泛研网

公众版 查看说明
全球科研项目数据库2026版

无监督层次表征学习模型及其在遥感影像解译中的应用

项目来源

国家自然科学基金(NSFC)

项目主持人

刘芳

项目受资助机构

西安电子科技大学

项目编号

61871310

立项年度

2018

立项时间

未公开

研究期限

未知 / 未知

项目级别

国家级

受资助金额

16.00万元

学科

信息科学-电子学与信息系统-信息获取与处理

学科代码

F-F01-F0113

基金类别

面上项目

无监督学习 ; SAR影像无监督分割 ; 极不匀质区域地物表示 ; 方向结构滤波器 ; 层次表征学习 ; unsupervised learning ; hierarchical representation learning ; unsupervised Segmentation of SAR Image ; representation of ground objects in extremely uneven regions ; directional structure filter

参与者

李婉；张雅科；崔元浩；黄欣研；骞晓雪；杨苗苗；李鹏芳；庞婷尹；李硕

参与机构

陕西科技大学

项目标书摘要：大场景高分辨SAR影像在语义空间中的区域图的作用下，被划分为混合集聚结构地物像素子空间、结构地物像素子空间和匀质地物像素子空间。针对大场景高分辨率SAR影像存在获取类别标签样本数据难的问题，通过对三个不同结构地物的像素子空间的无监督分割来解决SAR影像的无监督分割。由于极不匀质区域地物表示的复杂性与该区域目标大小和形状的多样性、相邻目标空间拓扑结构的多样性、目标散射特性和相干斑噪声等有密切的关系，与匀质区域和不匀质区域地物表示的复杂性相比，极不匀质区域地物表示的难度更大，导致混合集聚结构地物像素子空间的无监督分割更具有挑战性；因此，本项目拟重点研究极不匀质区域的无监督层次表征学习模型和方法来实现混合集聚结构地物像素子空间的无监督分割，为无监督表征学习探索出一条非深层且可解释的新的研究途径。

Application Abstract: Under the application of the regional map in the semantic space,the large scene and high-resolution SAR images are divided into hybrid aggregated,structural and homogeneous pixel-level subspaces.As it is difficult to obtain the labeled samples for the large scene and high-resolution SAR images,the unsupervised segmentation problem of the large scene and high-resolution SAR images is solved by the unsupervised segmentation of three structure pixel-level subspaces.The complexity of the object representation of the extremely unhomogeneous pixel-level subspace has a close relationship with the diversity of the size and shape of the target in the region,the diversity of spatial topological structures with adjacent targets,the target scattering characteristics and the speckle noise,etc.So compared with the homogeneous and unhomogeneous region,the extremely unhomogeneous region is more difficult to represent,which leads to the unsupervised segmentation of the hybird aggreated structure pixel-level subspace more challenging.Therefore,this project will focus on the unsupervised hierarchical represention learning model and method to realize the unsupervised segemention of the hybird aggregated stucture pixel-level subspace,and explore a new non-deep and interpretable apporach to realize unsupervised representation learning.

项目受资助省

陕西省

项目结题报告(全文)

由于SAR图像中极不匀质区域地物表示的复杂性与该区域目标大小和形状的多样性、相邻目标空间拓扑结构的多样性、目标散射特性和相干斑噪声等有密切的关系。针对SAR图像极不匀质区域地物的复杂形状结构表征难的问题，本项目首先构建了能捕捉方向信息的Ridgelet核函数和Curvelet核函数，并基于这些核函数设计了方向结构滤波器；接着，为了以无监督学习的方式获取初始结构滤波器，该项目利用了素描线段的方向和个数等信息，建立了基于素描方向信息的能量保真目标函数，并提出了相应优化方法。理论分析和实验表明本项目提出的基于素描方向信息的能量保真模型和优化方法在无需样本标签的条件下，能更好地求解出那些与原始影像块中蕴含的复杂地物结构尽量匹配的初始结构滤波器。在此基础上，针对目标形状边界的建模问题，本项目提出了以素描线段为单位构建能包含边界信息在内的几何结构块，设计了几何结构块中描述边界邻域内像素值之间约束关系的函数，建立了基于几何结构块约束关系的结构能量保真模型，并提出了相应优化方法。理论分析和实验表明与强调重构块与原始影像块的整体Frobenius范数最小化不同，本项目提出的模型在关注它们之间整体Frobenius范数较小化的同时，还充分考虑了目标形状边界的建模问题。最后，在语义空间和像素空间信息交互框架下，针对混合地物结构像素子空间的无监督分割问题，本项目提出了一种基于素描方向统计信息和特征学习的两阶段聚类的混合像素子空间分割方法，该方法将人工设计的特征和学习得到的特征通过两阶段聚类有机地结合起来，使得分割性能得到较大提升。

  • 排序方式：
  • 1
  • /

  • 1.Deep Learning in Visual Tracking: A Review

    • 关键词：
    • Visualization; Target tracking; Task analysis; Feature extraction; Deeplearning; Trajectory; Nonhomogeneous media; Deep learning (DL);multiple-object tracking (MOT); single-object tracking (SOT);MULTIPLE OBJECT TRACKING; CORRELATION FILTERS; NEURAL-NETWORKS; ROBUST;MULTITARGET; SYSTEM

    Deep learning (DL) has made breakthroughs in many computer vision tasks and also in visual tracking. From the beginning of the research on the automatic acquisition of high abstract feature representation, DL has gone deep into all aspects of tracking to date, to name a few, similarity metric, data association, and bounding box estimation. Also, pure DL-based trackers have obtained the state-of-the-art performance after the community's constant research. We believe that it is time to comprehensively review the development of DL research in visual tracking. In this article, we overview the critical improvements brought to the field by DL: deep feature representations, network architecture, and four crucial issues in visual tracking (spatiotemporal information integration, target-specific classification, target information update, and bounding box estimation). The scope of the survey of DL-based tracking covers two primary subtasks for the first time, single-object tracking and multiple-object tracking. Also, we analyze the performance of DL-based approaches and give meaningful conclusions. Finally, we provide several promising directions and tasks in visual tracking and relevant fields.

    ...

  • 2.Deep Learning in Visual Tracking: A Review

    • 关键词：
    • Visualization; Target tracking; Task analysis; Feature extraction; Deeplearning; Trajectory; Nonhomogeneous media; Deep learning (DL);multiple-object tracking (MOT); single-object tracking (SOT);MULTIPLE OBJECT TRACKING; CORRELATION FILTERS; NEURAL-NETWORKS; ROBUST;MULTITARGET; SYSTEM

    Deep learning (DL) has made breakthroughs in many computer vision tasks and also in visual tracking. From the beginning of the research on the automatic acquisition of high abstract feature representation, DL has gone deep into all aspects of tracking to date, to name a few, similarity metric, data association, and bounding box estimation. Also, pure DL-based trackers have obtained the state-of-the-art performance after the community's constant research. We believe that it is time to comprehensively review the development of DL research in visual tracking. In this article, we overview the critical improvements brought to the field by DL: deep feature representations, network architecture, and four crucial issues in visual tracking (spatiotemporal information integration, target-specific classification, target information update, and bounding box estimation). The scope of the survey of DL-based tracking covers two primary subtasks for the first time, single-object tracking and multiple-object tracking. Also, we analyze the performance of DL-based approaches and give meaningful conclusions. Finally, we provide several promising directions and tasks in visual tracking and relevant fields.

    ...
  • 排序方式：
  • 1
  • /