基于卷积神经网络的联合彩色图像和超分辨率的深度图

发布时间：2021-10-31 23:54

　　提高图像分辨率是当前数字图像处理领域的研究热点之一。超分辨率（SR）方法是一组信号处理算法,它允许从同一场景的单个或多个低分辨率（LR）图像生成高分辨率（HR）图像。不久前,深度神经网络（DNN）被引入到计算机视觉、机器翻译、自然语言处理、语音和音频识别、社会网络分析、生物信息学、医学图像分析和材料检验等领域。卷积神经网络（CNN）也被广泛应用于彩色图像和深度图的超分辨率问题,在相同场景的额外HR或LR彩色图像的引导下,可以从LR深度图中恢复高分辨率的深度图。本文提出了一种通过联合LR深度图和相应的LR强度图像重建HR深度图的算法。为解决图像超分辨率问题,提出了一种多尺度上采样的联合双分支网络（JDBNet）概念。该方法可以显著提高恢复的HR深度图像的质量。网络又细分为两个独立的网络—JDBNet1和JDBNet2。JDBNet1和JDBNet2的主要区别在于,JDBNet2有两个均方误差损失函数作为强度Y分支的最后一个输出层和深度映射D分支的最后一个输出层。这进而使得JDBNet2的性能优于JDBNet1。同一场景的低分辨率强度图像和低分辨率深度图是训练网络的输入数据。系统输出数据为... 

【文章来源】：哈尔滨工业大学黑龙江省 211工程院校 985工程院校

【文章页数】：96 页

【学位级别】：硕士

【文章目录】：
摘要
Abstract
Abbreviations
Chapter1 Introduction
    1.1 Source of the Project
    1.2 Research Background and Practical Significance of the Research
    1.3 Problem Statement
    1.4 Thesis Organization
Chapter2 Analysis of Methods for Solving Super-Resolution Problem
    2.1 General Classification of Super-Resolution Methods
    2.2 Single Image Super-Resolution
        2.2.1 Super-Resolution Methods based on Image Reconstruction Technology
        2.2.2 Super-Resolution Methods based on Machine Learning
    2.3 Multi-Frame Super-Resolution
        2.3.1 Super-Resolution Methods in Frequency Domain
        2.3.2 Super-Resolution Methods in Spatial Domain
    2.4 Research at Home and Abroad
        2.4.1 Research Abroad
        2.4.2 Research at Home
    2.5 Conclusion
Chapter3 Structure and Properties of Artificial Neural Network
    3.1 Artificial Neuron
        3.1.1 Artificial Neuron Model
        3.1.2 Activation Functions
    3.2 Artificial Neural Network
        3.2.1 Single-Layer Artificial Neural Network
        3.2.2 Multilayer Artificial Neural Network
    3.3 Convolutional Neural Network
        3.3.1 Convolutional Neural Network Structure
        3.3.2 Convolutional Neural Network Topology
    3.4 Training an Artificial Neural Networks
        3.4.1 Supervised Learning
        3.4.2 Unsupervised Learning
        3.4.3 The Process of Training Neural Network
    3.5 Loss Function
    3.6 Conclusion
Chapter4 Joint Double Branch Network
    4.1 Architecture of Network
    4.2 Formulation of JDBNet1 and JDBNet2
    4.3 Stages of Network
    4.4 Conclusion
Chapter5 Experimental Results and Analysis
    5.1 Training and Testing Details
        5.1.1 Implementation Tools
        5.1.2 Dataset Preparation
        5.1.3 Parameters of Network
    5.2 Results
    5.3 Anticipated Problems
Conclusions
References
Acknowledgements
Resume



本文编号：3469110