CN104867129A - Light field image segmentation method - Google Patents

Abstract

The invention discloses a light field image segmentation method, comprising the following steps: (100) parameterizing the collected light field information; (200) marking objects in any view in the light field; (300) selecting Objects are trained using machine learning methods to obtain a classifier; (400) using the classifier to segment the view of the entire light field. Through the above method, the present invention can directly use the geometric information contained in the light field information itself, the calculation amount is small, and a good segmentation effect can be achieved for various scenes.

Description

A Light Field Image Segmentation Method

technical field

The invention relates to the fields of image processing, light field and pattern recognition, in particular to a light field image segmentation method.

Background technique

In recent years, the progress made in the light field acquisition system has made it possible for light field technology to become the core technical means in the field of graphics and image technology in the future, and the era of light field photography is coming. Compared with a single image, the light field captures more directional scene information, which makes traditional image processing algorithms, computer vision, and some new related science and technology, etc., need to adapt to the new light field technology. With the development of light field technology, traditional image segmentation methods have great room for improvement. Since the light field information contains the geometric information of the scene, it seems possible to directly use the geometric information for image segmentation. At present, the image segmentation technology is still in the stage based on traditional images, and often only uses pixel color, depth, grayscale and other information. However, the research on light field image segmentation technology has just started.

At present, there are still many problems to be solved about image segmentation and light field image segmentation at home and abroad:

1) Traditional image segmentation cannot perform good analysis, identification, prediction, and segmentation when there are large differences between different types and small differences between the same type. For example, it cannot achieve good results when segmenting images of leaves, plants, etc.

2) When different objects have similar appearances, such as wooden walls and wooden benches, due to the loss of the geometric information of the image, traditional image segmentation technology is difficult to distinguish them, and the algorithm complexity is high. The effect is not satisfactory, and the accuracy is not high; in addition, the technology of geometric information restoration through traditional images has a large computational time complexity, and the results are often inaccurate, which cannot provide effective help for traditional image segmentation.

Contents of the invention

The main technical problem to be solved by the present invention is to provide a light field image segmentation method aimed at the deficiencies of the prior art, which can directly use the geometric information contained in the light field information itself, has a small amount of calculation, and can achieve Better segmentation effect.

In order to solve the above technical problems, a technical solution adopted by the present invention is to provide a light field image segmentation method, comprising the following steps:

(100) Parameterizing the collected light field information;

(200) Labeling objects in any view in the light field;

(300) Using a machine learning method to train the labeled objects to obtain a classifier;

(400) Segment the view of the entire light field with a classifier.

In a preferred embodiment of the present invention, the step (100) specifically includes: simply changing the coordinates of the light field information on the basis of lumens, which is described as follows:

If the ray L[s, t, x, y] is a ray defined by (s, t)∈Π, (x, y)∈Ω, (x, y) is the connection between the object and its vacuum projection on the plane Π The intersection of the line and the plane Ω, on the plane Π, x is consistent with s, and y is consistent with t; thus, the nuclear surface Ly, t and Lx, s of the light field are obtained, that is, the horizontal section and vertical section of the light field information; because There is a linear relationship between the viewing angle and the nuclear plane, which introduces the parameter "inconsistency", that is: the depth of the point projection in the scene on the plane determines the rate of change of the image in the view; thus, the coordinate information of the light field is normalized, and the introduction of Parameter "inconsistency".

In a preferred embodiment of the present invention, the step (200) specifically includes: selecting any view in the light field as a training sample, and marking different objects in the view with lines.

In a preferred embodiment of the present invention, the step (300) is specifically as follows: for the labeled object, select: RGB value, Hessian feature value, intensity standard deviation and inconsistency attribute in the image as training input, and use this get the classifier.

In a preferred embodiment of the present invention, the step (400) specifically includes: after the classification is completed, performing a grid search on the smallest segmented node, and remerging the over-refined classifications into the same classification.

The beneficial effects of the present invention are:

1) The present invention provides an image segmentation method for light field, so that the computer can directly use the geometric information contained in the light field information to perform image segmentation, and make full use of the characteristics of the light field to achieve a high-quality segmentation effect ;

2) For the light field of a certain scene, the present invention only needs to manually mark any view, and then conduct training to segment any view of the light field, with a small amount of calculation, low training cost and high efficiency;

3) The present invention can achieve ideal effects for various scenarios and has good adaptability;

4) When the present invention is used for image segmentation training, the selected training attribute set elements are less than traditional image segmentation, which further reduces the computational complexity and is suitable for large-scale image segmentation.

Description of drawings

Fig. 1 is the flowchart of a kind of light field image segmentation method of the present invention;

Fig. 2 is a view annotation diagram of a light field image segmentation method of the present invention;

Fig. 3 is a light field parameterized coordinate diagram of a light field image segmentation method of the present invention.

Detailed ways

The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, so as to define the protection scope of the present invention more clearly.

Please refer to Fig. 1-3, the embodiment of the present invention comprises:

A light field image segmentation method, comprising the following steps:

(100) Parameterize the collected light field information:

In the present invention, the parameterization of light field information is transformed on the basis of Lumigraph.

A 4D light field is defined in the ray space R, a group of rays emitted by the light point P (X, Y, Z) passes through two parallel planes Π and Ω, in the coordinate system R ³ , so that each ray L can be defined by the intersection points (s, t) and (x, y) of L with plane Π and plane Ω. The distance between plane Π and plane Ω is f > 0, and the respective coordinate systems are s, t and x, y; the unit vectors of the two coordinate systems are parallel, and the origin is on a straight line perpendicular to the two planes.

A ray L ₁ [s ₁ ,t,x ₁ ,y] is defined by (s ₁ ,t)∈Π，(x ₁ ,y)∈Ω, L ₁ [s ₁ ,t,0,0] is perpendicular to The light on the plane Π passing through the light point (s ₁ , t), similarly, on the plane Ω, x ₁ corresponds to s ₁ and y corresponds to t.

Another ray L ₂ [s ₂ ,t,x ₂ ,y] is defined by (s ₂ ,t)∈Π, (x ₂ ,y)∈Ω, and L ₂ [s ₂ ,t,0,0] is the vertical The light on the plane Π passing through the light point (s ₂ , t), similarly, on the plane Ω, x ₂ corresponds to s ₂ , and y corresponds to t.

And the distance between s ₁ and s ₂ is △s

Then the distance between x ₁ and x ₂ x ₂ - x ₁ = Δs,

Now, a light field can be represented as a function in the ray square:

Ly*,t*:(x,s) →L(x,y*,s,t*)

Ly*, t* and Lx*, s* are the nuclear planes of the light field, and they can be regarded as the horizontal section and the vertical section of the light field.

In a certain view of the same scene, the rate of change of the image depends on the depth of the scene projected on the planar image, that is, disparity.

So far, the parameterization of the light field is completed, and the parameters introduced are: nuclear surface, inconsistency.

(200) label arbitrary views of light field information:

Pick a view from any angle in the light field and annotate the different objects in that view with lines.

(300) For the objects marked in the view, the machine learning method (random forest method) is used for training:

Use the machine learning method (random forest method) to train the marked objects in the selected view, select the RGB value, Hessian eigenvalue, intensity standard deviation, disparity (inconsistency) attribute in the image as training input, and obtain Classifier.

In machine learning, a random forest is a classifier that contains multiple decision trees, and its output category is determined by the mode of the category output by individual trees. Leo Breiman and Adele Cutler developed algorithms to infer random forests. And "Random Forests" is their trademark. This term comes from random decision forests proposed by Tin Kam Ho of Bell Labs in 1995. This method combines Breimans' "Bootstrap aggregating" idea with Ho's "random subspace method" to build a collection of decision trees.

(400) classify the entire light field:

A classifier is used to segment the view of the entire light field. After the classification is over, a grid search is performed on the smallest segmented node, and the over-refined classifications are refused into the same classification to prevent the negative impact of over-classification on the results.

The invention discloses a light field image segmentation method, which has a small amount of calculation, is suitable for a light field with a large amount of data, fully utilizes the geometric information of the scene carried in the light field information, has a good segmentation effect, and improves the traditional image segmentation technology At the same time, it also caters to the development trend of light field technology. It can be applied to pattern recognition based on light field technology, video surveillance, image processing, etc.

The above is only an embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the description of the present invention and the contents of the accompanying drawings, or directly or indirectly used in other related technologies fields, all of which are equally included in the scope of patent protection of the present invention.

Claims (5)

1. a light field image dividing method, is characterized in that, comprises the following steps:

(100) parametrization is carried out to the field information gathered;

(200) object in view any in light field is marked;

(300) to the object of mark, use machine learning method to train, thus obtain sorter;

(400) split with the view of sorter to whole light field.

2. a kind of light field image dividing method according to claim 1, is characterized in that, described step (100) is specially: on the basis of lumen, and simply change the coordinate of field information, it is described below:

If light L [s, t, x, y] is by (s, t) light that ∈ Π, (x, y) ∈ Ω defines, (x, y) be the intersection point of the line that projects of object and its vacuum on plane Π and plane Ω, then, on plane Π, x and s is consistent, and y and t is consistent; Thus obtain light field core face Ly, t and Lx, s, that is, the horizontal section of field information and plumb cut; Because visual angle and core face exist linear relationship, introduce parameter " inconsistency " thus, that is: the degree of depth in the plane of the spot projection in scene determines the rate of change of image in view; Thus, standardization light field coordinate information, and introduce parameter " inconsistency ".

3. a kind of light field image dividing method according to claim 1, is characterized in that, described step (200) is specially: in light field, choose any view as training sample, marks the different object lines in this view.

4. a kind of light field image dividing method according to claim 1, it is characterized in that, described step (300) is specially: for the object of mark, select in image: rgb value, Hessian eigenwert, tension variance and inconsistency attribute as training input, and obtain sorter with this.

5. a kind of light field image dividing method according to claim 1, is characterized in that, described step (400) is specially: after classification terminates, and carries out grid search to smallest partition node, and the classification of undue refinement is fused into same classification again.