| 单幅图像的深度标签流形学习 |
| |
| 引用本文： | 叶华,谭冠政.单幅图像的深度标签流形学习[J].红外与激光工程,2018,47(6):626004-0626004(7). |
| |
| 作者姓名： | 叶华 谭冠政 |
| |
| 作者单位： | 1.中南大学 信息科学与工程学院,湖南 长沙 410083; |
| |
| 基金项目： | 国家自然科学基金（61403136）;湖南省自然科学基金（14JJ5008）;湖南省教育厅项目（16C1087） |
| |
| 摘 要： | 图像中背景与前景对象的空间位置决定了场景在图像中的相对深度，利用图像的局部特征相似性和流形结构的降维性能，并应用salient区域DCT高频系数分布的深度排序索引性能，定义出图像深度的马尔科夫概率图模型MRF。通过划分场景对象检测salient区域模糊度，最后估计得出图像场景的相对深度图。通过学习图像数据的流形嵌入对数据流形分布概率密度函数进行迁移，得出遵循相似流形分布的对象特征类别标记概率密度分布。进一步检测空间变化salient区的模糊程度，融合多尺度梯度幅度的高频离散余弦变换DCT系数特征，依据模糊变化高频特征计算深度标记索引确定深度标签的层级次序，融合类别标签以生成深度图。这种模型框架下检测单个图像中模糊和未模糊的区域，可获得图像中场景的相对深度，而无需了解相机设置或模糊类型的先验参数。在典型的深度图估计数据集中应用MRF深度图模型评测图像的深度估计性能，实验结果给出该方法在检测场景分布和划分场景深度次序上的准确率，验证了方法的有效性。
|
| 关 键 词： | 流形嵌入 深度排序 标记流形学习 离散余弦变换 模糊检测 |
| 收稿时间： | 2018-01-07 |
Manifold learning of depth label for single image |
| |
| Affiliation: | 1.School of Information Science and Engineering,Central South University,Changsha 410083,China;2.Hunan Province Cooperative Innovation Center for The Construction & Development of Dongting Lake Ecological Economic Zone,Hunan University of Arts and Science,Changde 415000,China |
| |
| Abstract: | The spatial position of the background and foreground determines the relative depth of the scene in the image. Using similarity characteristics of the local region of image and properties of dimensionality reduction of the manifold structure, the depth sortingindexing performance of the DCT high coefficientsfrequency distribution in the salient region was applied, the probability image map model of Markov Random Field(MRF) was defined to establish a relationship between the local feature and depth of different locations in the image. By segmenting the object, detecting relative blurring of the salient regions, and finally the relative depth map of the scene in the image was estimated. Through learning data embedding of manifold of the image, the probability density function of the data manifold distribution was migrated, the probability density function of category labels of object which followed similar manifold distributions was obtained. The blurred extent of salient regions was detected further, the high-frequency coefficient of discrete cosine transform(DCT) of multi-scale gradient amplitudes was fused, then depth mark index was calculated according to the high frequency characteristics of the fuzzy change to determine the hierarchical order of the depth tags, and the category tags were merged to generate a depth map. In this model framework, the blurred and unambiguous areas in a single image were detected to obtain the relative depth of the scene in the image, without knowing the priori settings of the camera or the type of blur. The depth estimation performance of the image was evaluated by using the MRF depth map model in a typical depth map estimation data set. The experimental results show the accuracy of the method in detecting scene distribution and ordering the depth of scene. It verifies the validity of the method. |
| |
| Keywords: | |
|
| 点击此处可从《红外与激光工程》浏览原始摘要信息 |
|
点击此处可从《红外与激光工程》下载全文 |
|
