A method for extracting shape features of a target object in a noise background

Abstract

This invention belongs to the field of computer image processing, specifically a method for extracting the shape features of a target object against a noisy background. Existing convex hull solving methods are highly sensitive to noise or outliers. In practical applications, due to image noise, the target object is affected by noise interference, and outliers exist in the target region's point set, leading to convex hull extraction errors and failing to effectively represent the geometric shape features of the target object. The following solution is proposed, comprising the following steps: S1, Forward Voting: Voting is conducted from the image space to the Hough space in 360 directions. This invention considers the interference of image noise on convex hull extraction, employing a voting mechanism and linear fitting technology. By finding the voting cell corresponding to each convex hull vertex, the coordinates of the convex hull vertex of the target region are backfitted based on the voting cell data. This allows for the extraction of the geometric shape features of the target object under noise interference and provides better robustness to noise.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer image processing, and particularly relates to a shape feature extraction method of a target object under a noise background. BACKGROUND

[0002] Extracting the geometric shape feature of the target object in the image is helpful for further analyzing, recognizing and understanding the target object. The convex hull is an effective geometric shape expression form and is often applied in computer vision image processing. The convex polygon surrounding the target object is extracted from the image, and the geometric shape feature of the target object is described by using the convex hull. For example, the target tracking, image registration, image segmentation, image clustering, salient object detection and other image processing applications are performed based on the convex hull feature.

[0003] The convex hull algorithm for calculating the point set of the target region has been widely studied. Common convex hull solving algorithms include the Graham scanning method, the Jarvis stepping method and the fast convex hull algorithm. The minimum convex polygon surrounding the target object is solved by using the convex hull algorithm, and the geometric shape feature of the target object is represented.

[0004] The current convex hull solving method is very sensitive to noise or outliers. In actual application, due to the existence of image noise, the target object is disturbed by the noise, and the point set of the target region has outliers, which leads to the error of the convex hull extraction and cannot effectively represent the geometric shape feature of the target object. SUMMARY

[0005] The present application aims at solving the problem in the prior art that the convex hull solving method is very sensitive to noise or outliers, in actual application, due to the existence of image noise, the target object is disturbed by the noise, and the point set of the target region has outliers, which leads to the error of the convex hull extraction and cannot effectively represent the geometric shape feature of the target object, and proposes a shape feature extraction method of a target object under a noise background.

[0006] In order to achieve the above object, the present application adopts the following technical scheme:

[0007] A shape feature extraction method of a target object under a noise background comprises the following steps:

[0008] S1, forward voting: voting from the image space to the Hough space in 360 directions;

[0009] S2, searching for the lower boundary cell: after the target object is voted, the target object presents a strip shape in the Hough space, and the lower boundary cells of the strip are searched, the cells correspond to the convex hull vertices of the target object, and there are 360 lower boundary cells in total;

[0010] S3, cell segmentation: according to the voting value of the lower boundary cell, 360 continuous boundary cells are segmented. Several groups of cells are obtained, and each group of cells corresponds to a convex hull vertex of the target object;

[0011] S4, reverse fitting: for each group of cells, a linear function f is fitted based on the voting angle θ value and the voting distance ρ value of the cell. The convex hull vertex coordinate (x, y) is just the fitting function coefficient.

[0012] Preferably, S1 is specifically: Hough voting is performed in 360 directions, and the Hough transform formula used is: ρ = x*cosθ + y*sinθ, wherein θ is the voting angle, and the range of the voting angle is 0≤θ<360°; ρ is the voting distance of the pixel point (x, y) in the θ direction.

[0013] Preferably, S2 is specifically: in the Hough space, the lower boundary cells of the strip are searched, each boundary cell includes: voting angle θ, voting distance ρ and voting value v, and 360 lower boundary cells of the strip are stored in an array according to θ from small to large.

[0014] Preferably, S3 is specifically: a threshold T is selected, and 360 lower boundary cells are segmented according to the voting value v, for example, if the v value of a certain cell is greater than T, the cell is a segmentation point, and the continuous cells between two segmentation points form a group, which correspond to a convex hull vertex of the target object.

[0015] Preferably, S4 is specifically: for each group of segmented boundary cells, the ρ / cosθ and tanθ values are calculated based on the θ and ρ values of the cell, and a linear function f is fitted using the calculated values, and the fitted function is represented as: Then the corresponding convex hull vertex coordinate is: x i =f0, y i =f1, and the vertex coordinate is just the fitting function system.

[0016] Preferably, in S1, the forward voting: voting is performed from the image space to the Hough space in 360 directions, and the voting angle θ ranges from 0 to 360°, and the angle resolution is 1°.

[0017] Preferably, S1 is specifically: transformation is performed from the image space to the Hough space in 360 directions, the relationship between the convex hull vertex in the image space and the cell in the Hough space is analyzed, and the selected Hough transform formula is:

[0018] ρ = x*cosθ + y*sinθ θ∈[0 360) (1)

[0019] Where, theta is the voting angle, the range of voting angle is 0<=theta<360 degree. rho is the voting distance of pixel point (x, y) in theta direction. Hough space is represented by a 2D accumulation array, denoted as: H(theta, rho).

[0020] Preferably, in S2, the cell corresponds to the convex hull vertex of the target object, and each boundary cell comprises a voting angle theta, a voting distance rho and a voting value v, represented by a triple (theta, rho, v), wherein theta is the voting angle 0<=theta<360, rho is the voting distance corresponding to the boundary cell in the theta voting direction, and v is the voting value corresponding to the boundary cell.

[0021] In the present application, the shape feature extraction method of the target object in the noise background has the following beneficial effects:

[0022] The present application takes into account the interference of image noise on convex hull extraction, uses the voting mechanism and linear fitting technology, finds the voting cell corresponding to each convex hull vertex, reversely fits the convex hull vertex coordinates of the target region according to the voting cell data, and can extract the geometric shape feature of the target object under the interference of noise, and has better robustness to noise. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The flow chart of the shape feature extraction of the target object of the present application;

[0024] Figure 2 The lower boundary cell of the strip of the embodiment of the present application;

[0025] Figure 3 The voting value distribution of the boundary cell of the embodiment of the present application;

[0026] Figure 4 The shape feature extraction result of the target object of the first kind of the present application;

[0027] Figure 5 The shape feature extraction result of the target object of the second kind of the present application. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all.

[0029] REFERENCE Figures 1-5 A shape feature extraction method of a target object in a noise background, comprising the following steps:

[0030] S1, forward voting. In 360 directions, vote from image space to Hough space. The voting angle θ ranges from [0 360°) with an angle resolution of 1°.

[0031] S2, search lower boundary cells. After all target region pixels in the image vote, the cells with non-zero vote values present as bands in Hough space. In each column of Hough space, search the lower boundary cells of the bands, which correspond to the convex hull vertices of the target object. There are 360 lower boundary cells in total.

[0032] S3, cell segmentation. According to the size of the vote value v of the lower boundary cells, segment the 360 consecutive boundary cells. Get several groups of cells, each group of cells corresponding to a convex hull vertex of the target object.

[0033] S4, inverse fitting. After segmentation, each group contains multiple cells. Based on the voting angle θ value and the voting distance ρ value of each cell, perform transformation to construct a linear function relationship. According to the transformed values, inversely fit a linear function f. The convex hull vertex coordinates (x, y) are just the zeroth term and first term coefficients of the fitted linear function f.

[0034] Further, the S1 is specifically: in 360 directions, transform from image space to Hough space, analyze the relationship between the convex hull vertices in the image space and the cells in the Hough space. The selected Hough transform formula is:

[0035] ρ = x cos θ + y sin θ θ ∈ [0 360) (1)

[0036] Where θ is the voting angle, and the range of the voting angle is 0≤θ<360°. ρ is the voting distance of the pixel point (x, y) in the θ direction. The Hough space is represented by a 2D accumulation array, denoted as: H(θ, ρ).

[0037] Further, the S2 is specifically: after all target region pixels in the image vote, present as bands in Hough space. In each column of Hough space, first find the cell corresponding to the maximum vote value; then search downward from the maximum vote value cell, if the vote value of a certain cell is less than 3, the cell is taken as the lower boundary cell. The searched lower boundary cells of the bands are as shown in Figure 2 The 360 lower boundary cells are shown in red.

[0038] These cells correspond to the convex hull vertices of the target object. Each boundary cell includes a voting angle θ, a voting distance ρ and a voting value v, which are represented by a triple (θ, ρ, v). Where θ is the voting angle 0≤θ<360, ρ is the voting distance of the boundary cell corresponding in the voting direction of θ, and v is the voting value of the boundary cell corresponding.

[0039] There are 360 boundary cells in total, which are stored in an array according to the ascending order of θ.

[0040] Further, the S3 is specifically: selecting a threshold T (such as T=10), dividing the 360 lower boundary cells according to the voting value v, and finding the cells corresponding to different convex hull vertices.

[0041] The voting value v distribution of the boundary cells is shown in Figure 3 Most of the voting values of the boundary cells are small, and only a few have peak values. If the v value of a certain cell is greater than T, the cell is taken as a division point. The continuous cells between two division points form a group, which correspond to a convex hull vertex of the target object.

[0042] Further, the S4 is specifically: for each group of boundary cells after division, a linear function f is fitted based on the θ and ρ values of the cells, and the coefficients of the linear function f are the coordinates of the convex hull vertex.

[0043] The voting angle θ and voting distance ρ of the boundary cells in the ith group satisfy the following formula:

[0044]

[0045] Where x i and y i are the coordinates of the convex hull vertex corresponding to the ith group of boundary cells. As can be seen from formula (2), the values of ρ / cosθ and tanθ are in a linear function relationship. For the boundary cells in the same group, the values of ρ / cosθ and tanθ are calculated respectively. Using the calculated values, a linear function f is fitted. The fitted function is represented as:

[0046]

[0047] Where f0 and f1 are the coefficients of the fitted linear function respectively. Comparing formula (2) and (3), we get:

[0048] x i = f0

[0049] y i = f1 (4)

[0050] Therefore, the coordinates of the convex hull vertex are exactly the zero-order term and the first-order term coefficients of the fitted linear function f.

[0051] The border cells are stored in order, and the fitted convex hull vertices are also in order. By connecting the vertices in order, a convex polygon of the target object is formed, as shown in Figure 4 Even if the image background contains noise interference, the convex polygon surrounding the target object can be accurately extracted by using the voting mechanism and fitting technique, and the geometric shape of the traffic sign is well represented, and has good anti-noise performance.

[0052] Figure 5 The second extracted geometric shape feature of the target object is shown. In the image, in addition to the tree leaves, there are other noise interferences. By using the technical method provided by the present application, the convex polygon of the target object can be accurately extracted, and the geometric shape of the leaves is represented.

[0053] The above description is only the preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art, according to the technical range disclosed by the present application and the inventive concept of the present application, should be covered in the protection scope of the present application.

Claims

1. A method for extracting shape features of a target object in a noise background, characterized in that, The method comprises the following steps: S1, forward voting: voting from the image space to the Hough space in 360 directions; The S1 is specifically: Hough voting in 360 directions, and a Hough transform formula used is: ρ=x*cosθ+y*sinθ, wherein θ is a voting angle, and the range of the voting angle is 0≤θ<360°; and ρ is a voting distance of a pixel point (x, y) in the θ direction; In the S1, the forward voting is: voting from the image space to the Hough space in 360 directions, and the range of the voting angle θ is [0 360°), and the angle resolution is 1°; The S1 is specifically: transforming from the image space to the Hough space in 360 directions, analyzing the relationship between the convex hull vertex in the image space and the unit cell in the Hough space, and selecting a Hough transform formula: ρ=x*cosθ+y*sinθ θ∈[0 2π) (1) Wherein θ is a voting angle, and the range of the voting angle is 0≤θ<360°; and ρ is a voting distance of a pixel point (x, y) in the θ direction; and the Hough space is represented by a 2D accumulation array, denoted as: H(θ, ρ); S2, searching for lower boundary unit cells: after the target object is voted, the target object presents a strip shape in the Hough space, and the lower boundary unit cells of the strip are searched, and the unit cells correspond to the convex hull vertex of the target object, and there are 360 lower boundary unit cells in total; The S2 is specifically: searching for the lower boundary unit cells of the strip in the Hough space, and each boundary unit cell comprises a voting angle θ, a voting distance ρ and a voting value v, and the 360 lower boundary unit cells of the strip are stored in an array according to the increasing order of θ; In the S2, the unit cells correspond to the convex hull vertex of the target object, each boundary unit cell comprises a voting angle θ, a voting distance ρ and a voting value v, and a triplet (θ, ρ, v) is used, wherein θ is a voting angle 0≤θ<360, ρ is a voting distance corresponding to the boundary unit cell in the θ voting direction, and v is a voting value corresponding to the boundary unit cell; S3, unit cell segmentation: according to the voting value of the lower boundary unit cell, 360 continuously stored boundary unit cells are segmented; several groups of unit cells are obtained, and each group of unit cells corresponds to a convex hull vertex of the target object; The S3 is specifically: a threshold T is selected, 360 lower boundary unit cells are segmented according to the voting value v, if the v value of a certain unit cell is greater than T, the unit cell is a segmentation point, and the continuous unit cells between two segmentation points form a group, and they correspond to a convex hull vertex of the target object; S4, reverse fitting: for each group of unit cells, a linear function f is fitted based on the voting angle θ value and the voting distance ρ value of the unit cell; and the convex hull vertex coordinate (x, y) is just the function coefficient of the fitting. The S4 is specifically: for each group of boundary cells after segmentation, based on the theta and rho values of the cells, the rho / cos theta and tan theta values are calculated, and a linear function f is fitted using the calculated values, and the fitted function is represented as: Then the corresponding convex hull vertex coordinates are: x i =f0, y i =f1, the vertex coordinates are just the fitted function system.

Images