發(fā)布時(shí)間：2023-12-02 08:07

　　

【文章頁數(shù)】：77 頁

【學(xué)位級別】：碩士

【文章目錄】：
Abstract
1 Introduction
    1.1 Motivation and Objectives
    1.2 State of the Art
        1.2.1 Aerial Scene Classification
        1.2.2 Deep Convolutional Neural Networks (CNN)
        1.2.3 Visualizing Deep CNN
    1.3 Outline of the Thesis
    1.4 Summary
2 Convolutional Nueral Network
    2.1 Introduction to Network Composition
        2.1.1 Convolutional Layer
        2.1.2 Pooling Layer
        2.1.3 Fully-connected Layer
        2.1.4 Activation Function
    2.2 Architectures of Deep CNNs
        2.2.1 AlexNet
        2.2.2 VGGNet
        2.2.3 GoogLeNet
        2.2.4 ResNet
    2.3 Summary
3 Review of CNN Visualization Strategies
    3.1 Deconvolutional Network (Deconvnet)
    3.2 Guided backpropagation
    3.3 Class Activation Map(CAM)
    3.4 Summary
4 Methodology of L1-regularized CAM
    4.1 Ll-regularized CAM(L1-CAM)
    4.2 L1-CAM of Intermediate Layers(inter-L1-CAM)
    4.3 Summary
5 Experiment and Result Analysis
    5.1 Datasets of Aerial Scenes
        5.1.1 UC-Merced Land Use Dataset
        5.1.2 Aerial Image Dataset
    5.2 Evaluation of Ll-regularized CAM
        5.2.1 Experimental Setup
        5.2.2 Results on UCM and AID datasets
    5.3 Evaluation of inter-Ll-CAM
        5.3.1 Experimental Setup
        5.3.2 Results on UCM and AID datasets
    5.4 Summary
6 Application
    6.1 Introduction of LAHNet
        6.1.1 Architecture of LAHNet
        6.1.2 Experiment Setup
        6.1.3 Results of LAHNet
    6.2 Weakly Supervised Scene Localization
        6.2.1 Experimental Setup
        6.2.2 Results of Scene Localization
    6.3 Summary
7 Concluding Remarks
    7.1 Discussion and Conclusion
    7.2 Outlook
References
Acknowledgements



本文編號：3869256