WO2016184099A1 - 基于光场数据分布的深度估计方法 - Google Patents

基于光场数据分布的深度估计方法 Download PDF

Info

Publication number
WO2016184099A1
WO2016184099A1 PCT/CN2015/098117 CN2015098117W WO2016184099A1 WO 2016184099 A1 WO2016184099 A1 WO 2016184099A1 CN 2015098117 W CN2015098117 W CN 2015098117W WO 2016184099 A1 WO2016184099 A1 WO 2016184099A1
Authority
WO
WIPO (PCT)
Prior art keywords
depth
light field
scene
pixel
macro
Prior art date
Application number
PCT/CN2015/098117
Other languages
English (en)
French (fr)
Inventor
金欣
许娅彤
戴琼海
Original Assignee
清华大学深圳研究生院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 清华大学深圳研究生院 filed Critical 清华大学深圳研究生院
Publication of WO2016184099A1 publication Critical patent/WO2016184099A1/zh
Priority to US15/809,769 priority Critical patent/US10346997B2/en

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/50Depth or shape recovery
    • G06T7/55Depth or shape recovery from multiple images
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/50Depth or shape recovery
    • G06T7/55Depth or shape recovery from multiple images
    • G06T7/557Depth or shape recovery from multiple images from light fields, e.g. from plenoptic cameras
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T1/00General purpose image data processing
    • G06T1/0007Image acquisition
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/50Depth or shape recovery
    • G06T7/529Depth or shape recovery from texture
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/50Depth or shape recovery
    • G06T7/55Depth or shape recovery from multiple images
    • G06T7/571Depth or shape recovery from multiple images from focus
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N23/00Cameras or camera modules comprising electronic image sensors; Control thereof
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04NPICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N5/00Details of television systems
    • H04N5/222Studio circuitry; Studio devices; Studio equipment
    • H04N5/2224Studio circuitry; Studio devices; Studio equipment related to virtual studio applications
    • H04N5/2226Determination of depth image, e.g. for foreground/background separation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10028Range image; Depth image; 3D point clouds
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10052Images from lightfield camera

Definitions

  • the invention relates to the field of computer vision and digital image processing, and in particular to a depth estimation method based on light field data distribution.
  • Stereo matching algorithm Existing depth estimation methods based on light field cameras can be roughly divided into two categories: stereo matching algorithm and light field analysis algorithm.
  • the traditional stereo matching algorithm directly calculates the depth using the correlation between the sub-aperture images acquired by the light field camera.
  • Such algorithms are generally computationally complex, and because the low resolution of the subaperture image does not meet the accuracy requirements required for algorithm matching, the resulting depth results are of poor quality.
  • Other improved stereo matching algorithms such as considering the linearity of light propagation, still limit the performance of depth estimation by using only the correlation information of each viewpoint image in the light field data.
  • the light field analysis method attempts to estimate the depth by using both the consistency of the viewpoint images and the focal length information contained in the light field data.
  • This type of algorithm defines different cost equations for different clues, and combines the depth estimates obtained by the two clues to complement each other to improve the accuracy of the final result.
  • the depth estimated by this type of algorithm lacks detailed information, and there is still room for improvement in accuracy and consistency.
  • the idea of the present invention is to fully extract the focal depth-related tensor estimation scene depth from a series of refocused light field images obtained by changing the pixel distribution of the input light field image with sufficient reference to the light field data characteristics. Further, the tensor with depth trend and the gradient information of the scene center sub-aperture texture map are used to establish a multi-confidence model to measure the initial depth quality of each point, and the length complement is used to optimize the preliminary estimation result, and the data is calculated by using the light field camera. The purpose of high quality depth images.
  • Step S2 is repeated to obtain the scene depth of all macro pixels.
  • the pixel distribution of the input light field image is adjusted using a point spread function.
  • the method further includes the step S3 of globally optimizing the scene depth obtained in step S2 according to the reliability model.
  • the step S3 of globally optimizing the scene depth obtained in step S2 according to the credibility model includes: using the scene depth obtained in step S2 as an initial input.
  • the Markov random field is used for optimization.
  • the specific optimization methods include: depth assessment of each point according to the credibility model, using inaccurate depth estimation to correct inaccurate depth, and improving homogeneity region depth estimation. Consistency and preserve depth boundaries.
  • the credibility model is a multi-level credibility model including a first part for measuring the accuracy of the depth of the scene. And a second part for measuring the consistency of the depth of the scene in the non-boundary region and the abruptness of the boundary region.
  • the first part of the multivariate credibility model is C 1 (x, y),
  • R z* (x, y) and R z' (x, y) are the minimum and minimum values of the intensity range R z (x, y) as a function of the depth of the scene, z* and z, respectively. 'The depth of the scene corresponding to the minimum point and the minimum point.
  • the second part of the multivariate credibility model is based on gradient information of the central subaperture texture map; the depth estimation method further includes separately acquiring a step of a central subaperture texture map of the plurality of refocused light field images, and a step of calculating credibility through the second portion of the multivariate credibility model using the acquired central subaperture texture map.
  • the present invention extracts the focal depth-dependent tensor estimated scene depth from a series of refocused light field images obtained by transforming the input image pixel distribution.
  • the tensor with depth trend and the gradient information of the scene center subaperture texture map are used to define the accuracy and consistency of the multivariate credibility model to measure the initial depth to further optimize the depth estimation.
  • the scene texture and spatial information collected by a light field camera such as Lytro can be fully utilized to obtain a scene depth estimation with rich details, clear features, high accuracy and consistency.
  • FIG. 1 is a flow chart of some embodiments of a depth estimation method based on light field data distribution according to the present invention.
  • some embodiments of a depth estimation method based on a light field data distribution include the following steps:
  • S1 Adjust a pixel distribution of the input light field image to generate a plurality of refocused light field images of different focal lengths.
  • the input light field image is first subjected to pre-correction processing to remove peripheral points of each macro pixel that fail to capture valid data information, thereby preventing meaningless pixel values from interfering with subsequent processing.
  • the point spread function (PSF) is used to adjust the pixel position distribution of the corrected light field image L o , and the refocusing process on the input light field image is as follows:
  • the macro pixel corresponds to a point in the actual scene, and the intensity range of the macro pixel is a variation range of the intensity values of all points in the macro pixel.
  • Each microlens of the microlens array of the light field camera represents a subaperture at a certain angle with respect to the main lens of the camera.
  • the macro pixel in the light field data corresponds to a point in the actual scene, and the macro pixel includes the corresponding scene point.
  • the angle information of the entire microlens array projection is correspondingly recorded at each point in the macro pixel, that is, the intensity value and the distribution position of each point.
  • L z the focal plane of a series of light field images
  • the intensity values of the points in the macro pixel are constantly changing, and the intensity range of the entire macro pixel is also changed. Therefore, the depth is determined by using the intensity range of the macro pixel as the depth-dependent tensor.
  • the macro pixel intensity range is extracted as follows:
  • I(x, y, u, v) is the intensity value of a point (u, v) in the microlens (corresponding to the macro pixel (x, y) in the image plane L z ) at coordinates (x, y)
  • M Represents a collection of all points within the microlens.
  • the point is accurately projected on the image plane through the sub-aperture at each angle, that is, the projection of each angle accurately reflects the texture value of the point, so
  • the intensity range of each point in the corresponding macro pixel is the smallest - the macro pixel intensity range is the smallest.
  • the focal length of the light field image L z focusing on the scene point reflects the depth information of the point, thereby obtaining an initial estimate of the scene depth D initial (x, y) of the macro pixel (x, y),
  • step S2 the scene depth of all macro pixels can be obtained.
  • step S3. Perform global optimization on the depth of the scene obtained in step S2 according to the credibility model.
  • the multivariate credibility model includes a first portion for measuring the accuracy of the depth of the scene, and a measure of the consistency of the scene depth in the non-boundary region and the abruptness of the boundary region. The second part.
  • the multi-confidence model is established as follows: First, define the unitary credibility (ie, the first part of the multi-level credibility model) to measure the accuracy of the depth estimation, and extract the intensity range R z of each point by analysis ( x, y) With the trend of the depth D (variation curve), it is found that the Euclidean distance between the minimum point and the minimum point of the curve is positively correlated with the accuracy of the point depth estimate D initial (x, y). Therefore, the credibility C 1 corresponding to the accuracy given to the depth estimation of each point is as follows:
  • R z* (x, y) and R z' (x, y) are the minimum and minimum values of the intensity curve R z (x, y) as a function of depth D, respectively, z* and z 'For their respective depths.
  • elemental credibility ie the second part of the multivariate credibility model to measure the consistency of the estimated depth D initial in the non-boundary region and the abrupt change of the boundary region, according to the gradient information of the central subaperture texture image.
  • the non-boundary area changes gently, and the characteristics of the mutation in the boundary area define another unit of confidence C 2 as follows:
  • the second portion of the multivariate credibility model is based on gradient information of the central subaperture texture map.
  • the depth estimation method further comprises the steps of respectively acquiring a central subaperture texture map of the plurality of refocused light field images, and passing the obtained multi-trust reliability model with the obtained central subaperture texture map
  • the second part is the step of calculating the credibility. Specifically, since the angle information is recorded at each point in the macro pixel, the image formed by the center point of each macro pixel is the central subaperture texture map.
  • the step of global optimization comprises: using the scene depth D initial obtained in step S2 as an initial input, optimized using a Markov Random Field (MRF).
  • MRF Markov Random Field
  • the optimization principle is to improve the accuracy and consistency of the depth estimation and to preserve clear boundary features.
  • the specific optimization method includes: determining the depth of each point according to the credibility model, using the high-accuracy depth estimation to correct the inaccurate depth, improving the consistency of the homogeneity depth estimation and retaining the depth boundary.
  • the final depth estimate D final is as follows:
  • ⁇ flat and ⁇ smooth are the parameters of the Laplace constraint and the second-order differential term, which respectively limit the smoothness and continuity of the final depth estimate D final .
  • the error between D final and the constraint term can be calculated, and the error matrix is constructed to minimize the formula (8), thereby further optimizing the depth estimation result.
  • global optimization by using Markov random field is only a preferred manner, and the present invention can also adopt other methods for global optimization, such as image-cut-based multi-mark optimization, joint discrete continuous optimization, and the like.
  • Some embodiments described above fully refer to the characteristics of the light field data, extract the focal depth-related tensor estimation scene depth from a series of refocused light field images obtained by transforming the input light field image pixel distribution, and utilize the variation trend of the tensor with depth.
  • the gradient information of the scene center subaperture texture map defines the multivariate credibility model to measure the accuracy and consistency of the initial depth to further optimize the depth estimate.
  • the scene texture and spatial information collected by a light field camera such as Lytro can be fully utilized to obtain indoor and outdoor scene depth estimation with rich details, clear features, high accuracy and consistency.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Studio Devices (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

一种基于光场数据分布的深度估计方法,其包括以下步骤:S1、调整输入光场图像的像素分布,生成多个不同焦距的重聚焦光场图像;S2、针对所述多个重聚焦光场图像,分别提取同一个宏像素的强度范围,进而选出最小强度范围对应的重聚焦光场图像,以该重聚焦光场图像的焦距作为该宏像素的场景深度;所述宏像素对应实际场景中的一点,所述宏像素的强度范围为该宏像素内所有点的强度值的变化范围;以及重复所述步骤S2,获得所有宏像素的场景深度。利用本发明的方法能够充分借助类似Lytro等光场相机采集的场景纹理及空间信息,获得细节丰富,特征清晰,准确性、一致性高的场景深度估计。

Description

基于光场数据分布的深度估计方法 技术领域
本发明涉及计算机视觉与数字图像处理领域,特别涉及一种基于光场数据分布的深度估计方法。
背景技术
基于光场成像理论,新近发布的光场相机(如Lytro和RayTrix)获得了巨大的商业成功,并引发研究学界的广泛关注。普通用户利用单台相机除了能实现重聚焦与视点移动功能,也可以借助相关算法从所拍摄的光场图像中估算场景深度。该方法计算深度具有价廉、便捷等优势。
现存基于光场相机的深度估计方法大体可分为两类:立体匹配算法与光场分析算法。传统的立体匹配算法直接利用光场相机采集的各子孔径图像间的相关性计算深度。然而,这类算法通常计算复杂度较高,并且因为子孔径图像的低分辨率无法满足算法匹配所需精度要求,所以得到的深度结果质量较差。另一些改良的立体匹配算法,比如考虑光线传播的直线性,却仍旧因为只利用到光场数据中各视点图像的相关性信息,限制了深度估计的性能。
光场分析方法试图同时利用光场数据中包含的各视点图像的一致性与焦距信息这两种线索来估算深度。该类算法为不同线索定义不同的代价方程,并融合两种线索得到的深度估计进行互补来提升最终结果的准确度。但是,由该类算法估计的深度缺乏细节信息,在准确性与一致性上还有待提升。
发明内容
本发明的目的是提供一种基于光场数据分布的深度估计方法,以解决现有光场相机深度估计方法存在的上述技术问题中的至少一种。
本发明的思路是充分参考光场数据特性,从改变输入光场图像像素分布所得的一系列重聚焦光场图像中提取与焦距相关张量预估场景深度。进一步,还利用该张量随深度的变化趋势与场景中心子孔径纹理图的梯度信息建立多元可信度模型衡量各点初始深度质量,取长补短从而优化初步估计结果,达到利用光场相机采集数据计算高质量深度图像的目的。
本发明提供的一种基于光场数据分布的深度估计方法包括以下步骤:
S1、调整输入光场图像的像素分布,生成多个不同焦距的重聚焦光场图像;
S2、针对所述多个重聚焦光场图像,分别提取同一个宏像素的强度范围,进而选出最小强度范围对应的重聚焦光场图像,以该重聚焦光场图像的焦距作为该宏像素的场景深度;所述宏像素对应实际场景中的一点,所述宏像素的强度范围为该宏像素内所有点的强度值的变化范围;以及
重复所述步骤S2,获得所有宏像素的场景深度。
在上述的基于光场数据分布的深度估计方法中,优选地,所述步骤S1中,采用点扩散函数调整输入光场图像的像素分布。
在上述的基于光场数据分布的深度估计方法中,优选地,进一步还包括依据可信度模型对步骤S2获得的场景深度进行全局优化的步骤S3。
在上述的基于光场数据分布的深度估计方法中,优选地,所述的依据可信度模型对步骤S2获得的场景深度进行全局优化的步骤S3包括:以步骤S2获得的场景深度作为初始输入,利用马尔科夫随机场进行优化,具体的优化方法包括:依据所述可信度模型对各点的深度评估,利用准确性高的深度估计修正不准确的深度,提升同质区域深度估计的一致性并保留深度边界。
在上述的基于光场数据分布的深度估计方法中,优选地,所述可信度模型为多元可信度模型,该多元可信度模型包括用于衡量所述场景深度的准确性的第一部分,以及用于衡量所述场景深度在非边界区域的一致性与边界区域的突变性的第二部分。
在上述的基于光场数据分布的深度估计方法中,优选地,所述多元可信度模型的第一部分为C1(x,y),
Figure PCTCN2015098117-appb-000001
其中,Rz*(x,y)和Rz’(x,y)分别是强度范围Rz(x,y)随场景深度的变化曲线的最小值点和极小值点,z*和z’为最小值点和极小值点对应的场景深度。
在上述的基于光场数据分布的深度估计方法中,优选地,所述多元可信度模型的第二部分以中心子孔径纹理图的梯度信息为基础;所述深度估计方法进一步还包括分别获取所述多个重聚焦光场图像的中心子孔径纹理图的步骤,以及用获取的中心子孔径纹理图通过所述多元可信度模型的第二部分计算可信度的步骤。
本发明从变换输入图像像素分布所得的一系列重聚焦光场图像中提取与焦距相关张量估计场景深度。在更优的方案中,还利用该张量随深度的变化趋势与场景中心子孔径纹理图的梯度信息定义多元可信度模型衡量初始深度的准确性与一致性来进一步优化深度估计。利用本发明的方法能够充分借助类似Lytro等光场相机采集的场景纹理及空间信息,获得细节丰富,特征清晰,准确性、一致性高的场景深度估计。
附图说明
图1为本发明基于光场数据分布的深度估计方法一些实施例的流程图。
具体实施方式
下面结合附图和实施例对本发明进一步说明。这些更详细的描述旨在帮助理解本发明,而不应被用于限制本发明。根据本发明公开的内容,本领域技术人员明白,可以不需要一些或者所有这些特定细节即可实施本发明。而在其它情况下,为了避免将发明创造淡化,未详细描述众所周知的操作过程。
参照图1,一些实施例基于光场数据分布的深度估计方法包括以下步骤:
S1、调整输入光场图像的像素分布,生成多个不同焦距的重聚焦光场图像。
具体地,输入光场图像首先经过预矫正处理,将每个宏像素中未能捕捉到有效数据信息的外围点去除,防止无意义像素值干扰后续处理。采用点扩散函数(Point Spread Function,PSF)调整矫正后的光场图像Lo的像素位置分布,实现对输入光场图像的重聚焦处理如下:
Figure PCTCN2015098117-appb-000002
生成聚焦平面由近及远变化的一系列光场图像Lz(z=1,2,3...),其中z为预设的深度层次,x,y与u,v分别为图像平面上的空间坐标与角度坐标。
S2、针对所述多个重聚焦光场图像,分别提取同一个宏像素的强度范围,进而选出最小强度范围对应的重聚焦光场图像,以该重聚焦光场图像的焦距作为该宏像素的场景深度。所述宏像素对应实际场景中的一点,所述宏像素的强度范围为该宏像素内所有点的强度值的变化范围。
光场相机的微透镜阵列的每个微透镜相对于相机的主透镜而言代表某角度上的子孔径,光场数据中宏像素对应实际场景中的一点,宏像素包含了其相应场景点通过整个微透镜阵列投影的角度信息,并对应记录在宏像素内各点上,即各点的强度值与分布位置。根据光场成像理论,图像宏像素中各点强度反映场景点通过不同位置微透镜的投影,在一系列光场图像Lz(z=1,2,3...)的聚焦平面由近及远变化时,宏像素中各点的强度值在不断变化,导致整个宏像素的强度范围也随之变化,因此,以宏像素的强度范围作为与深度相关张量来估计深度。具体地,提取宏像素强度范围如下:
Figure PCTCN2015098117-appb-000003
其中I(x,y,u,v)是位于坐标(x,y)的微透镜(对应图像平面Lz中的宏像素(x,y))内一点(u,v) 的强度值,M表示该微透镜内所有点的集合。
由光场成像原理分析可知当场景点恰好位于相机聚焦平面上时,该点透过位于各角度的子孔径被准确投影在图像平面上,即各角度投影均准确反映了该点的纹理值,因此其对应宏像素内各点强度的变化范围最小——宏像素强度范围最小。聚焦该场景点的光场图像Lz的焦距反映了该点的深度信息,由此得到宏像素(x,y)的场景深度初始估计Dinitial(x,y),
Figure PCTCN2015098117-appb-000004
重复上述步骤S2,即可获得所有宏像素的场景深度。
S3、依据可信度模型对步骤S2获得的场景深度进行全局优化。
在优选的实施例中,所述多元可信度模型包括用于衡量所述场景深度的准确性的第一部分,以及用于衡量所述场景深度在非边界区域的一致性与边界区域的突变性的第二部分。
在更优选实施例中,多元可信度模型建立如下:首先定义一元可信度(即多元可信度模型的第一部分)以衡量深度估计的准确性,通过分析提取各点强度范围Rz(x,y)随深度D的变化趋势(变化曲线),发现该曲线的最小值点与极小值点的欧氏距离与该点深度估计Dinitial(x,y)的准确性存在正相关关系,由此对各点深度估计赋予与准确性相应的可信度C1如下:
Figure PCTCN2015098117-appb-000005
其中,Rz*(x,y)和Rz’(x,y)分别是强度范围Rz(x,y)随深度D的变化曲线的最小值点与极小值点,z*和z’为其各自对应的深度。其次,定义另一元可信度(即多元可信度模型的第二部分)以衡量估计深度Dinitial在非边界区域的一致性与边界区域的突变性,根据中心子孔径纹理图像的梯度信息在非边界区域变化平缓、在边界区域突变的特性,定义另一元可信度C2如下:
C2(x,y)=(Gradx(x,y)+Grady(x,y))/2         (5)
Figure PCTCN2015098117-appb-000006
                                                                   (7)
Figure PCTCN2015098117-appb-000007
其中Gradx和Grady分别为图像平面x,y方向上的梯度值,F(x,y)为中心子孔径纹理图(x,y)处像素的强度值。最后,结合可信度C1和C2建立多元可信度模型C如下:
C(x,y)=C1(x,y)·C2(x,y)         (6)
在上述更优选实施例中,所述多元可信度模型的第二部分以中心子孔径纹理图的梯度信息为基础。相应地,所述深度估计方法进一步还包括分别获取所述多个重聚焦光场图像的中心子孔径纹理图的步骤,以及用获取的中心子孔径纹理图通过所述多元可信度模型的第二部分计算可信度的步骤。具体地,由于宏像素内各点记录了角度信息,因此各宏像素中心点所构成的图像即为中心子孔径纹理图。
在一些实施例中,全局优化的步骤包括:以步骤S2获得的场景深度Dinitial作为初始输入,利用马尔科夫随机场(Markov Random Field,MRF)进行优化。优化原则是:提高深度估计的准确性与一致性,并保留清晰的边界特征。具体的优化方法包括:依据所述可信度模型对各点的深度评估,利用准确性高的深度估计修正不准确的深度,提升同质区域深度估计的一致性并保留深度边界。全局优化后得最终深度估计Dfinal,过程如下:
Figure PCTCN2015098117-appb-000008
其中,λflat和λsmooth为拉普拉斯约束项和二阶微分项的参数,分别限制了最终深度估计Dfinal的平滑性与连续性。此外还可以计算Dfinal与约束项间的误差,构造误差矩阵来最小化公式(8),由此进一步优化深度估计结果。可以理解地,利用马尔科夫随机场进行全局优化只是一个优选方式,本发明还可以采用其它方式来进行全局优化,例如基于图像切割的多标记优化,联合离散连续优化等。
上述一些实施例充分参考光场数据特性,从变换输入光场图像像素分布所得的一系列重聚焦光场图像中提取与焦距相关张量估计场景深度,并利用该张量随深度的变化趋势与场景中心子孔径纹理图的梯度信息定义多元可信度模型衡量初始深度的准确性与一致性来进一步优化深度估计。利用一些实施例能够充分借助类似Lytro等光场相机采集的场景纹理及空间信息,获得细节丰富,特征清晰,准确性、一致性高的室内外场景深度估计。

Claims (7)

  1. 一种基于光场数据分布的深度估计方法,其特征在于,包括以下步骤:
    S1、调整输入光场图像的像素分布,生成多个不同焦距的重聚焦光场图像;
    S2、针对所述多个重聚焦光场图像,分别提取同一个宏像素的强度范围,进而选出最小强度范围对应的重聚焦光场图像,以该重聚焦光场图像的焦距作为该宏像素的场景深度;所述宏像素对应实际场景中的一点,所述宏像素的强度范围为该宏像素内所有点的强度值的变化范围;以及
    重复所述步骤S2,获得所有宏像素的场景深度。
  2. 根据权利要求1所述的基于光场数据分布的深度估计方法,其特征在于,所述步骤S1中,采用点扩散函数调整输入光场图像的像素分布。
  3. 根据权利要求1所述的基于光场数据分布的深度估计方法,其特征在于,进一步还包括依据可信度模型对步骤S2获得的场景深度进行全局优化的步骤S3。
  4. 根据权利要求3所述的基于光场数据分布的深度估计方法,其特征在于,所述的依据可信度模型对步骤S2获得的场景深度进行全局优化的步骤S3包括:以步骤S2获得的场景深度作为初始输入,利用马尔科夫随机场进行优化,具体的优化方法包括:依据所述可信度模型对各点的深度评估,利用准确性高的深度估计修正不准确的深度,提升同质区域深度估计的一致性并保留深度边界。
  5. 根据权利要求3所述的基于光场数据分布的深度估计方法,其特征在于,所述可信度模型为多元可信度模型,该多元可信度模型包括用于衡量所述场景深度的准确性的第一部分,以及用于衡量所述场景深度在非边界区域的一致性与边界区域的突变性的第二部分。
  6. 根据权利要求5所述的基于光场数据分布的深度估计方法,其特征在于,
    所述多元可信度模型的第一部分为C1(x,y),
    Figure PCTCN2015098117-appb-100001
    其中,Rz*(x,y)和Rz’(x,y)分别是强度范围Rz(x,y)随场景深度的变化曲线的最小值点和极小值点,z*和z’为最小值点和极小值点对应的场景深度。
  7. 根据权利要求5所述的基于光场数据分布的深度估计方法,其特征在于,
    所述多元可信度模型的第二部分以中心子孔径纹理图的梯度信息为基础;
    所述深度估计方法进一步还包括分别获取所述多个重聚焦光场图像的中心子孔径纹理图的步骤,以及用获取的中心子孔径纹理图通过所述多元可信度模型的第二部分计算可信度的步骤。
PCT/CN2015/098117 2015-05-15 2015-12-21 基于光场数据分布的深度估计方法 WO2016184099A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US15/809,769 US10346997B2 (en) 2015-05-15 2017-11-10 Depth estimation method based on light-field data distribution

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201510251234.8 2015-05-15
CN201510251234.8A CN104899870B (zh) 2015-05-15 2015-05-15 基于光场数据分布的深度估计方法

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US15/809,769 Continuation US10346997B2 (en) 2015-05-15 2017-11-10 Depth estimation method based on light-field data distribution

Publications (1)

Publication Number Publication Date
WO2016184099A1 true WO2016184099A1 (zh) 2016-11-24

Family

ID=54032515

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/098117 WO2016184099A1 (zh) 2015-05-15 2015-12-21 基于光场数据分布的深度估计方法

Country Status (3)

Country Link
US (1) US10346997B2 (zh)
CN (1) CN104899870B (zh)
WO (1) WO2016184099A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109615650A (zh) * 2018-11-22 2019-04-12 浙江工商大学 一种基于变分法和遮挡互补的光场流估计方法
CN110415209A (zh) * 2019-06-12 2019-11-05 东北农业大学 一种基于光场视觉深度估计的奶牛采食量监测方法
CN112288669A (zh) * 2020-11-08 2021-01-29 西北工业大学 一种基于光场成像的点云地图获取方法
CN117274067A (zh) * 2023-11-22 2023-12-22 浙江优众新材料科技有限公司 一种基于强化学习的光场图像盲超分辨处理方法与系统

Families Citing this family (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104899870B (zh) 2015-05-15 2017-08-25 清华大学深圳研究生院 基于光场数据分布的深度估计方法
CN106384338B (zh) * 2016-09-13 2019-03-15 清华大学深圳研究生院 一种基于形态学的光场深度图像的增强方法
CN107038719A (zh) * 2017-03-22 2017-08-11 清华大学深圳研究生院 基于光场图像角度域像素的深度估计方法及系统
CN111406412B (zh) 2017-04-11 2021-09-03 杜比实验室特许公司 分层的增强型娱乐体验
CN107702660A (zh) * 2017-04-22 2018-02-16 广州首感光电科技有限公司 基于光场显微成像的微尺度三维工业检测系统
CN107610170B (zh) * 2017-08-04 2020-05-19 中国科学院自动化研究所 多目图像重聚焦的深度获取方法及系统
CN107909578A (zh) * 2017-10-30 2018-04-13 上海理工大学 基于六边形拼接算法的光场图像重聚焦方法
CN107726993B (zh) * 2017-10-30 2019-07-19 上海理工大学 基于光场图像宏像素最大梯度区域的颗粒深度测量方法
CN108230223A (zh) * 2017-12-28 2018-06-29 清华大学 基于卷积神经网络的光场角度超分辨率方法及装置
CN111869205B (zh) 2018-01-19 2022-06-10 Pcms控股公司 具有变化位置的多焦平面
CN108596965B (zh) * 2018-03-16 2021-06-04 天津大学 一种光场图像深度估计方法
CN110349196B (zh) * 2018-04-03 2024-03-29 联发科技股份有限公司 深度融合的方法和装置
EP4270944A3 (en) * 2018-07-06 2024-01-03 InterDigital VC Holdings, Inc. Method and system for forming extended focal planes for large viewpoint changes
CN109064505B (zh) * 2018-07-26 2020-12-25 清华大学深圳研究生院 一种基于滑动窗口张量提取的深度估计方法
CN110120071B (zh) * 2019-05-15 2023-03-24 南京工程学院 一种面向光场图像的深度估计方法
CN111679337B (zh) * 2019-10-15 2022-06-10 上海大学 一种水下主动激光扫描成像系统中散射背景抑制方法
CN111784620B (zh) * 2020-07-06 2023-05-16 太原科技大学 空间信息引导角度信息的光场相机全聚焦图像融合算法
CN112215879A (zh) * 2020-09-25 2021-01-12 北京交通大学 一种光场极平面图像的深度提取方法
CN112288789B (zh) * 2020-10-26 2024-03-29 杭州电子科技大学 基于遮挡区域迭代优化的光场深度自监督学习方法
CN113298943A (zh) * 2021-06-10 2021-08-24 西北工业大学 一种基于光场成像的esdf地图构建方法
CN113808019A (zh) * 2021-09-14 2021-12-17 广东三水合肥工业大学研究院 一种非接触测量系统及方法
CN114511605B (zh) * 2022-04-18 2022-09-02 清华大学 光场深度估计方法、装置、电子设备及存储介质
CN115100269B (zh) * 2022-06-28 2024-04-23 电子科技大学 一种光场图像深度估计方法、系统、电子设备及存储介质
CN116449049A (zh) * 2023-03-29 2023-07-18 南京航空航天大学 一种基于多色光深度编码和光场相机的三维流场测试方法及系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102314683A (zh) * 2011-07-15 2012-01-11 清华大学 一种非平面图像传感器的计算成像方法和成像装置
CN104079827A (zh) * 2014-06-27 2014-10-01 中国科学院自动化研究所 一种光场成像自动重对焦方法
CN104463949A (zh) * 2014-10-24 2015-03-25 郑州大学 一种基于光场数字重聚焦的快速三维重建方法及其系统
US9025895B2 (en) * 2011-09-28 2015-05-05 Pelican Imaging Corporation Systems and methods for decoding refocusable light field image files
CN104899870A (zh) * 2015-05-15 2015-09-09 清华大学深圳研究生院 基于光场数据分布的深度估计方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102809918B (zh) * 2012-08-08 2014-11-05 浙江大学 基于多层空间光调制器的高分辨全息三维显示装置和方法
US9658443B2 (en) * 2013-03-15 2017-05-23 The Board Of Trustees Of The Leland Stanford Junior University Optics apparatus with detection of light rays received at different angles for output indicative of aliased views
US9524556B2 (en) * 2014-05-20 2016-12-20 Nokia Technologies Oy Method, apparatus and computer program product for depth estimation
CN104050662B (zh) * 2014-05-30 2017-04-12 清华大学深圳研究生院 一种用光场相机一次成像直接获取深度图的方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102314683A (zh) * 2011-07-15 2012-01-11 清华大学 一种非平面图像传感器的计算成像方法和成像装置
US9025895B2 (en) * 2011-09-28 2015-05-05 Pelican Imaging Corporation Systems and methods for decoding refocusable light field image files
CN104079827A (zh) * 2014-06-27 2014-10-01 中国科学院自动化研究所 一种光场成像自动重对焦方法
CN104463949A (zh) * 2014-10-24 2015-03-25 郑州大学 一种基于光场数字重聚焦的快速三维重建方法及其系统
CN104899870A (zh) * 2015-05-15 2015-09-09 清华大学深圳研究生院 基于光场数据分布的深度估计方法

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109615650A (zh) * 2018-11-22 2019-04-12 浙江工商大学 一种基于变分法和遮挡互补的光场流估计方法
CN109615650B (zh) * 2018-11-22 2022-11-25 浙江工商大学 一种基于变分法和遮挡互补的光场流估计方法
CN110415209A (zh) * 2019-06-12 2019-11-05 东北农业大学 一种基于光场视觉深度估计的奶牛采食量监测方法
CN112288669A (zh) * 2020-11-08 2021-01-29 西北工业大学 一种基于光场成像的点云地图获取方法
CN112288669B (zh) * 2020-11-08 2024-01-19 西北工业大学 一种基于光场成像的点云地图获取方法
CN117274067A (zh) * 2023-11-22 2023-12-22 浙江优众新材料科技有限公司 一种基于强化学习的光场图像盲超分辨处理方法与系统

Also Published As

Publication number Publication date
US10346997B2 (en) 2019-07-09
CN104899870B (zh) 2017-08-25
CN104899870A (zh) 2015-09-09
US20180114328A1 (en) 2018-04-26

Similar Documents

Publication Publication Date Title
WO2016184099A1 (zh) 基于光场数据分布的深度估计方法
EP3158532B1 (en) Local adaptive histogram equalization
CN107995424B (zh) 基于深度图的光场全聚焦图像生成方法
CN107316326B (zh) 应用于双目立体视觉的基于边的视差图计算方法和装置
WO2018024006A1 (zh) 一种聚焦型光场相机的渲染方法和系统
WO2014044126A1 (zh) 坐标获取装置、实时三维重建系统和方法、立体交互设备
WO2020007320A1 (zh) 多视角图像的融合方法、装置、计算机设备和存储介质
US8144974B2 (en) Image processing apparatus, method, and program
CN109345502B (zh) 一种基于视差图立体结构信息提取的立体图像质量评价方法
CN107084680B (zh) 一种基于机器单目视觉的目标深度测量方法
TWI529661B (zh) 快速建立景深圖的方法及影像處理裝置
CN110120071B (zh) 一种面向光场图像的深度估计方法
WO2015184978A1 (zh) 摄像机控制方法、装置及摄像机
TW201214335A (en) Method and arrangement for multi-camera calibration
WO2018053952A1 (zh) 一种基于场景样本库的影视图像深度提取方法
KR101714224B1 (ko) 센서 융합 기반 3차원 영상 복원 장치 및 방법
US8340399B2 (en) Method for determining a depth map from images, device for determining a depth map
US9635245B2 (en) Focus position detection device and focus position detection method
CN115375745A (zh) 基于偏振微透镜光场影像视差角的绝对深度测量方法
TWI538476B (zh) 立體攝影系統及其方法
JP7312026B2 (ja) 画像処理装置、画像処理方法およびプログラム
CN112132925A (zh) 用于重建水下图像颜色的方法和装置
JP2009293971A (ja) 距離測定装置および方法並びにプログラム
JP2017011397A (ja) 画像処理装置、画像処理方法、及びプログラム
CN114998522A (zh) 多视连续光场影像室内场景稠密点云准确提取方法及系统

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15892473

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 15892473

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 04.05.2018)

122 Ep: pct application non-entry in european phase

Ref document number: 15892473

Country of ref document: EP

Kind code of ref document: A1