Homogenized image feature extraction method with gradient value decreasing step by step

Abstract

The invention discloses a homogenized image feature extraction method with gradient values decreasing step by step, which belongs to the field of image processing, and comprises the following steps: traversing a whole disparity map, and obtaining a point with the maximum gradient value in the disparity map as a first feature point; generating a circular domain and carrying out gradient zero setting on all other areas except the first feature point in the circular domain; carrying out circular domain diffusion in the whole disparity map by taking the obtained circular domain as a reference, independently calculating a gradient value in each diffused circular domain, and carrying out feature point selection; and selecting feature points in scattered areas which are not covered by the diffused circular domain in the disparity map until no new feature points are generated. According to the method, part of feature points can also be extracted in a flat region with small depth change through gradient value step-by-step descending homogenization feature extraction, and point cloud holes are prevented from being generated in the region.

Description

Technical Field

[0001] The invention relates to the field of image processing, and in particular to a method for extracting uniform image features with gradually decreasing gradient values. Background Art

[0002] In traditional thinking, the selection of three-dimensional spatial feature points in the disparity map is a fixed process, which is similar to the feature detection in the stereo matching system, and the target is all relatively "strong" feature points. The difference is that the feature detection link in the stereo matching system selects the feature points with large and obvious pixel transitions, while the feature extraction link in the depth value calculation part obtains the feature points in the disparity map with large gradient values ​​or obvious gradient value changes.

[0003] The traditional feature point selection operation is easy to understand and relatively simple. It will mechanically move closer to the areas with large gradient values ​​or obvious gradient value changes in the disparity map, and then select feature points in such areas to form a three-dimensional point cloud. For some disparity maps with a single structural change law and evenly distributed texture features, the traditional feature point selection can obtain feature points with good distribution effects and texture information that can be fully represented. The final scattered feature point set is involved in all areas of the disparity map, so it can cover all three-dimensional information of the scene to be reconstructed, and it can also be directly applied to the next step of three-dimensional coordinate calculation. However, when the traditional feature point selection is faced with the disparity map of unstructured weak texture large-scale scenes, the following disadvantages will occur: First, since such scenes cover a large area of ​​weak texture, when using traditional feature point selection to work in the disparity map of such scenes, the feature points that can be selected are often "snatched" by strong texture areas with obvious texture information and large depth changes, while large areas of weak texture areas can only have a very small number of feature points or even no feature points. When the weak texture area lacks the distribution of feature points, it will directly lead to the subsequent reconstruction task losing the three-dimensional spatial information in the area and no point cloud effect will be produced.

[0004] Second, feature points tend to be concentrated in a small area.

[0005] Third, some areas with strong textures are not taken seriously due to the “snatching” between feature points.

[0006] Therefore, if the traditional feature point selection is performed directly on the disparity map of the scene to be reconstructed in this paper, the obtained feature point set will not achieve the ideal effect in terms of distribution or position, and will directly waste the work done in the stereo matching link and the good disparity map effect obtained in this paper. Summary of the invention

[0007] In view of the above-mentioned deficiencies in the prior art, the present invention provides a homogenized image feature extraction method with a gradually decreasing gradient value.

[0008] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: A method for extracting features from a homogenized image with a gradient value decreasing step by step, comprising the following steps: S1, traverse the entire disparity map, and obtain the point with the largest gradient value in the disparity map as the first feature point; S2, generating a circular region and setting the gradient of all regions except the first feature point in the circular region to zero; S3, using the circular domain obtained in S2 as a reference, performing circular domain diffusion in the entire disparity map, independently calculating the gradient value in each diffused circular domain and selecting feature points; S5. Select feature points in scattered areas of the disparity map that are not covered by the diffused circular region until no new feature points are generated. Furthermore, the specific method of calculating the gradient value in S1 is: S11. Based on the Sobel operator, the Sobel convolution kernel is used to calculate the gradient components of the image in the horizontal and vertical directions respectively. The Sobel convolution kernel is defined as:

[0009] and Horizontal and vertical The convolution kernel; the gradient components in the horizontal and vertical directions The calculation formula is:

[0010] in, For the image The pixel value of is the image pixel coordinate, is the position of the image in the corresponding convolution kernel; S12, calculating the gradient amplitude according to the Euclidean norm of the gradient of the image in the horizontal and vertical directions, and the specific calculation method is:

[0011] In the formula, is the gradient amplitude.

[0012] Furthermore, the specific method of generating the circular domain in S2 is: A circular region is formed with the first feature point obtained as the center and the distance between the two feature points closest to the first feature point as the radius.

[0013] Furthermore, in S2, the specific method of setting the gradient value to zero is: Assume that the feature points in the disparity map are , and the corresponding circle radius is , pixel coordinates , the gradient zeroing operation is expressed as:

[0014] In the formula, is the original gradient value, is the gradient value after suppression, the condition Indicates that the point is inside the circle. is the radius of the circle; Indicates that the gradient value of the feature point itself is retained.

[0015] Furthermore, the S3 specifically includes the following steps: S31, setting the circular region obtained in S2 as the basis, and the size range of the entire disparity map as the total area; S32, diffusing the circular domain within the disparity map with the obtained circular domain range size, until the entire circular domain of the same size cannot be accommodated, and a plurality of diffused circular domains are obtained; S33, performing independent gradient value calculation within each diffused circular domain, and obtaining the point with the largest gradient value in the corresponding circular domain as the first feature point of the corresponding circular domain; S34, taking the first feature point of the corresponding circular domain as the center and the distance between any two feature points other than the first feature point of the corresponding circular domain as the radius, setting the gradient value in the area other than the first feature point of the corresponding circular domain to zero, thereby obtaining feature points containing three-dimensional information and with depth changes.

[0016] Furthermore, the S4 specifically includes the following steps: S41, forming a new circular domain with the coordinates of any point in the scattered area as the center and the distance between the two nearest feature points other than the point as the radius; S42, using the formed new circular domain to traverse the scattered areas; S43, when traversing to any point, if the depth change gradient values ​​of the upper, lower, left and right extremes of the circular domain boundary are not equal to zero, then the point is also classified as a feature point. The boundary gradient non-zero condition can be expressed as:

[0017] in, They are the depth change gradient values ​​at the extreme top, extreme bottom, extreme left, and extreme right.

[0018] The present invention has the following beneficial effects: The gradient value gradually decreases and the homogenization feature extraction algorithm takes into account that the feature points are taken in the area with large depth changes, and the texture information covered is more comprehensive. The gradient zeroing operation also avoids the dense distribution of feature points. At the same time, the gradient value gradually decreases and the homogenization feature extraction can also extract some feature points in the flat area with small depth changes, avoiding the generation of point cloud holes in such areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The figure is a flow chart of a homogenized image feature extraction method with gradually decreasing gradient values ​​according to the present invention.

[0020] Figure 2 Schematic diagram of feature point selection results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0022] like Figure 1 As shown, this paper proposes a homogenized feature extraction algorithm with a gradient value decreasing step by step in the feature extraction link of the disparity map. This paper explains the calculation steps of the homogenized feature extraction algorithm with a gradient value decreasing step by step as follows: The entire disparity map is traversed without blind spots to find the point with the largest gradient value in the map, which is set as And define it as a feature point.

[0023] The gradient calculation is based on the Sobel operator. First, the Sobel convolution kernel is used to calculate the gradient components of the image in the horizontal and vertical directions respectively:

[0024] in Indicates that the image is at position The pixel value, u and v values ​​represent the corresponding positions of the convolution kernel, such as , then when u=-1, v=1 that is In the convolution kernel, it corresponds to 1. is -1, then with the pixel Around Convolution is performed on pixel values ​​in a range.

[0025] and for The convolution kernel is defined as:

[0026] The gradient magnitude is the Euclidean norm of the horizontal and vertical gradients: .

[0027] by The distance between the nearest two feature points is A circular area is formed with a radius of All other areas outside the point are subjected to gradient zeroing operation. is the Euclidean distance.

[0028] In the circular region of the selected feature points, the gradient value of the non-feature point area is suppressed to avoid excessive aggregation of feature points. Suppose the feature point in the disparity map is , and the corresponding circle radius is , then the gradient zeroing operation can be expressed as: ; In the formula, is the original gradient value; is the gradient value after suppression; condition , indicates that the point is within the circle, is the radius of the circle; Ensure that the gradient value of the feature point itself is preserved.

[0029] Based on the size range of the circle, the size range of the entire disparity map is the total area. The diffusion operation of the circle is performed inside the disparity map until the entire circle of this size can no longer be accommodated. In this paper, the size of the disparity map is assumed to be The size of the circular domain is , then the number of circular diffusions in the disparity map is The remaining disparity map parts will be scattered areas.

[0030] In each diffused circular domain, the gradient value calculation operation is performed independently of each other, and the point with the largest gradient value in the circular domain is found and classified as a feature point. In this paper, the feature point selected in this step is set as ,in After the feature point is selected, the gradient of the circle where the point is located is also set to zero. However, the zeroing operation in this step is not equivalent to directly setting the gradient of the diffused circle to zero. is the center of the circle, the Euclidean distance The radius is set to zero. This step of zeroing operation avoids the selection of feature point distribution areas being too regular and mechanical, which leads to the omission of excellent feature points containing rich three-dimensional information and large depth changes.

[0031] A circular domain is also established, and the center of the circle is the coordinates of any point in the scattered area remaining in the disparity map. In this paper, the coordinates of this point are set as , and at the same time A circular domain is formed with a radius of . The circular domain is traversed through the scattered areas. When the depth gradient values ​​of the extreme upper, extreme lower, extreme left, and extreme right on the boundary of the circular domain are not equal to zero at a certain point, the point is also classified as a feature point. In this paper, it is set as , and loop until no new feature points are generated.

[0032] The selection of feature points for homogenization feature extraction with gradually decreasing gradient values ​​is completed.

[0033] The specific schematic diagram of the homogenization feature extraction algorithm with gradually decreasing gradient values ​​in this paper is as follows Figure 2 As shown in the figure. The gradient value gradually decreases and the homogenization feature extraction algorithm considers that the feature points are best taken in the area with large depth changes, because such areas contain a lot of texture information that cannot be abandoned. The gradient zeroing operation also avoids the dense distribution of feature points. At the same time, the gradient value gradually decreases and the homogenization feature extraction can also extract some feature points in the flat area with small depth changes, avoiding the generation of point cloud holes in such areas.

[0034] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0035] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for extracting features from a homogenized image with a gradient value decreasing step by step, characterized in that: The steps include: S1, traverse the entire disparity map, and obtain the point with the largest gradient value in the disparity map as the first feature point; S2, generating a circular region and setting the gradient of all regions except the first feature point in the circular region to zero; S3, using the circular domain obtained in S2 as a reference, performing circular domain diffusion in the entire disparity map, independently calculating the gradient value in each diffused circular domain and selecting feature points; S4. Select feature points in scattered areas of the disparity map that are not covered by the diffused circular region until no new feature points are generated.

2. The method for extracting uniform image features with gradually decreasing gradient values ​​according to claim 1, characterized in that: The specific method of calculating the gradient value in S1 is: S11. Based on the Sobel operator, the Sobel convolution kernel is used to calculate the gradient components of the image in the horizontal and vertical directions respectively. The Sobel convolution kernel is defined as: and Horizontal and vertical directions The convolution kernel; the gradient components in the horizontal and vertical directions The calculation formula is: in, For images in The pixel value of is the image pixel coordinate, is the position of the image in the corresponding convolution kernel; S12, calculating the gradient amplitude according to the Euclidean norm of the gradient of the image in the horizontal and vertical directions, and the specific calculation method is: In the formula, is the gradient amplitude.

3. The method for extracting uniform image features with gradually decreasing gradient values ​​according to claim 1, characterized in that: The specific method of generating the circular domain in S2 is: A circular region is formed with the first feature point obtained as the center and the distance between the two feature points closest to the first feature point as the radius.

4. The method for extracting uniform image features with gradually decreasing gradient values ​​according to claim 1, characterized in that: In S2, the specific method of setting the gradient value to zero is: Assume that the feature points in the disparity map are , and the corresponding circle radius is , pixel coordinates , the gradient zeroing operation can be expressed as: In the formula, is the original gradient value, is the gradient value after suppression, the condition Indicates that the point is within the circle. Indicates that the gradient value of the feature point itself is retained.

5. The method for extracting uniform image features with gradually decreasing gradient values ​​according to claim 1, characterized in that: The S3 specifically includes the following steps: S31, setting the circular region obtained in S2 as the basis, and the size range of the entire disparity map as the total area; S32, diffusing the circular domain within the disparity map with the obtained circular domain range size, until the entire circular domain of the same size cannot be accommodated, and a plurality of diffused circular domains are obtained; S33, performing independent gradient value calculation within each diffused circular domain, and obtaining the point with the largest gradient value in the corresponding circular domain as the first feature point of the corresponding circular domain; S34, taking the first feature point of the corresponding circular domain as the center and the distance between any two feature points other than the first feature point of the corresponding circular domain as the radius, setting the gradient value in the area other than the first feature point of the corresponding circular domain to zero, thereby obtaining feature points containing three-dimensional information and with depth changes.

6. The method for extracting uniform image features with gradually decreasing gradient values ​​according to claim 1, characterized in that: The S4 specifically includes the following steps: S41, forming a new circular domain with the coordinates of any point in the scattered area as the center and the distance between the two nearest feature points other than the point as the radius; S42, using the formed new circular domain to traverse the scattered areas; S43, when traversing to any point, if the depth change gradient values ​​of the upper, lower, left and right extremes of the circular domain boundary are not equal to zero, then the point is also classified as a feature point. The boundary gradient non-zero condition can be expressed as: in, They are the depth change gradient values ​​at the extreme top, extreme bottom, extreme left, and extreme right.

Images