A grayscale image feature extraction system

Abstract

The present invention discloses a grayscale image feature extraction system, comprising: a wavelet image pyramid module, which is used to perform layer-by-layer decomposition on a grayscale image based on discrete wavelet transform, extract the low-frequency information representing the global part and the high-frequency information representing the details in the image, and use these as multiple resolution representations of the image; a feature point detection module, which is used to detect the positions of feature points in the image as the position information of the features; a feature description module, which is used to generate corresponding descriptors as the description information of the features after performing principal direction normalization on the neighborhoods of the detected feature points, and finally output the positions of the feature points and the descriptors as features. The present invention can effectively solve the problem of being difficult to capture both the details and global information of an image simultaneously, and at the same time reduce the sensitivity of feature extraction to image noise.

Description

Technical Field

[0001] The present invention relates to the technical field of feature extraction, and in particular to a grayscale image feature extraction system. Background Art

[0002] Most traditional image feature extraction methods rely on pixel-level operations, such as gradient-based, edge-based, or texture-based analysis. However, these methods usually ignore the multi-scale characteristics of images and are difficult to effectively process local details and global structures in images. Especially in complex natural images, traditional methods often struggle to capture both the details and global information of images simultaneously, resulting in insufficiently comprehensive feature representation. In addition, traditional methods are highly sensitive to noise and are easily interfered by image noise, affecting the accuracy of feature extraction. Moreover, the features generated by the current FAST feature point detection method do not have rotational invariance, resulting in the inability of originally matching features to correctly match after the image rotates. Summary of the Invention

[0003] The purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art, and provide a grayscale image feature extraction system, which can effectively solve the problem of difficulty in simultaneously capturing the details and global information of an image, while reducing the sensitivity of feature extraction to image noise, and making the generated features have rotational invariance, effectively improving the quality of feature matching.

[0004] To achieve the above purpose, the technical solution provided by the present invention is: a grayscale image feature extraction system, comprising:

[0005] A wavelet image pyramid module, which is used to perform layer-by-layer decomposition of the grayscale image based on discrete wavelet transform, extract the low-frequency information representing the global part and the high-frequency information representing the details in the image, and use these as multiple resolution representations of the image;

[0006] A feature point detection module, which is used to detect the position of the feature points in the image as the position information of the features;

[0007] A feature description module, which is used to generate corresponding descriptors as the description information of the features after performing principal direction normalization on the neighborhood of the detected feature points, and finally output the position and descriptor of the feature points as the features.

[0008] Further, the wavelet image pyramid module specifically performs the following operations:

[0009] Perform wavelet image pyramid representation of the grayscale image f(x, y) based on discrete wavelet transform, where the image A j (x, y) that retains the low-frequency information and the images H j (x, y), V j (x, y), Dj The specific situation of (x, y) is as follows:

[0010]

[0011] In the formula, f(i, k) represents the gray value of the image f(x, y) at the coordinate (i, k), n and m respectively represent the length and width of the image f(x, y), x and y respectively represent the horizontal and vertical coordinates in the image, · represents the multiplication operation, and the image H j (x, y) represents the horizontal detail, and the image V j (x, y) represents the vertical detail, and the image D j (x, y) represents the diagonal detail, g(·) represents the low-pass filter, and h h (·), h v (·), h d (·) respectively represent the high-pass filters corresponding to the images H j (x, y), V j (x, y), D j (x, y). The specific type of filter is determined by the selected wavelet basis;

[0012] Subsequently, the images A j (x, y), H j (x, y), V j (x, y), D j (x, y) and f(x, y) are input into the feature point detection module.

[0013] Furthermore, the feature point detection module specifically performs the following operations:

[0014] Judge all pixel points in all the input images. For each pixel point p in the image f(x, y), the processing process is as follows: Select 16 pixel points on the ring centered on it for judgment, as shown in the following formula:

[0015]

[0016] In the formula, I p is the gray value of the pixel point p, and I n is the gray value of the nth pixel point on the ring, n = 1, 2, 3,..., 16, t is an adjustable threshold, and S n is the judgment result;

[0017] When there are N' consecutive pixel points among the surrounding 16 pixel points with |S n |≠0, the pixel point p is determined as a feature point, denoted as p c , and the coordinate of this feature point is denoted as L c =(x, y);

[0018] For image A j (x, y), H j (x, y), V j (x, y) and D j Perform feature point judgment on the pixel points in (x, y), and execute the same processing process as the image f(x, y). Denote the pixel points determined as feature points as p jc , and denote the coordinates of this feature point as L jc =(x, y). Since each wavelet transform will halve the length and width of the image, in order to represent p jc on f(x, y), multiply the horizontal and vertical coordinates of p jc by 2 j times, and denote it as p · jc ;

[0019] Compare the obtained feature point p · jc with p c to see if they are at the same pixel coordinates. If not, then retain both p · jc and p c , and denote the retained feature points as p l , and denote the coordinates of this feature point p l as L l =(x, y); if so, then delete p · jc and retain p c , and denote the remaining feature points as p l , and denote the coordinates of this feature point p l as L l =(x, y).

[0020] Furthermore, the feature description module specifically performs the following operations:

[0021] First, perform normalization based on the regional main direction on the detected feature point p l . Specifically: for each feature point, select a region with a length and width of H pixels around it as a window, define this region as D, and calculate the partial derivatives of each pixel point in the horizontal and vertical directions to obtain the gradient components G x , G y :

[0022]

[0023] In the formula, G x represents the partial derivative of the gray value I(x, y) at (x, y) with respect to x, and G yDenote the partial derivative of the grayscale value \(I(x,y)\) at \((x,y)\) with respect to \(y\);

[0024] Calculate the gradient magnitude \(m(x,y)\) and gradient direction \(\theta(x,y)\) for each pixel point:

[0025]

[0026] Quantize the gradient direction \(\theta(x,y)\) into a set angle interval. Divide \(360^{\circ}\) into \(M\) equal parts. In this way, the gradient direction of each pixel is mapped into a direction interval, and through weighted averaging, use the gradient magnitude \(m(x,y)\) as the weight, and each direction interval is represented by the cumulative sum of the gradient magnitudes of the included directions; the main direction \(\theta\) m is the direction with the largest cumulative sum of gradient magnitudes, that is, the interval corresponding to the maximum value:

[0027]

[0028] In the formula, \(i'\) represents the \(i'\)-th pixel point belonging to a certain direction interval;

[0029] Subsequently, normalize the main direction \(\theta\) m Rotate the region \(D\) by \(-\theta\) m degrees to obtain the region \(D'\) · and generate descriptors based on the pixel points within \(D'\) · as follows:

[0030] First, divide \(D'\) · into several sub-regions Within calculate the gradient direction and gradient magnitude \(\theta\) nj (x,y), \(m\) nj (x,y) of the \(j\)-th pixel point, and map the gradient direction of each pixel point within into \(M\) direction intervals, and calculate the mean values of the gradient direction and gradient magnitude within each direction interval and as the local direction descriptor \(N\) am of this direction interval:

[0031]

[0032] Arrange the \(M\) local direction descriptors \(N\) am in a clockwise direction to obtain the local descriptor \(N\) of the sub-region a :

[0033]

[0034] Arrange the local descriptors \(N\) aConcatenate them in order from left to right and from top to bottom into a complete narrative sub - N:

[0035]

[0036] Finally, the coordinates L l of the feature point p l =(x, y) and the narrative sub - N are used as a finally obtained feature.

[0037] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0038] 1. Using a wavelet image pyramid to simultaneously capture the details and global information of an image, with comprehensive feature expression; in addition, it has low sensitivity to noise, is not easily interfered by image noise, ensures the accuracy of feature extraction, and makes the generated features have rotational invariance, effectively improving the quality of feature matching.

[0039] 2. Adding main - direction normalization in the feature narrative module, making the generated narrative sub - N have rotational invariance and improving the quality of feature narration. Brief Description of the Drawings

[0040] Figure 1 is a framework diagram of the system of the present invention.

[0041] Figure 2 is a schematic diagram of feature - point detection.

[0042] Figure 3 is a schematic diagram of direction normalization. Detailed Embodiments

[0043] The present invention will be further described in detail below in conjunction with embodiments and the drawings, but the embodiments of the present invention are not limited thereto.

[0044] As Figures 1 to 3 shown, this embodiment discloses a grayscale - image feature - extraction system, including:

[0045] A wavelet image pyramid module, which is used to perform layer - by - layer decomposition on a grayscale image based on discrete wavelet transform, extract the low - frequency information representing the global part and the high - frequency information representing the details of the image, and use them as multiple - resolution representations of the image;

[0046] A feature - point detection module, which is used to detect the position of feature points in the image as the position information of the feature;

[0047] A feature - narrative module, which is used to generate a corresponding narrative sub - N after performing main - direction normalization on the neighborhood of the detected feature points as the narrative information of the feature, and finally output the position of the feature points and the narrative sub - N as the feature.

[0048] Specifically, the wavelet image pyramid module performs the following operations:

[0049] Perform a wavelet image pyramid representation of the grayscale image f(x,y) based on discrete wavelet transform. Among them, the image A j (x,y) that retains the low-frequency information and the image H j (x,y), V j (x,y), D j (x,y) are as follows:

[0050]

[0051] In the formula, f(i,k) represents the grayscale value of the image f(x,y) at the coordinate (i,k), n and m respectively represent the length and width of the image f(x,y), x and y respectively represent the horizontal and vertical coordinates in the image, · represents the multiplication operation, and the image H j (x,y) represents the horizontal detail, the image V j (x,y) represents the vertical detail, and the image D j (x,y) represents the diagonal detail. g(·) represents the low-pass filter, and h h (·), h v (·), h d (·) respectively represent the high-pass filters corresponding to the images H j (x,y), V j (x,y), D j (x,y). Specifically, what kind of filter it is is determined by the selected wavelet basis;

[0052] Subsequently, input the images A j (x,y), H j (x,y), V j (x,y), D j (x,y) and f(x,y) into the feature point detection module.

[0053] Specifically, the feature point detection module performs the following operations:

[0054] Judge all pixel points in all the input images. For each pixel point p in the image f(x,y), the processing process is as follows: Select 16 pixel points on the ring centered on it (see Figure 2 shown) for judgment, as shown in the following formula:

[0055]

[0056] In the formula, I p is the grayscale value of the pixel point p, and I nis the grayscale value of the nth pixel on the ring, where n = 1, 2, 3,..., 16, t is an adjustable threshold, and S n is the judgment result;

[0057] When there are N' consecutive pixels among the surrounding 16 pixels where |S n | ≠ 0, the pixel p is determined as a feature point, denoted as p c , and the coordinates of this feature point are denoted as L c = (x, y);

[0058] For image A j (x, y), H j (x, y), V j (x, y), and D j (x, y), perform feature point judgment on the pixels in them, execute the same processing process as the image f(x, y), and denote the pixels determined as feature points as p jc , and the coordinates of this feature point are denoted as L jc = (x, y). Since each wavelet transform will halve the length and width of the image, in order to represent p jc on f(x, y), multiply the horizontal and vertical coordinates of p jc by 2 j times, denoted as p · jc ;

[0059] Compare the obtained feature point p · jc with p c to see if they are at the same pixel coordinates. If not, then keep both p · jc and p c , and denote the retained feature points as p l , and the coordinates of this feature point p l are denoted as L l = (x, y); if so, then delete p · jc and keep p c , and denote the remaining feature points as p l , and the coordinates of this feature point p l are denoted as L l = (x, y).

[0060] Specifically, the feature description module specifically performs the following operations:

[0061] First, for the detected feature point p lPerform normalization based on the principal direction of the region. Specifically: For each feature point, select a region with a length and width of H pixels around it as a window, define this region as D, and calculate the partial derivatives of each pixel point in the horizontal and vertical directions to obtain the gradient components G x and G y :

[0062]

[0063] In the formula, G x represents the partial derivative of the gray value I(x, y) at (x, y) with respect to x, and G y represents the partial derivative of the gray value I(x, y) at (x, y) with respect to y; Calculate the gradient magnitude m(x, y) and gradient direction θ(x, y) of each pixel point:

[0064]

[0065] Quantize the gradient direction θ(x, y) into a set angle interval. Divide 360° into M equal parts, so that the gradient direction of each pixel is mapped into a direction interval, and through weighted averaging, use the gradient magnitude m(x, y) as the weight, and each direction interval is represented by the cumulative sum of the gradient magnitudes of the included directions; The principal direction θ m is the direction with the largest cumulative sum of gradient magnitudes, that is, the interval corresponding to the maximum value:

[0066]

[0067] In the formula, i′ represents the i′-th pixel point belonging to a certain direction interval;

[0068] Subsequently, normalize the principal direction θ m , rotate the region D by -θ m degrees to obtain the region D · (see Figure 3 shown), and generate a descriptor based on the pixel points within D · as follows:

[0069] First, divide D · into several sub-regions Within , calculate the gradient direction and gradient magnitude θ nj (x, y), m nj (x, y) of the j-th pixel point, and map the gradient direction of each pixel point within into M direction intervals, and calculate the mean values of the gradient direction and gradient magnitude in each direction interval and as the local direction descriptor N am of this direction interval:

[0070]

[0071] Arrange M local direction descriptors N am in a clockwise direction to obtain the local descriptor N of the sub-region a :

[0072]

[0073] Concatenate the local descriptors N of all sub-regions a in the order from left to right and from top to bottom into a complete descriptor N:

[0074]

[0075] Finally, take the coordinates L l =(x, y) of the feature point p l and the descriptor N as a finally obtained feature.

[0076] As described above, this is only the specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims described.

Claims

1. A grayscale image feature extraction system, characterized in that: include: The wavelet image pyramid module is used to decompose grayscale images layer by layer based on discrete wavelet transform, extracting low-frequency information representing the global image and high-frequency information representing the details, and using this information as a representation of the image at multiple resolutions; Feature point detection module, used to detect the position of feature points in the image as feature position information; The feature description module is used to normalize the main direction of the neighborhood of the detected feature point, generate the corresponding description sub-item as the description information of the feature, and finally output the position of the feature point and the description sub-item as the feature; The feature point detection module specifically performs the following operations: All pixels in all input images are judged. For each pixel p in the image f(x, y), the processing process is: select 16 pixels on the circle with it as the center for judgment, as expressed by the following formula: Where, I p is the grayscale value of pixel p, I n is the grayscale value of the nth pixel on the ring, n=1,2,3,...,16, t is an adjustable threshold, S n To judge the result; When the surrounding 16 pixels have N' consecutive pixels |S n When |≠0, the pixel point p is determined to be a feature point, denoted as p c , the coordinates of the feature point are marked as L c =(x,y); For the jth transformation, the low-frequency information of image A is retained. j (x,y) and the image H that retains high-frequency information j (x,y),V j (x,y),D j The pixel points in (x, y) are judged as feature points, and the same processing process as the image f(x, y) is performed. The pixel points judged as feature points are recorded as p jc , the coordinates of the feature point are marked as L jc =(x,y), since each wavelet transform will shorten the length and width of the image by half, in order to make p jc Can be expressed on f(x,y), p jc The horizontal and vertical coordinates are enlarged by 2 j times, denoted as p · jc ; The obtained feature point p · jc With p c Compare to see if they are at the same pixel coordinates, if not, then p · jc and p c All are retained, and the retained feature points are recorded as p l , the feature point p l The coordinates are marked as L l =(x,y); if yes, delete p · jc Keep p c , and record the remaining feature points as p l , the feature point p l The coordinates are marked as L l =(x,y); The feature description module specifically performs the following operations: First, the detected feature point p l Normalization based on the main direction of the region is performed. Specifically, for each feature point, a region with a length and width of H pixels is selected around it as a window, and the region is defined as D. The partial derivatives of each pixel in the horizontal and vertical directions are calculated to obtain the gradient component G. x , G y : Where G x Represents the partial derivative of the gray value I(x,y) at (x,y) with respect to x, G y Represents the partial derivative of the grayscale value I(x,y) at (x,y) with respect to y; Calculate the gradient magnitude m(x,y) and gradient direction θ(x,y) for each pixel: The gradient direction θ(x,y) is quantized to a set angle interval and divided into M parts by 360°. In this way, the gradient direction of each pixel is mapped to a direction interval, and the gradient amplitude m(x,y) is used as the weight by weighted averaging. Each direction interval is represented by the cumulative sum of the gradient amplitudes of the directions included. The main direction θ m It is the direction where the cumulative sum of the gradient amplitude is the largest, that is, the interval corresponding to the maximum value: Where i′ represents the i′th pixel point belonging to a certain direction interval; Then the main direction θ m Normalize and rotate region D by -θ m After the degree, we get area D · , and based on D · The narrator is generated for the pixel points within, as follows: First, D · Divided into several sub-areas exist Calculate the gradient direction and gradient amplitude θ of the jth pixel point nj (x,y),m nj (x,y), put The gradient direction of each pixel is mapped to M direction intervals, and the mean of the gradient direction and gradient amplitude in each direction interval is calculated. and As the local direction descriptor N of the direction interval am : Describe the M local direction sub- am Arrange in a clockwise direction to obtain sub-areas The local narrator N a : The local description sub-regions N a Concatenate from left to right and from top to bottom to form a complete narrative sub-N: Finally, the feature point p l Coordinates L l =(x,y) and the descriptor N as the final feature.

2. A grayscale image feature extraction system according to claim 1, characterized in that: The wavelet image pyramid module specifically performs the following operations: The grayscale image f(x,y) is represented by a wavelet image pyramid based on discrete wavelet transform, where the jth transform retains the low-frequency information of the image A j (x,y) and the image H that retains high-frequency information j (x,y),V j (x,y),D j The specific situation of (x,y) is as follows: Where f(i,k) represents the grayscale value of the image f(x,y) at coordinate (i,k), n and m represent the length and width of the image f(x,y), x and y represent the horizontal and vertical coordinates in the image, and represents the multiplication operation. The image H j (x,y) represents the horizontal details, image V j (x,y) represents the vertical details, image D j (x,y) represents the diagonal details, g(·) represents the low-pass filter, and h h (·),h v (·),h d (·) respectively represent the corresponding images H j (x,y),V j (x,y),D j (x,y) high-pass filter, the specific type of filter is determined by the selected wavelet basis; Then image A j (x,y),H j (x,y),V j (x,y),D j (x,y) and f(x,y) are input to the feature point detection module.

Images