 Network Architecture
        2.4.1 VGG16
        2.4.2 Residual Nets
    2.5 Existing SAR Classification Algorithms and Limitations
        2.5.1 Limitations of Existing Algorithms
Chapter 3 Setup & Algorithm Evaluation
    3.1 Introduction
    3.2 Setup
    3.3 Dataset Augmentation
    3.4 Confusion Matrix and Evaluation Metrics
    3.5 Fine Tuning ResNet50 and VGG16 for Ship Classification
        3.5.1 ResNet50 for Ship Classification
        3.5.2 VGG16 for Ship Classification
    3.6 A-ConvNets and CNN-MR
        3.6.1 A-ConvNets
        3.6.2 CNN-MR
        3.6.3 Comparison of ResNet50, Fine-Tuned VGG16 and A-ConvNets
    3.7 Detection and Classification Constraints
        3.7.1 YoloV2
        3.7.2 YoloV2 training
        3.7.3 Methodology
Chapter 4 Classification Algorithm Design and Development
    4.1 Initial Architecture
        4.1.1 Initial Model Training
    4.2 Hyperparameter tuning and Model Pipeline
        4.2.1 M1,M2, M3
        4.2.2 Training and Testing M1,M2, M3
    4.3 CiNet
    4.4 Results
Chapter 5 Conclusion and Future Work
    5.1 Summary
    5.2 Conclusion
    5.3 Future Work
Reference
Acknowledgement
List of Publications



本文編號(hào)：3203313