WO2019214568A1 - 一种基于深度的光场拼接方法 - Google Patents

一种基于深度的光场拼接方法 Download PDF

Info

Publication number
WO2019214568A1
WO2019214568A1 PCT/CN2019/085643 CN2019085643W WO2019214568A1 WO 2019214568 A1 WO2019214568 A1 WO 2019214568A1 CN 2019085643 W CN2019085643 W CN 2019085643W WO 2019214568 A1 WO2019214568 A1 WO 2019214568A1
Authority
WO
WIPO (PCT)
Prior art keywords
light field
depth
feature point
transformation matrix
point pairs
Prior art date
Application number
PCT/CN2019/085643
Other languages
English (en)
French (fr)
Inventor
金欣
王培�
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 WO2019214568A1 publication Critical patent/WO2019214568A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00Geometric image transformations in the plane of the image
    • G06T3/40Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G06T3/4038Image mosaicing, e.g. composing plane images from plane sub-images
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/23Clustering techniques
    • G06F18/232Non-hierarchical techniques
    • G06F18/2321Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/10Segmentation; Edge detection
    • G06T7/162Segmentation; Edge detection involving graph-based methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/30Determination of transform parameters for the alignment of images, i.e. image registration
    • G06T7/33Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
    • G06T7/337Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods involving reference images or patches
    • 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
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20212Image combination
    • G06T2207/20221Image fusion; Image merging

Definitions

  • the invention relates to the field of computer vision and digital image processing, and in particular to a depth-based light field splicing method.
  • Light field imaging records light information from all directions by adding a microlens array between the main lens and the sensor to obtain a complete information optical radiation field.
  • the resolution of optical sensors continues to rise and the field of light cameras becomes more market-oriented, the practical value of light field imaging technology is becoming higher and higher.
  • the microlens array-based light field camera can record the spatial position information and direction information of the three-dimensional scene at the same time. Therefore, the light field data collected by the light field camera has a wide range of applications, such as refocusing. , depth estimation, significance test, etc.
  • the field of view of the handheld all-optical camera is small.
  • the light field splicing method that does not depend on the light field structure setting can improve the field of view of the light field camera.
  • the existing method of light field splicing mainly uses the feature extraction and matching method to calculate the transformation matrix between adjacent input light fields, perform light field registration, and find the light by constructing the energy loss function of the overlapping light field overlap region.
  • the optimal stitching of the field realizes the fusion of the light field; this method has certain limitations. Although it can realize the light field splicing with less parallax, once the parallax of the input light field data changes greatly, it will produce misalignment and ghosting. Waiting for errors, you can't get the correct stitching results.
  • Another method is to reduce the influence of parallax on the splicing result by using image splicing method combined with parallax tolerance. This method improves the splicing accuracy to a certain extent, but can not completely offset the influence of parallax, and Separating the image of the corresponding sub-aperture image of the light field alone will introduce the problem that the entire light field is inconsistent in the spatial domain and the angular domain.
  • the present invention provides a depth-based light field splicing method, which solves the problem of dislocation and ghosting caused by large parallax changes, and realizes a precise parallax-tolerant light field splicing method.
  • the invention discloses a depth-based light field splicing method, comprising the following steps:
  • A1 inputting a light field to be spliced and a sub-aperture image of the light field, and performing a light field depth estimation on the sub-aperture image of the light field to obtain a depth map of the light field;
  • A2 extracting feature points in the subaperture image of the light field, matching the feature points to obtain feature point pairs, and screening the feature point pairs to obtain matching feature point pairs;
  • A3 4D meshing the light field to be spliced, predicting the global homography transformation matrix according to the matching feature point pairs; and establishing a weight matrix according to the depth and position relationship between the feature point and the grid center point;
  • the optimal transformation matrix and the weight matrix are used to predict the optimal homography transformation matrix of each grid, and the light field is mapped according to the optimal homography transformation matrix of each grid in the light field;
  • A4 The light field is fused to obtain the result of the light field splicing.
  • step A2 specifically includes:
  • A21 extracting feature points in the subaperture image of the light field, and matching the feature points to obtain feature point pairs;
  • A22 performing feature clustering on the depth map to obtain a depth level map of the light field
  • A23 Grouping the feature point pairs according to the depth level map, respectively filtering the feature point pairs of each group, and combining the selected feature point pairs to obtain matching feature point pairs.
  • step A3 specifically includes:
  • A31 4D meshing the light field to be spliced
  • step A32 determining whether each of the grids after 4D meshing includes different depth layers, and if so, dividing the network again by the depth layer; otherwise, directly proceeding to step A33;
  • A33 predicting a global homography transformation matrix according to matching feature point pairs
  • A34 establishing a weight matrix according to the depth and position relationship between the feature point and the center of the grid point;
  • A35 predicting an optimal homography transformation matrix of each grid according to a global homography transformation matrix and a weight matrix
  • A36 The light field is mapped according to the optimal homography transformation matrix of each grid in the light field.
  • the invention has the beneficial effects that the depth-based light field splicing method of the present invention uses the optimal homography transformation matrix of each grid instead of the global homography transformation matrix to map the light field.
  • the optimal homography transformation matrix of each mesh is combined with the global homography transformation matrix and introduced
  • the weight matrix established by the depth map is used for prediction, which solves the problem of misalignment and ghosting caused by large parallax changes, and realizes accurate parallax-tolerant light field splicing method; thereby further realizing more accurate large parallax light.
  • the field splicing ensures the consistency of the spatial and angular domains of the spliced light field, thereby expanding the viewing angle of the light field.
  • the depth map is obtained by characterizing the depth map, and the feature point pairs are grouped according to the depth level map, and the feature point pairs of each group are separately screened to avoid parallax caused by
  • the erroneous deletion of the feature point pairs ensures that sufficient and effective matching feature point pairs can be obtained, which lays a good foundation for the subsequent prediction of the global homography transformation matrix and the optimal homography transformation matrix of each grid, and further improves Accuracy during the registration of the light field.
  • FIG. 1 is a schematic flow chart of a depth-based light field splicing method according to a preferred embodiment of the present invention.
  • a preferred embodiment of the present invention discloses a depth-based light field splicing method, including the following steps:
  • A1 inputting a light field to be spliced and a sub-aperture image of the light field, and performing a light field depth estimation on the sub-aperture image of the light field to obtain a depth map of the light field;
  • the light field to be spliced is input, and the sub-aperture image of the light field is obtained by decoding and pre-processing; and the depth map of the light field is obtained by using the optical field depth estimation for the sub-aperture image of the light field.
  • L r (x, y, u, v) is the reference light field
  • L w (x, y, u, v) is the light field to be spliced
  • S r (u 0 , v 0 ) and S w (u 0 , v 0 ) are subaperture images of the light field at the viewing angle (u 0 , v 0 ).
  • the depth image of the light field is obtained using the light field depth estimation method as D(x, y).
  • A2 extracting feature points in the subaperture image of the light field, matching the feature points to obtain feature point pairs, and screening the feature point pairs to obtain matching feature point pairs;
  • step A2 includes the following steps:
  • A21 extracting feature points in the subaperture image of the light field, and matching the feature points to obtain feature point pairs;
  • the feature points in the subaperture image of the light field are extracted by using the SIFT feature extraction method, and the feature points are matched to obtain a coarse matching feature point pair, namely:
  • S r (u 0 , v 0 ) and S w (u 0 , v 0 ) are sub-aperture images of the light field at the viewing angle (u 0 , v 0 ), and ⁇ F is composed of feature points extracted by SIFT set.
  • A22 performing feature clustering on the depth map to obtain a depth level map of the light field
  • the depth map is layered by using feature clustering, the main depth layer is retained, and small depth changes are discarded, so that the details of the depth map obtained by the depth estimation algorithm may be inaccurate.
  • the method for obtaining the depth map D l of the light field by using the k-means feature clustering method is as follows:
  • S i is the i-th depth layer where the pixel is located, and is generated by clustering:
  • D(x, y) is the depth map obtained by the optical field depth estimation method
  • ⁇ i is the cluster center
  • K is the number of clusters (corresponding to the depth layer in the depth map)
  • D l (x, y) is the depth map obtained.
  • A23 Group the feature point pairs according to the depth hierarchy diagram, and then filter the feature point pairs of each group separately, and combine the selected feature point pairs to obtain matching feature point pairs.
  • the feature point pairs of the rough matching are filtered according to the depth level map.
  • the main steps are: grouping the feature point pairs according to the depth degree map D l (x, y), and then respectively A set of feature points are screened using the Continuous Consistent Sampling Detection (RANSAC) algorithm, the outliers are eliminated, and finally the selected pairs of selected feature points are combined to obtain the final valid matching feature point pairs, namely:
  • RANSAC Continuous Consistent Sampling Detection
  • P is a feature point to be screened
  • S i is the i-th depth layer where the pixel is located
  • K is the number of clusters
  • ⁇ F is the set of feature point pairs extracted by SIFT
  • ⁇ r is the feature point after the screening A collection of pairs.
  • A3 4D meshing the light field to be spliced, predicting the global homography transformation matrix according to the matching feature point pairs; and establishing a weight matrix according to the depth and position relationship between the feature point and the grid center point;
  • the optimal transformation matrix and the weight matrix are used to predict the optimal homography transformation matrix of each grid, and the light field is mapped according to the optimal homography transformation matrix of each grid in the light field;
  • step A3 includes the following steps:
  • A31 4D meshing the light field to be spliced
  • the input light field is segmented into a regular four-dimensional solid mesh to improve the degree of freedom in the light field registration process.
  • step A32 determining whether each of the grids after 4D meshing includes different depth layers, and if so, dividing the network again by the depth layer; otherwise, directly proceeding to step A33;
  • A33 predicting a global homography transformation matrix according to matching feature point pairs
  • A34 establishing a weight matrix according to the depth and position relationship between the feature point and the center of the grid point of each grid;
  • the weight matrix w i is:
  • ⁇ and ⁇ are proportional coefficients
  • ⁇ [0,1] is the minimum threshold of the weight matrix wi
  • (x*,y*) is the position coordinate of the center point of the mesh
  • (x i ,y i ) is the feature point
  • D l is the depth hierarchy in step A2.
  • the D l in the w i formula of the weight matrix may also be calculated by using the depth map D in step A1, that is, the weight matrix w i is:
  • ⁇ and ⁇ are proportional coefficients
  • ⁇ [0,1] is the minimum threshold of the weight matrix wi
  • (x*,y*) is the position coordinate of the center point of the mesh
  • (x i ,y i ) is the feature point
  • D is the depth map of the light field in step A1.
  • A35 predicting an optimal homography transformation matrix of each grid according to a global homography transformation matrix and a weight matrix
  • the method for predicting the optimal homography transformation matrix of each grid by the depth-based light field motion model is as follows:
  • w i is a weight matrix, which is related to the depth and position of the feature point and the grid;
  • is the 5-dimensional light field global homography transformation matrix;
  • the matrix A ⁇ R 4N ⁇ 25 can be obtained by matrix transformation;
  • the matrix A ⁇ R 4N ⁇ 25 is obtained by matrix transformation:
  • a ⁇ R 4N ⁇ 25 has four linear independent row vectors, so at least six pairs of matching feature points are needed. To enhance the robustness, more pairs of matching feature points can be used.
  • A36 The light field is mapped according to the optimal homography transformation matrix of each grid in the light field.
  • each grid is mapped:
  • M is the mesh after dividing the input light field, and the M' mapped grid. It is the optimal homography transformation matrix of the light field corresponding to each grid.
  • the light field is mapped according to the optimal homography transformation matrix of the light field grid, and the pixel coverage area caused by the parallax is determined according to the depth map obtained by the optical field depth in step A1 or the step A2.
  • the depth hierarchy obtained by feature clustering selects the pixel with the smallest depth as the final pixel value of the pixel coverage position.
  • A4 The light field image fusion is obtained, and the light field stitching result is obtained.
  • the light field is fused by using a 4D graph cutting method to obtain a light field splicing result.
  • the 4D map cut is a four-dimensional multi-resolution map cut.
  • the four-dimensional graph cut is specifically: mapping the entire 4D light field to a weighted undirected graph, finding the optimal dividing line to ensure the continuity of the space and angle of the entire light field, so p' is the pixel p in the energy optimization function.
  • the spatial and angular dimensions are adjacent to each other; the multi-resolution map is specifically as follows: first, the light field data is downsampled in spatial resolution, and then the graph cut is performed to obtain a split line at a low resolution, according to low resolution. The dividing line under the rate limits the high-resolution cutting area, and finally the image is cut at high resolution to obtain the optimal suture.
  • the 4D graph-cut is specifically: firstly, the 4D light field is mapped into a weighted undirected graph, and then the energy optimization function is calculated:
  • multi-resolution 4D image cutting is adopted, and the specific steps are: first, down-sampling the spatial resolution of the light field, and then performing The graph cuts the partition line at low resolution, limits the high-resolution map cut region according to the split line at low resolution, and finally performs the graph cut to obtain the optimal suture at high resolution.
  • the light field splicing method of the present invention combines the idea of using a local homography transformation matrix to replace the global homography transformation matrix, which significantly improves the flexibility in the light field registration process, thereby achieving more accurate light field splicing in the detail part.
  • the problem of fruit dislocation and ghosting caused by large parallax changes is solved, and a precise parallax-tolerant light field splicing method is realized.
  • the depth map estimated by the light field camera's own light field data guides the screening of the feature point pairs, avoiding the erroneous deletion of the feature point pairs due to the parallax, thereby ensuring sufficient and effective matching feature point pairs.
  • the image cutting algorithm is used to find the optimal stitching to realize the light field fusion, and further correct the small misalignment generated during the stitching process to achieve more accurate light field stitching.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Processing (AREA)

Abstract

一种基于深度的光场拼接方法,包括:输入待拼接的光场以及该光场的子孔径图像,对光场的子孔径图像进行光场深度估计得到光场的深度图;提取光场的子孔径图像中的特征点,对特征点进行匹配得到特征点对,对特征点对进行筛选得到匹配特征点对;将待拼接的光场进行4D网格化,根据匹配特征点对预测全局单应性变换矩阵;并根据特征点与网格中心点的深度和位置关系建立权值矩阵;再根据全局单应性变换矩阵和权值矩阵来预测每个网格的最优单应性变换矩阵,并根据光场中每个网格的最优单应性变换矩阵映射光场对光场进行融合,得到光场拼接结果。该方法解决了较大视差变化造成结果错位和重影的问题,实现了精确的视差容忍的光场拼接。

Description

一种基于深度的光场拼接方法 技术领域
本发明涉及计算机视觉与数字图像处理领域,尤其涉及一种基于深度的光场拼接方法。
背景技术
光场成像通过在主透镜和传感器中间加入一个微透镜阵列,记录了来自各个方向的光线信息,从而得到完整信息的光辐射场。随着光学传感器的分辨率持续上升和光场相机逐步市场化,光场成像技术的实用价值变得越来越高。相比于传统的数码相机,基于微透镜阵列的光场相机由于可以同时记录三维场景的空间位置信息和方向信息,因此基于该光场相机采集到的光场数据有着广泛的应用,如重聚焦、深度估计、显著性检验等等。然而由于光学结构的限制,手持式全光相机的视场角较小。不依赖于光场结构设置的光场拼接方法可以提高光场相机的视场角。
现有的光场拼接方法主要是通过特征提取和匹配的方法,计算相邻输入光场间的变换矩阵,进行光场配准,通过构造配准后光场重叠区域的能量损失函数来寻找光场的最优缝合线,实现光场融合;这种方法存在一定的局限性,虽然其可以实现视差较小的光场拼接,一旦输入光场数据视差变化较大,将会产生错位,重影等错误,无法得到正确的拼接结果。
另外一种方法是利用结合视差容忍的图像拼接方法来减小视差对拼接结果带来的影响,这种方式在一定程度上提高了拼接的精确度,但是不能完全抵消视差带来的影响,且单独对光场相应子孔径图像进行图像拼接会引入整个光场在空间域和角度域不一致的问题。
以上背景技术内容的公开仅用于辅助理解本发明的构思及技术方案,其并不必然属于本专利申请的现有技术,在没有明确的证据表明上述内容在本专利申请的申请日已经公开的情况下,上述背景技术不应当用于评价本申请的新颖性和创造性。
发明内容
为了解决上述技术问题,本发明提出一种基于深度的光场拼接方法,解决了较大视差变化造成结果错位和重影的问题,实现了精确的视差容忍的光场拼接方法。
为了达到上述目的,本发明采用以下技术方案:
本发明公开了一种基于深度的光场拼接方法,包括以下步骤:
A1:输入待拼接的光场以及该光场的子孔径图像,对光场的子孔径图像进行光场深度估计得到光场的深度图;
A2:提取光场的子孔径图像中的特征点,对特征点进行匹配得到特征点对,对特征点对进行筛选得到匹配特征点对;
A3:将待拼接的光场进行4D网格化,根据匹配特征点对预测全局单应性变换矩阵;并根据特征点与网格中心点的深度和位置关系建立权值矩阵;再根据全局单应性变换矩阵和权值矩阵来预测每个网格的最优单应性变换矩阵,并根据光场中每个网格的最优单应性变换矩阵映射光场;
A4:对光场进行融合,得到光场拼接结果。
进一步地,步骤A2具体包括:
A21:提取光场的子孔径图像中的特征点,对特征点进行匹配得到特征点对;
A22:对所述深度图进行特征聚类得到光场的深度层次图;
A23:根据深度层次图对特征点对进行分组,分别对各组的特征点对进行筛选,合并筛选后的特征点对得到匹配特征点对。
进一步地,步骤A3中具体包括:
A31:将待拼接的光场进行4D网格化;
A32:判断4D网格化后的每一个网格中是否包含不同深度层,如果是,则按深度层再次分割网络;否则直接进入步骤A33;
A33:根据匹配特征点对预测全局单应性变换矩阵;
A34:根据特征点与网格点中心的深度和位置关系建立权值矩阵;
A35:根据全局单应性变换矩阵和权值矩阵来预测每个网格的最优单应性变换矩阵;
A36:根据光场中每个网格的最优单应性变换矩阵映射光场。
与现有技术相比,本发明的有益效果在于:本发明的基于深度的光场拼接方法采用每个网格的最优单应性变换矩阵替代全局单应性变换矩阵来对光场进行映射,显著提升了光场配准过程中的灵活度,实现细节部分更精确的光场拼接;而且其中的每个网格的最优单应性变换矩阵是通过结合全局单应性变换矩阵以及引入了深度图而建立的权值矩阵来进行预测的,解决了较大视差变化造成结果错位和重影的问题,实现了精确的视差容忍的光场拼接方法;从而进一步实现更精准的大视差光场拼接,且保证了拼接光场空间域和角度域一致性,从而扩大光场的视角。
在进一步的方案中,通过对深度图进行特征聚类得到深度层次图,并根据该深度层次图来指导特征点对进行分组,再对各组的特征点对分别进行筛选,避免了由于视差导致的特征点对的误删除,从而保证能够得到足够且有效的匹配特征点对,为后续预测全局单应性变换矩阵以及每个网格的最优单应性变换矩阵打下良好基础,更进一步提高光场配准过程中的准确性。
附图说明
图1是本发明优选实施例的基于深度的光场拼接方法的流程示意图。
具体实施方式
下面对照附图并结合优选的实施方式对本发明作进一步说明。
如图1所示,本发明优选实施例公开了一种基于深度的光场拼接方法,包括以下步骤:
A1:输入待拼接的光场以及光场的子孔径图像,对光场的子孔径图像进行光场深度估计得到光场的深度图;
具体地,输入待拼接的光场,进行解码和预处理得到光场的子孔径图像;并对光场的子孔径图像使用光场深度估计得到光场的深度图。
在本实施例中,以两个光场的拼接为例,L r(x,y,u,v)为参考光场,L w(x,y,u,v)为待拼接光场,多个光场的拼接可以使用相同的方法扩展。对输入的光场数据进行解码和预处理,得到光场的子孔径图像方法如下:
S r(u 0,v 0)={L r(x,y,u,v)|u=u 0,v=v 0}
S w(u 0,v 0)={L w(x,y,u,v)|u=u 0,v=v 0}
其中S r(u 0,v 0)和S w(u 0,v 0)是光场在视角(u 0,v 0)处的子孔径图像。
在本实施例中,使用光场深度估计方法得到光场的深度图像为D(x,y)。
A2:提取光场的子孔径图像中的特征点,对特征点进行匹配得到特征点对,对特征点对进行筛选得到匹配特征点对;
具体地,步骤A2包括以下步骤:
A21:提取光场的子孔径图像中的特征点,对特征点进行匹配得到特征点对;
本实施例中,使用SIFT特征提取方法提取光场的子孔径图像中的特征点,并对特征点进行匹配,得到粗匹配的特征点对,即:
Figure PCTCN2019085643-appb-000001
其中,S r(u 0,v 0)和S w(u 0,v 0)是光场在视角(u 0,v 0)处的子孔径图像,Ω F为SIFT提取的特征点对组成的集合。
A22:对所述深度图进行特征聚类得到光场的深度层次图;
在本实施例中,通过使用特征聚类的方法对深度图进行分层,保留主要的深度层,舍弃小的深度变化,从而可以避免深度估计算法得到的深度图的细节部分可能存在不准确的区域而对光场拼接造成的影响。
本实施例中,使用k-means特征聚类的方法得到光场的深度层次图D l的方法如下:
Figure PCTCN2019085643-appb-000002
其中S i为像素所在的第i深度层,通过聚类的方法生成:
Figure PCTCN2019085643-appb-000003
其中D(x,y)是使用光场深度估计方法得到的深度图,μ i是聚类中心,K是聚类个数(对应于深度层次图里面的深度层数),D l(x,y)是得到的深度层次图。
A23:根据深度层次图对特征点对进行分组,再分别对各组的特征点对进行 筛选,合并筛选后的特征点对得到匹配特征点对。
在本实施例中,基于深度层次图的指导下对粗匹配的特征点对进行筛选,主要步骤为:根据深度度层次图D l(x,y)对特征点对进行分组,然后分别对每一组的特征点对使用连续一致抽样检测(RANSAC)算法进行筛选,剔除离群点,最后合并每一组筛选后的特征点对,得到最终有效的匹配特征点对,即:
Figure PCTCN2019085643-appb-000004
其中,P是一对待筛选的特征点,S i为像素所在的第i深度层,K是聚类个数,Ω F为SIFT提取的特征点对组成的集合,Ω r为筛选过后的特征点对组成的集合。
A3:将待拼接的光场进行4D网格化,根据匹配特征点对预测全局单应性变换矩阵;并根据特征点与网格中心点的深度和位置关系建立权值矩阵;再根据全局单应性变换矩阵和权值矩阵来预测每个网格的最优单应性变换矩阵,根据光场中每个网格的最优单应性变换矩阵映射光场;
具体地,步骤A3包括以下步骤:
A31:将待拼接的光场进行4D网格化;
在本实施例中,将输入光场分割成规则的四维立体网格,提升光场配准过程中的自由度。
A32:判断4D网格化后的每一个网格中是否包含不同深度层,如果是,则按深度层再次分割网络;否则直接进入步骤A33;
A33:根据匹配特征点对预测全局单应性变换矩阵;
在本实施例中,预测的5维光场全局单应性变换矩阵Η的计算公式为:
P'=HP
其中P(u,v,x,y,1)和P'(u',v',x',y',1)是通过步骤A2所得到的匹配特征点对;
也即Η的计算公式为:
Figure PCTCN2019085643-appb-000005
A34:根据特征点与每个网格的网格点中心的深度和位置关系建立权值矩阵;
在本实施例中,权值矩阵w i为:
Figure PCTCN2019085643-appb-000006
其中α、β是比例系数,η∈[0,1]是权值矩阵 wi的最小阈值,(x*,y*)是网格中心点的位置坐标,(x i,y i)是特征点所在的位置坐标,D l是步骤A2中的深度层次图。
在其他一些实施例中,上述权值矩阵的w i公式中的D l也可以采用步骤A1中的深度图D来进行计算,也即权值矩阵w i为:
Figure PCTCN2019085643-appb-000007
其中α、β是比例系数,η∈[0,1]是权值矩阵 wi的最小阈值,(x*,y*)是网格中心点的位置坐标,(x i,y i)是特征点所在的位置坐标,D是步骤A1中的光场的深度图。
A35:根据全局单应性变换矩阵和权值矩阵来预测每个网格的最优单应性变换矩阵;
在本实施例中,通过基于深度的光场运动模型预测每个网格的最优单应性变换矩阵方法如下:
Figure PCTCN2019085643-appb-000008
其中w i是权值矩阵,与特征点和网格所在的深度及位置有关;Η是5维光场全局单应性变换矩阵;矩阵A∈R 4N×25可以通过矩阵变换得到;
在本实施例中,矩阵A∈R 4N×25由矩阵变换得到:
04×1=A×Η
即:
Figure PCTCN2019085643-appb-000009
其中A∈R 4N×25有四个线性独立的行向量,因此最少需要六对匹配的特征点,为了增强鲁棒性,可以使用更多对的匹配特征点。
A36:根据光场中每个网格的最优单应性变换矩阵映射光场。
在本实施例中,对每一个网格进行映射:
Figure PCTCN2019085643-appb-000010
其中M是将输入的光场分割后的网格,M'映射后的网格,
Figure PCTCN2019085643-appb-000011
是每个网格对应的光场最优单应性变换矩阵。
在本实施例中,根据光场网格最优单应性变换矩阵映射光场,针对视差造成的像素覆盖区域,则根据步骤A1中的光场深度估计得到的深度图或者步骤A2中的进行特征聚类得到的深度层次图挑选深度最小的像素作为像素覆盖位置的最终像素值。
A4:对光场图像融合,得到光场拼接结果。
具体地,采用4D图割方法对光场进行融合,得到光场拼接结果。
其中,4D图割是四维多分辨率图割。四维图割具体为:将整个4D光场映射到一个带权无向图,寻找最优的分割线使其保证整个光场的空间和角度的连续性,因此能量优化函数中p'是像素p空间维度和角度维度均相邻的像素;多分辨率图割具体为:首先对光场数据进行空间分辨率上的降采样,然后进行图割,得到低分辨率下的分割线,根据低分辨率下的分割线限制高分辨率的图割区域,最终在高分辨率下进行图割得到最优缝合线。
在本实施例中4D graph-cut具体为:首先将4D光场映射到一个带权无向图中,其次计算能量优化函数:
Figure PCTCN2019085643-appb-000012
其中p、p'是光场重叠区域中相邻的像素,R(p)是区域项,B(p,p')是边界项。最后通过最小化能量优化函数找到光场拼接的最优缝合线,实现光场融合。
在本实施例中,为了加快4D图割(4D graph-cut)处理4D光场,采用多分辨率的4D图割,具体步骤为:首先对光场进行空间分辨率上的降采样,然后进行图割,得到低分辨率下的分割线,根据低分辨率下的分割线限制高分辨率的图割区域,最终在高分辨率下进行图割得到最优缝合线。
本发明的光场拼接方法结合了使用局部单应性变换矩阵来替代全局单应性变换矩阵的想法,显著提升了光场配准过程中的灵活度,从而实现细节部分更精确的光场拼接;并结合了通过引入深度图配合全局单应性变换矩阵的使用,解决了较大视差变化造成的果错位和重影的问题,实现了精确的视差容忍的光场拼接方法。通过光场相机自身的光场数据估计出的深度图来指导特征点对的筛选,避免了由于视差导致的特征点对的误删除,从而保证得到足够且有效的匹配特征点对。通过图割算法寻找最优缝合线实现光场融合,进一步修正拼接过程中产生的小的错位,实现更精确的光场拼接。
以上内容是结合具体的优选实施方式对本发明所作的进一步详细说明,不能认定本发明的具体实施只局限于这些说明。对于本发明所属技术领域的技术人员来说,在不脱离本发明构思的前提下,还可以做出若干等同替代或明显变型,而且性能或用途相同,都应当视为属于本发明的保护范围。

Claims (10)

  1. 一种基于深度的光场拼接方法,其特征在于,包括以下步骤:
    A1:输入待拼接的光场以及该光场的子孔径图像,对光场的子孔径图像进行光场深度估计得到光场的深度图;
    A2:提取光场的子孔径图像中的特征点,对特征点进行匹配得到特征点对,对特征点对进行筛选得到匹配特征点对;
    A3:将待拼接的光场进行4D网格化,根据匹配特征点对预测全局单应性变换矩阵;并根据特征点与网格中心点的深度和位置关系建立权值矩阵;再根据全局单应性变换矩阵和权值矩阵来预测每个网格的最优单应性变换矩阵,并根据光场中每个网格的最优单应性变换矩阵映射光场;
    A4:对光场进行融合,得到光场拼接结果。
  2. 根据权利要求1所述的基于深度的光场拼接方法,其特征在于,步骤A2具体包括:
    A21:提取光场的子孔径图像中的特征点,对特征点进行匹配得到特征点对;
    A22:对所述深度图进行特征聚类得到光场的深度层次图;
    A23:根据深度层次图对特征点对进行分组,分别对各组的特征点对进行筛选,合并筛选后的特征点对得到匹配特征点对。
  3. 根据权利要求2所述的基于深度的光场拼接方法,其特征在于,
    其中步骤A21具体包括:使用SIFT特征提取方法提取光场的子孔径图像中的特征点,并对特征点进行匹配,得到粗匹配的特征点对:
    Figure PCTCN2019085643-appb-100001
    其中,S r(u 0,v 0)和S w(u 0,v 0)是光场在视角(u 0,v 0)处的子孔径图像,Ω F为SIFT提取的特征点对组成的集合;
    优选地,步骤A23具体包括:根据深度层次图D l(x,y)对特征点对进行分组,然后分别对每一组的特征点对使用连续一致抽样检测算法进行筛选,剔除离群点,最后合并每一组筛选后的特征点对,得到最终有效的匹配特征点对:
    Figure PCTCN2019085643-appb-100002
    其中,P是一对待筛选的特征点,S i为像素所在的第i深度层,K是聚类个数,Ω F为SIFT提取的特征点对组成的集合,Ω r为筛选过后的特征点对组成的集合。
  4. 根据权利要求2所述的基于深度的光场拼接方法,其特征在于,其中步骤A22中对所述深度图进行特征聚类得到光场的深度层次图D l为:
    Figure PCTCN2019085643-appb-100003
    其中,
    Figure PCTCN2019085643-appb-100004
    S i为像素所在的第i深度层,D(x,y)是步骤A1中的光场的深度图,μ i是聚类中心,K是聚类个数。
  5. 根据权利要求1所述的基于深度的光场拼接方法,其特征在于,步骤A3中根据匹配特征点对预测的全局单应性变换矩阵Η的计算公式为:
    P'=HP
    其中P(u,v,x,y,1)和P'(u',v',x',y',1)是通过步骤A2所得到的匹配特征点对。
  6. 根据权利要求1所述的基于深度的光场拼接方法,其特征在于,步骤A3中根据特征点与网格中心点的深度和位置关系建立权值矩阵w i为:
    Figure PCTCN2019085643-appb-100005
    其中α、β是比例系数,η∈[0,1]是权值矩阵 wi的最小阈值,(x*,y*)是网格中心点的位置坐标,(x i,y i)是特征点所在的位置坐标,D是步骤A1中的光场的深度图。
  7. 根据权利要求2至5任一项所述的基于深度的光场拼接方法,其特征在于,步骤A3中根据特征点与网格中心点的深度和位置关系建立权值矩阵w i为:
    Figure PCTCN2019085643-appb-100006
    其中α、β是比例系数,η∈[0,1]是权值矩阵 wi的最小阈值,(x*,y*)是网格中心点的位置坐标,(x i,y i)是特征点所在的位置坐标,D l是步骤A2中的光场 的深度层次图。
  8. 根据权利要求1所述的基于深度的光场拼接方法,其特征在于,步骤A3中根据全局单应性变换矩阵和权值矩阵来预测每个网格的最优单应性变换矩阵
    Figure PCTCN2019085643-appb-100007
    为:
    Figure PCTCN2019085643-appb-100008
    其中,w i是权值矩阵,Η是全局单应性变换矩阵,矩阵A∈R 4N×25由矩阵变换0 4×1=A×Η得到。
  9. 根据权利要求2至5任一项所述的基于深度的光场拼接方法,其特征在于,步骤A3中根据光场中每个网格的最优单应性变换矩阵映射光场时,针对视差造成的像素覆盖区域,根据步骤A1中的深度图或者步骤A2中的深度层次图挑选深度最小的像素作为像素覆盖位置的最终像素值。
  10. 根据权利要求1所述的基于深度的光场拼接方法,其特征在于,步骤A3中具体包括:
    A31:将待拼接的光场进行4D网格化;
    A32:判断4D网格化后的每一个网格中是否包含不同深度层,如果是,则按深度层再次分割网络;否则直接进入步骤A33;
    A33:根据匹配特征点对预测全局单应性变换矩阵;
    A34:根据特征点与网格点中心的深度和位置关系建立权值矩阵;
    A35:根据全局单应性变换矩阵和权值矩阵来预测每个网格的最优单应性变换矩阵;
    A36:根据光场中每个网格的最优单应性变换矩阵映射光场。
PCT/CN2019/085643 2018-05-07 2019-05-06 一种基于深度的光场拼接方法 WO2019214568A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810428591.0 2018-05-07
CN201810428591.0A CN108921781B (zh) 2018-05-07 2018-05-07 一种基于深度的光场拼接方法

Publications (1)

Publication Number Publication Date
WO2019214568A1 true WO2019214568A1 (zh) 2019-11-14

Family

ID=64402352

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/085643 WO2019214568A1 (zh) 2018-05-07 2019-05-06 一种基于深度的光场拼接方法

Country Status (2)

Country Link
CN (1) CN108921781B (zh)
WO (1) WO2019214568A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111340701A (zh) * 2020-02-24 2020-06-26 南京航空航天大学 一种基于聚类法筛选匹配点的电路板图像拼接方法
CN111507904A (zh) * 2020-04-22 2020-08-07 华中科技大学 一种微观打印图案的图像拼接方法及装置
CN111882487A (zh) * 2020-07-17 2020-11-03 北京信息科技大学 一种基于双平面平移变换的大视场光场数据融合方法
CN113191369A (zh) * 2021-04-09 2021-07-30 西安理工大学 一种基于光场角度域变化矩阵的特征点检测方法
CN113506214A (zh) * 2021-05-24 2021-10-15 南京莱斯信息技术股份有限公司 一种多路视频图像拼接方法
CN116934591A (zh) * 2023-06-28 2023-10-24 深圳市碧云祥电子有限公司 多尺度特征提取的图像拼接方法、装置、设备及存储介质
CN117221466A (zh) * 2023-11-09 2023-12-12 北京智汇云舟科技有限公司 基于网格变换的视频拼接方法及系统

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108921781B (zh) * 2018-05-07 2020-10-02 清华大学深圳研究生院 一种基于深度的光场拼接方法
CN110084749B (zh) * 2019-04-17 2023-03-31 清华大学深圳研究生院 一种焦距非一致的光场图像的拼接方法
CN110264403A (zh) * 2019-06-13 2019-09-20 中国科学技术大学 一种基于图像深度分层的去伪影图像拼接方法
CN110930310B (zh) * 2019-12-09 2023-04-07 中国科学技术大学 全景图像拼接方法
CN111161143A (zh) * 2019-12-16 2020-05-15 首都医科大学 一种光学定位技术辅助的手术视野全景拼接方法
CN112465704B (zh) * 2020-12-07 2024-02-06 清华大学深圳国际研究生院 一种全局-局部自适应优化的全景光场拼接方法
CN114373153B (zh) * 2022-01-12 2022-12-27 北京拙河科技有限公司 一种基于多尺度阵列相机的视频成像优化系统与方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106873301A (zh) * 2017-04-21 2017-06-20 北京理工大学 基于阵列相机对远距离小孔后方进行成像的系统及方法
CN106886979A (zh) * 2017-03-30 2017-06-23 深圳市未来媒体技术研究院 一种图像拼接装置及图像拼接方法
CN107403423A (zh) * 2017-08-02 2017-11-28 清华大学深圳研究生院 一种光场相机的合成孔径去遮挡方法
CN107578376A (zh) * 2017-08-29 2018-01-12 北京邮电大学 基于特征点聚类四叉划分和局部变换矩阵的图像拼接方法
CN108921781A (zh) * 2018-05-07 2018-11-30 清华大学深圳研究生院 一种基于深度的光场拼接方法

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CA2369710C (en) * 2002-01-30 2006-09-19 Anup Basu Method and apparatus for high resolution 3d scanning of objects having voids
CN101394573B (zh) * 2008-10-30 2010-06-16 清华大学 一种基于特征匹配的全景图生成方法及系统
CN101923709B (zh) * 2009-06-16 2013-06-26 日电(中国)有限公司 图像拼接方法与设备
CN102833487B (zh) * 2012-08-08 2015-01-28 中国科学院自动化研究所 面向视觉计算的光场成像装置和方法
US9332243B2 (en) * 2012-10-17 2016-05-03 DotProduct LLC Handheld portable optical scanner and method of using
US8978984B2 (en) * 2013-02-28 2015-03-17 Hand Held Products, Inc. Indicia reading terminals and methods for decoding decodable indicia employing light field imaging
CN106791869B (zh) * 2016-12-21 2019-08-27 中国科学技术大学 基于光场子孔径图像相对位置关系的快速运动搜索方法
CN106526867B (zh) * 2017-01-22 2018-10-30 网易(杭州)网络有限公司 影像画面的显示控制方法、装置及头戴式显示设备
CN107295264B (zh) * 2017-08-01 2019-09-06 清华大学深圳研究生院 一种基于单应性变换光场数据压缩方法
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

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106886979A (zh) * 2017-03-30 2017-06-23 深圳市未来媒体技术研究院 一种图像拼接装置及图像拼接方法
CN106873301A (zh) * 2017-04-21 2017-06-20 北京理工大学 基于阵列相机对远距离小孔后方进行成像的系统及方法
CN107403423A (zh) * 2017-08-02 2017-11-28 清华大学深圳研究生院 一种光场相机的合成孔径去遮挡方法
CN107578376A (zh) * 2017-08-29 2018-01-12 北京邮电大学 基于特征点聚类四叉划分和局部变换矩阵的图像拼接方法
CN108921781A (zh) * 2018-05-07 2018-11-30 清华大学深圳研究生院 一种基于深度的光场拼接方法

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111340701A (zh) * 2020-02-24 2020-06-26 南京航空航天大学 一种基于聚类法筛选匹配点的电路板图像拼接方法
CN111340701B (zh) * 2020-02-24 2022-06-28 南京航空航天大学 一种基于聚类法筛选匹配点的电路板图像拼接方法
CN111507904A (zh) * 2020-04-22 2020-08-07 华中科技大学 一种微观打印图案的图像拼接方法及装置
CN111507904B (zh) * 2020-04-22 2023-06-02 华中科技大学 一种微观打印图案的图像拼接方法及装置
CN111882487A (zh) * 2020-07-17 2020-11-03 北京信息科技大学 一种基于双平面平移变换的大视场光场数据融合方法
CN113191369A (zh) * 2021-04-09 2021-07-30 西安理工大学 一种基于光场角度域变化矩阵的特征点检测方法
CN113191369B (zh) * 2021-04-09 2024-02-09 西安理工大学 一种基于光场角度域变化矩阵的特征点检测方法
CN113506214A (zh) * 2021-05-24 2021-10-15 南京莱斯信息技术股份有限公司 一种多路视频图像拼接方法
CN113506214B (zh) * 2021-05-24 2023-07-21 南京莱斯信息技术股份有限公司 一种多路视频图像拼接方法
CN116934591A (zh) * 2023-06-28 2023-10-24 深圳市碧云祥电子有限公司 多尺度特征提取的图像拼接方法、装置、设备及存储介质
CN117221466A (zh) * 2023-11-09 2023-12-12 北京智汇云舟科技有限公司 基于网格变换的视频拼接方法及系统
CN117221466B (zh) * 2023-11-09 2024-01-23 北京智汇云舟科技有限公司 基于网格变换的视频拼接方法及系统

Also Published As

Publication number Publication date
CN108921781A (zh) 2018-11-30
CN108921781B (zh) 2020-10-02

Similar Documents

Publication Publication Date Title
WO2019214568A1 (zh) 一种基于深度的光场拼接方法
CN112435325B (zh) 基于vi-slam和深度估计网络的无人机场景稠密重建方法
CN107659774B (zh) 一种基于多尺度相机阵列的视频成像系统及视频处理方法
CA3121440C (en) Assembly body change detection method, device and medium based on attention mechanism
CN108074218B (zh) 基于光场采集装置的图像超分辨率方法及装置
CN109064410B (zh) 一种基于超像素的光场图像拼接方法
CN111343367B (zh) 一种十亿像素虚拟现实视频采集装置、系统与方法
CN109816708B (zh) 基于倾斜航空影像的建筑物纹理提取方法
CN107316275A (zh) 一种光流辅助的大尺度显微图像拼接算法
CN115205489A (zh) 一种大场景下的三维重建方法、系统及装置
CN109118544B (zh) 基于透视变换的合成孔径成像方法
CN107341815B (zh) 基于多目立体视觉场景流的剧烈运动检测方法
JP6174104B2 (ja) 室内2d平面図の生成方法、装置及びシステム
CN105005964A (zh) 基于视频序列影像的地理场景全景图快速生成方法
CN115953535A (zh) 三维重建方法、装置、计算设备和存储介质
CN111860651B (zh) 一种基于单目视觉的移动机器人半稠密地图构建方法
CN108663026A (zh) 一种振动测量方法
WO2021035627A1 (zh) 获取深度图的方法、装置及计算机存储介质
CN112465704A (zh) 一种全局-局部自适应优化的全景光场拼接方法
CN115619623A (zh) 一种基于移动最小二乘变换的平行鱼眼相机图像拼接方法
Gao et al. Multi-source data-based 3D digital preservation of largescale ancient chinese architecture: A case report
CN113436130A (zh) 一种非结构光场智能感知系统与装置
CN116132610A (zh) 一种综采面视频拼接方法及系统
CN108090930A (zh) 基于双目立体相机的障碍物视觉检测系统及方法
Zhou et al. Video stabilization and completion using two cameras

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: 19800597

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: 19800597

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 01.04.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19800597

Country of ref document: EP

Kind code of ref document: A1