Feature point detection method based on edge saliency and scale sensitivity

Abstract

The invention discloses a feature point detection method based on edge saliency and scale sensitivity. The method comprises the following steps of performing Gaussian smooth denoising on an input image; calculating the edge saliency of the pixel points through a Laplace operator; calculating a texture change degree based on the local gradient magnitude; generating a scale sensitivity weighted value through multi-scale analysis; candidate points are screened in combination with edge saliency, texture weighting and scale weighted values, and mismatching points are removed through non-maximum suppression; and finally, key points are enhanced and marked. According to the method, the edge structure, the texture information and the multi-scale features are fused, the robustness and accuracy of feature point detection are improved, and the method is particularly suitable for image matching and target recognition tasks in complex scenes.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and digital image processing technology, specifically relating to a feature point detection method that combines edge saliency and scale sensitivity, which can be applied to image matching, 3D reconstruction, target tracking and augmented reality. Background Technology

[0002] In the field of computer vision, feature point detection is a key technology for image matching, 3D reconstruction, and object recognition, and its performance directly affects the accuracy of subsequent algorithms. Traditional methods such as SIFT and SURF detect feature points by constructing scale space. Although they have rotation and scale invariance, they have two significant drawbacks: first, they rely on global gradient statistical features, which can easily lead to false detections in areas with repetitive or weak textures; second, the Difference of Gaussian (DoG) operator is sensitive to noise, and its stability drops sharply in low-light or high dynamic range scenes. In recent years, deep learning-based detectors (such as SuperPoint) have improved semantic perception capabilities, but their high model complexity and reliance on large amounts of labeled data make them difficult to deploy on edge devices with limited computing resources.

[0003] Another type of edge-feature-based method (such as the Harris corner detector) locates feature points through corner response functions. While computationally efficient, it only considers local grayscale changes and ignores two key factors: first, the significant differences in edge structures (e.g., insufficient distinction between straight and curved edges); and second, the adaptive fusion of multi-scale features. Existing improvements, such as the FAST detector, accelerate the judgment process through machine learning, but sacrifice the geometric consistency of feature points, leading to a significant decrease in matching accuracy when the image is deformed or the viewpoint changes. Furthermore, existing methods generally use fixed thresholds to select feature points, making it difficult to adapt to dynamic changes in edge strength and texture complexity under different scenes. Summary of the Invention

[0004] The purpose of this invention is to propose a feature point detection method based on edge saliency and scale sensitivity to solve the problems of poor adaptability to complex scenes and insufficient robustness of feature point detection in existing technologies, thereby improving the accuracy of image matching and target recognition.

[0005] To achieve the above objectives, this invention provides a feature point detection method based on edge saliency and scale sensitivity, comprising the following steps:

[0006] S1, acquire the input image and perform Gaussian smoothing on the image to remove noise;

[0007] S2, calculate the edge saliency of each pixel, and calculate the second derivative of the image using the Laplacian operator to obtain the intensity of change of the image at that point;

[0008] S3, based on the edge saliency value, further calculates the texture change degree of each pixel, and describes the texture change by calculating the gradient magnitude in the local region;

[0009] S4. The image is blurred through multi-scale analysis to generate image features at different scales and calculate the scale sensitivity weighting value.

[0010] S5, combining the edge saliency, texture weighting and scale weighting values ​​from steps S2, S3 and S4, candidate points are screened and mismatched points are removed by non-maximum suppression;

[0011] S6 enhances the final selected feature points and marks them as key points.

[0012] A further technical solution is that, in S1, the Gaussian smoothing process includes:

[0013] The image is convolved using a standard Gaussian filter to smooth it and remove noise.

[0014] A further technical solution is that, in S2, the step of calculating the edge saliency includes:

[0015] The second derivative of the image is calculated using the Laplacian operator, and the edge saliency value is calculated using the following formula:

[0016]

[0017] in, This represents the gradient value of the current pixel. This is a local window.

[0018] A further technical solution is that, in S3, the step of calculating the texture variation degree includes:

[0019] The texture variation is calculated using the following formula by calculating the gradient information of each pixel in the image and the gradient magnitude of its neighborhood:

[0020]

[0021] in, Represents pixels The gradient magnitude.

[0022] A further technical solution is that, in S4, the scale sensitivity weighting step includes:

[0023] By performing multi-scale blurring on the image, image features at different scales are generated, and a weighted value is calculated for each scale. Combining edge saliency and texture weighting, the scale weighting is calculated using the following formula:

[0024]

[0025] in, For scale levels, The marginal significance value, This represents the texture variation degree.

[0026] A further technical solution is that, in S5, the non-maximum suppression step includes:

[0027] For each candidate point, compare it with other points in its neighborhood, retain the local maximum significance points, and remove non-maximum points.

[0028] A further technical solution is that, in S5, the step of removing mismatched points includes:

[0029] If the significance value of the edges around a candidate point is low, the point is considered to be located in a flat region or an insignificant edge and is therefore eliminated.

[0030] If the texture variation of a candidate point is small, the point is considered to be located in a low-texture region and needs to be removed.

[0031] If a point shows significant changes across multiple scales, it is considered a mismatch and is removed. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This invention provides a feature point detection method based on edge saliency and scale sensitivity.

[0034] Step-by-step diagram.

[0035] Figure 2 This is a structural diagram of the edge saliency calculation module provided by the present invention.

[0036] Figure 3 This is a structural diagram of the texture variation calculation module provided by the present invention.

[0037] Figure 4 This is a structural diagram of the scale sensitivity weighting value calculation module provided by the present invention.

[0038] Figure 5 This is a visual comparison of the method of the present invention with ORB and SIFT feature point extraction methods on infrared images.

[0039] Figure 6 This is a visual comparison of the method of this invention with ORB and SIFT feature point extraction methods on visible light images. Detailed Implementation

[0040] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0041] In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0042] Please see Figures 1-6 This invention provides a feature point detection method based on edge saliency and scale sensitivity, comprising the following steps:

[0043] (1) Obtain the input image and perform Gaussian smoothing on the image to remove noise;

[0044] (2) Calculate the edge saliency of each pixel, and calculate the second derivative of the image using the Laplacian operator to obtain the intensity of change of the image at that point;

[0045] (3) Based on the edge saliency value, the texture change degree of each pixel is further calculated, and the texture change is described by calculating the gradient magnitude in the local area;

[0046] (4) The image is blurred by multi-scale analysis to generate image features at different scales and the scale sensitivity weighting value is calculated.

[0047] (5) Combining the edge saliency, texture weighting, and scale weighting values ​​from steps S2, S3, and S4, the results are...

[0048] Candidate points are filtered and mismatched points are removed by nonmaximum suppression;

[0049] (6) Enhance the final selected feature points and mark them as key points.

[0050] In a preferred embodiment of the present invention, the Gaussian smoothing process in step (1) includes the following specific steps:

[0051] (1-1) First, construct a two-dimensional Gaussian kernel function matrix. This kernel function has a bell-shaped distribution in the spatial domain, with the largest weight at the center and decreasing exponentially outwards. Its mathematical expression is:

[0052]

[0053] in The standard deviation of the Gaussian kernel (in pixels) is set to 1.0; Represents the position coordinates (in pixels) relative to the center within the kernel. In actual implementation, a 5×5 discretized window is used (i.e., All kernel coefficients need to be normalized so that their sum is 1.

[0054] (1-2) For the input image The convolution operation is performed, and the output value of each pixel is the weighted sum of its 5×5 neighboring pixels and the Gaussian kernel coefficients. The calculation process is expressed as follows:

[0055]

[0056] in The smoothed image is represented by a mirror-fill method for boundary processing (i.e., symmetrically expanding pixel values ​​at image boundaries) to ensure that the output image size remains consistent with the input. This step effectively suppresses high-frequency noise in the image, providing a stable input for subsequent feature extraction.

[0057] In a preferred embodiment of the present invention, the edge saliency calculation in step (2) includes the following specific steps:

[0058] (2-1) The second derivative of the image is calculated using the discrete Laplacian operator, passing through the center pixel. The local curvature change is estimated by the difference between the pixel and its four neighboring pixels, and its calculation expression is as follows:

[0059]

[0060] in Indicates position The Laplacian response value at the specified location is calculated by using a convolution kernel of [0 1 0; 1 -4 1; 0 1 0] to filter the image, which can effectively highlight the edges and corners in the image.

[0061] (2-2) Local normalization is performed on the Laplace response within a 3×3 window. (Right now picture

[0062] Calculate relative edge strength within the prime neighborhood and add a minimum value. To prevent division by zero, the calculation process is as follows:

[0063]

[0064] in The resulting edge saliency map The value ranges from [0,1], with values ​​closer to 1 indicating more significant edge features at that location. After calculation, histogram equalization can be performed to further enhance feature contrast.

[0065] In a preferred embodiment of the present invention, the texture variation calculation in step (3) includes the following specific steps:

[0066] (3-1) The Sobel operator is used to calculate the gradient field of the image, using the horizontal direction ( Direction) and vertical direction ( The kernel is used for filtering by a convolution kernel in the direction of ( ), and its kernel matrix is ​​represented as:

[0067]

[0068] The gradient components are then calculated as follows:

[0069]

[0070] in This represents the convolution operation. and These represent the gradient components in the horizontal and vertical directions, respectively (unit: gray level / pixel).

[0071] (3-2) Calculate the gradient magnitude of each pixel and obtain the local gradient intensity using the Euclidean distance formula:

[0072]

[0073] in Indicates position The gradient vector at that point, This represents the magnitude of the vector. The larger the value, the more drastic the texture change at that point.

[0074] (3-3) The texture complexity is described by accumulating gradient magnitudes within a 3×3 neighborhood, and the calculation expression is as follows:

[0075]

[0076] in The definition is the same as in step (2-2). This step essentially involves performing a local integration operation on the gradient magnitude map to obtain the texture variation. Reflecting on The value indicates the richness of texture in the local area centered on the model; a larger value indicates more obvious texture features.

[0077] In a preferred embodiment of the present invention, the scale sensitivity weighting in step (4) includes the following specific steps:

[0078] (4-1) Construct a Gaussian-scale space and use different standard deviations A Gaussian kernel is used to generate a multi-scale image set, with the scale parameter chosen as follows:

[0079]

[0080] The images at each scale are obtained by Gaussian blurring:

[0081]

[0082] The Gaussian kernel size is based on The principles are automatically determined, ensuring coverage of 99.7% of the energy distribution. This step constructs a multi-scale representation of the image, capable of simultaneously capturing feature structures of different sizes.

[0083] (4-2) After calculating the edge saliency and texture variation at each scale level, scale-weighted calculation is performed:

[0084]

[0085] in It is a scale-adaptive weighting factor that appropriately attenuates the contribution of large-scale features; and Representing scale The weighted value represents the saliency of the features at different scales, as well as the degree of texture variation.

[0086] (4-3) Obtaining the optimal feature response through cross-scale maximum fusion:

[0087]

[0088] This operation takes the maximum value along the scale dimension, retaining the most significant feature response at each location. The final output is a normalized feature map:

[0089]

[0090] The denominator is the maximum response value of the entire image, ensuring the comparability of feature strengths after normalization.

[0091] In a preferred embodiment of the present invention, nonmaximum suppression in step (5) includes the following specific steps:

[0092] (5-1) First, filter responses that exceed the set threshold. Candidate points are selected to form an initial set of feature points:

[0093]

[0094] The threshold This threshold is used to filter out candidate points with weak responses while preserving potential feature regions. In practical applications, this threshold can be dynamically adjusted based on the image content.

[0095] (5-2) In the 5×5 neighborhood of each candidate point (i.e. Non-maximum suppression is performed within the pixel range, retaining only the local maximum response points:

[0096]

[0097] This step ensures that the feature points have sufficient spatial discriminative power, avoiding excessive clustering of feature points. A neighborhood radius of 2 pixels corresponds to a 5×5 search window.

[0098] In a preferred embodiment of the present invention, the feature point enhancement in step (6) includes the following specific steps:

[0099] (6-1) Perform linear normalization on the filtered feature point response values ​​to make them evenly distributed in the [0,1] interval:

[0100]

[0101] This process can eliminate the influence of overall brightness differences between images on feature point selection.

[0102] (6-2) When marking key points, two conditions must be met: the normalized response value must be greater than the quality level threshold. and maintain the minimum distance from the marked point. Pixels:

[0103]

[0104] in Controlling the quality level of feature points, Ensure that the feature points are spatially uniformly distributed. This step outputs the final set of keypoints. It can be used as input features for subsequent image matching or target recognition.

[0105] Beneficial effects

[0106] The feature point detection method based on edge saliency and scale sensitivity provided by this invention has significant technical advantages. By innovatively integrating edge structure, local texture, and multi-scale features, this method significantly improves the accuracy and robustness of feature detection while maintaining high computational efficiency. Specifically, it employs a multi-scale weighting mechanism of edge saliency and texture variability, improving the detection accuracy of feature points by more than 12.6% in complex scenarios such as low light and noise interference; and it effectively extracts feature points through a three-level screening strategy (edge ​​strength, texture complexity, and scale consistency), demonstrating superior adaptability, particularly in professional fields such as multispectral fusion and remote sensing images. The technical solution of this invention has outstanding value in engineering applications. Actual tests show that in scenarios requiring high-precision feature matching, such as AR navigation and autonomous driving, this method improves the feature repetition rate by more than 17.8% compared to traditional algorithms, and the key parameters all have clear physical meanings, facilitating adjustment and optimization in different application scenarios.

Claims

1. A feature point detection method based on edge saliency and scale sensitivity, characterized in that, Includes the following steps: S1: Acquire the input image and perform Gaussian smoothing on the image to remove noise. S2, calculate the edge saliency of each pixel, and calculate the second derivative of the image using the Laplacian operator to obtain the intensity of change of the image at that point; S3, based on the edge saliency value, further calculates the texture variation degree of each pixel, and describes the texture variation by calculating the gradient magnitude in the local region; S4: The image is blurred through multi-scale analysis to generate image features at different scales and calculate the scale sensitivity weighting value. S5, combining the edge saliency, texture weighting and scale weighting values ​​from steps S2, S3 and S4, candidate points are screened and mismatched points are removed by non-maximum suppression; S6 enhances the final selected feature points and marks them as key points.

2. The feature point detection method based on edge saliency and scale sensitivity according to claim 1, characterized in that, In S1, the Gaussian smoothing process includes the following steps: The image is convolved using a standard Gaussian filter to smooth it and remove noise.

3. The feature point detection method based on edge saliency and scale sensitivity according to claim 1, characterized in that, In S2, the steps for calculating the edge saliency include: The second derivative of the image is calculated using the Laplacian operator, and then the edge saliency value is calculated: Where G(x,y) is the gradient value of the current pixel, and W(x,y) is the local window.

4. The feature point detection method based on edge saliency and scale sensitivity according to claim 1, characterized in that, In S3, the steps for calculating the texture variation include: The texture variation is calculated using the following formula by calculating the gradient information of each pixel in the image and the gradient magnitude of its neighborhood: in, This represents the gradient magnitude of pixel (x', y').

5. The feature point detection method based on edge saliency and scale sensitivity according to claim 1, characterized in that, In S4, the scale sensitivity weighting step includes: By performing multi-scale blurring on the image, image features at different scales are generated, and a weighted value is calculated for each scale. Combining edge saliency and texture weighting, the scale weighting is calculated using the following formula: Where σ is the scale level, S e (x,y) represents the edge saliency value, and T(x,y) represents the texture variation.

6. The feature point detection method based on edge saliency and scale sensitivity according to claim 5, characterized in that, In S5, the non-maximum suppression step includes: For each candidate point, compare it with other points in its neighborhood, retain the local maximum significance points, and remove non-maximum points.

7. The method according to claim 6, characterized in that, The step of removing mismatched points includes: if the edge significance value around a candidate point is low, the point is considered to be located in a flat region or an insignificant edge, and is thus removed.

8. The method according to claim 6, characterized in that, The step of removing mismatched points includes: if the texture change of a candidate point is small, the point is considered to be located in a low-texture area and needs to be removed.

9. The method according to claim 6, characterized in that, The steps for removing mismatched points include: if a point shows significant changes across multiple scales, then the point is considered a mismatched point and is removed.

Images