A method for constructing a positive sample dataset for astronomical image denoising applications

Abstract

The application relates to an astronomical image denoising application-oriented positive sample data set construction method, relates to the field of image processing, and solves the problems that the data set for the astronomical image denoising and stray light removal task is relatively scarce and the processing performance needs to be improved. The method comprises the following steps: collecting and cutting star maps in different time periods; star point identification and positioning; generation and matching of double-triangle structures; star map alignment; and selection of positive and negative samples of the data set. The application is suitable for model training data construction of astronomical image denoising and stray light removal processing. In the subgraph cutting process, the un-contaminated or lightly contaminated areas are preferentially retained, the effectiveness and pertinence of the data are improved, and finally, the appropriate images are selected as the positive and negative samples of the data set by accurately identifying the star points, constructing the double-triangle structures, matching the star points by using a geometric hash algorithm and aligning the images. Under the same astronomical observation conditions, the denoising effect of the method is better, and the data processing efficiency is higher.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image denoising technology, and particularly relates to a positive sample data set construction method for astronomical image denoising application. BACKGROUND

[0002] Denoising and destraylighting of astronomical images is an important task in the field of computer vision, and is widely used in celestial body recognition, astronomical data processing, observation accuracy improvement and space exploration, etc. The existing data sets have good denoising effect in the case of weak stray light interference, but still face great challenges for astronomical images with strong stray light and complex noise.

[0003] Due to the relative lack of data sets for astronomical images used in denoising tasks, the training and performance improvement of existing algorithms are limited. The acquisition process of astronomical images is limited by observation equipment, observation conditions and optical properties of celestial bodies, and it is difficult to obtain a large amount of high-quality, noise-free real data at the same time. In order to solve this problem, some existing researches usually use artificial simulation of noise or stray light sources on clean star maps to generate training data. These synthetic data can help the network learn denoising features to some extent, but since the noise types in real astronomical images are complex and diverse, including photoelectric noise, atmospheric disturbance, thermal noise generated by equipment and stray light interference, these synthetic data are difficult to fully cover all cases in real observation environment. Therefore, although the existing data sets can play a certain role in training denoising models, they are difficult to fully meet the needs of real astronomical scenarios, resulting in suboptimal performance of the models in actual application, especially when dealing with complex weak signal celestial bodies and strong stray light interference, there is still a lot of room for improvement in denoising effect.

[0004] The present application proposes a method of constructing a data set using real images taken by an astronomical telescope. The star maps come from multiple observations of the same observation area at different time periods, so that the noise and stray light in the data set are closer to the real observation environment. This method not only reduces the complexity of data acquisition, but also effectively improves the performance of the denoising model when processing real astronomical images. SUMMARY

[0005] The present application is to solve the problems of the relative lack of data sets for astronomical image denoising and stray light removal tasks and the need to improve processing performance in the prior art, and provides a positive sample data set construction method for astronomical image denoising applications. The method of the present application is mainly applicable to the construction of model training data for astronomical image denoising and stray light removal processing. In the subgraph cropping process, the present application preferentially retains uncontaminated or lightly contaminated areas, improving the effectiveness and relevance of the data. Subsequently, the star points are accurately identified and located by a local maximum value detection algorithm, a double-triangle structure is constructed, and a geometric hashing algorithm is used for star point matching and image alignment, and finally the appropriate image is selected as the positive and negative samples of the data set. Under the same astronomical observation conditions, the data set constructed by the present application can achieve better denoising effect and higher data processing efficiency.

[0006] The positive sample data set construction method for astronomical image denoising applications comprises the following steps:

[0007] Step one, obtain star maps of different time periods, crop the same area from the star maps to obtain two subgraphs;

[0008] Step two, star point identification and positioning;

[0009] For each cropped subgraph, star point identification and positioning are performed by local maximum value detection;

[0010] Step three, generation and matching of double-triangle structure;

[0011] The star points identified in the two subgraphs are represented by a double-triangle structure to represent the geometric relationship between the star points; and the corresponding star point matching in the two subgraphs is realized;

[0012] Step four, according to the star point matching result in step three, complete the star map alignment;

[0013] Step five, selection of positive and negative samples of the data set;

[0014] According to the star map alignment result in step four, the uncontaminated area and the strongly contaminated area in the star map are cropped to generate the data set.

[0015] The present application has the following advantages:

[0016] The present application starts from star point recognition and geometric relationship modeling, processes the subgraph area of the astronomical image, so as to subsequent star map alignment and data set construction. The present application cuts the star map shot at different time periods, adopts local maximum value detection algorithm to accurately recognize and locate the star points in the subgraph, and improves the accuracy of star point detection through dynamic adaptive threshold method. After the star point recognition, the double triangle structure is constructed according to the recognition result, four star points are selected and two non-intersecting triangles are constructed, the geometric relationship features between the star points are extracted, including the side length and the angle, and the geometric hash code is generated to realize the efficient star point matching. After the star point matching is completed, the coordinates of the star points in the two star maps are uniformly transformed by using the affine transformation, so that the matched star points can be accurately aligned in the original coordinate system, thereby improving the accuracy of the star map alignment. After the star map alignment is completed, the seriously polluted area and the relatively clean area are cut according to the light pollution degree, so as to generate diversified training data set for the astronomical image denoising model.

[0017] The present application solves the problems of low star point matching accuracy and insufficient generalization ability of the denoising model under the conditions of light pollution and noise by cutting the star map based on the real scene, accurately recognizing the star points, modeling the geometric relationship and affine alignment, and simultaneously adopting the cutting strategy of different areas. Finally, the constructed denoising data set can provide higher denoising performance and better star point retention effect in the actual application of the model, and the processing efficiency and quality of the astronomical image are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The flowchart of the positive sample data set construction method for the astronomical image denoising application of the present application.

[0019] Figure 2 The flowchart of the double triangle structure matching method.

[0020] Figure 3 The overall schematic diagram of the double triangle structure matching star map. Wherein, (a) represents the strong pollution and weak pollution star map, (b) represents the image after detecting the star points by using the local maximum value detection algorithm, c(a) represents the schematic diagram of the double triangle structure, and (a) represents the aligned image after the affine transformation.

[0021] Figure 4 The visual quality comparison diagram of the denoising image of the denoising network trained by different data sets. a represents the original star map, b represents the denoising image of the network trained by the data set simulating the astronomical noise and stray light, and c represents the denoising image of the network trained by the data set of the present application.

[0022] Figure 5 The visual quality comparison diagram of the denoising image of the denoising network trained by different data sets. a represents the original star map, b represents the denoising image of the network trained by the data set simulating the astronomical noise and stray light, and c represents the denoising image of the network trained by the data set of the present application.

[0023] Figure 6 A visual quality comparison chart of denoised images trained by different data sets. a represents the original star map, b represents the denoised image of the network trained by the simulation data set, and c represents the denoised image of the network trained by the data set of the application. DETAILED DESCRIPTION

[0024] DETAILED DESCRIPTION Figures 1 to 6 To illustrate this embodiment, a positive sample data set construction method for astronomical image denoising application is described. For the star map to be processed, the recognition accuracy of star points is improved by using a local maximum value detection algorithm with adaptive threshold. Subsequently, a geometric relationship model between star points is constructed based on the star point detection results, and efficient star point alignment is achieved through geometric hash matching, thereby ensuring the alignment accuracy and consistency of the star map. Based on the alignment results, different noise and stray light pollution regions are further cropped to provide high-quality samples for the astronomical image denoising data set, so that the denoising performance of the subsequent model is better and the detail preservation effect of star points is better.

[0025] As shown in Figure 1 The positive sample data set construction method for astronomical image denoising application described in this embodiment mainly includes the following steps:

[0026] Step 1: Star map collection and cropping at different times;

[0027] The same region of the star map at different times is cropped, and the part with less light pollution is preferentially retained, so that the retained image region is relatively clean and the number of star points is moderate.

[0028] In this embodiment, the same region is selected from two star maps F1 and F2 of the same sky region taken at different times to obtain sub-maps G1 and G2. When cropping, the part with less light pollution is preferentially retained, so that the retained image region is relatively clean and the number of star points is moderate.

[0029] Step 2: Star point recognition and positioning;

[0030] According to the subgraph after cutting according to step one, the star points are accurately identified and positioned by a local maximum detection algorithm, and the coordinate information of the identified star points is retained. Since the star points in the astronomical image are represented as high-brightness point objects, the local maximum value of the brightness value significantly higher than the surrounding pixels can be found by scanning the neighborhood of the image, and the potential star point position is identified, which has good robustness especially in a more complex background. For each local maximum point, its pixel coordinates in the image are directly obtained as the accurate position of the star point, and the star point coordinates in the subgraph G1 are defined as (x i ,y i ), and the star point coordinates in the subgraph G2 are defined as (x i ’,y i ’).

[0031] Step three, generation and matching of double triangle structure (generation of local star point topology);

[0032] The star points identified in the two subgraphs are represented by a double triangle structure to represent the geometric relationship between the star points.

[0033] First, by selecting four star points and constructing two non-intersecting triangles therefrom, the geometric relationship between the star points is formed.

[0034] Suppose that in a graph, there are multiple star points, and the coordinates of the star points are (x i ,y i ). The distance between the four star points is calculated by the following formula:

[0035]

[0036] Where d ij represents the Euclidean distance between star point i and star point j.

[0037] Suppose that the four star points in the subgraph G1 are A(x1,y1), B(x2,y2), C(x3,y3), and D(x4,y4). By comparing the distances between the four star points, the three star points A(x1,y1), B(x2,y2), and C(x3,y3) with the shortest distance are selected to form the first triangle, and the remaining star point D(x4,y4) automatically becomes a vertex in the second triangle.

[0038] Then, by calculating the distance from the remaining star point D(x4,y4) to each vertex of the first triangle, the two vertices with shorter distances are selected to ensure that the second triangle is not too distorted and the geometric relationship between the star points is more stable and natural. The distance from point D(x4,y4) to each vertex of triangle ABC is calculated:

[0039]

[0040] Select two points with shorter distance, say A and B, then the second triangle is DAB. Similarly, we can assume that the four star points in G2 are E, F, G, and H, respectively. The three closest points E, F, and G form the first triangle EFG, and the remaining star point H forms the second triangle HEF with the two closest points E and F.

[0041] Two triangles are formed by four star points in each subgraph, and their geometric features such as edge length and angle are recorded. By calculating the length of the three edges, the shape characteristics of the triangle can be obtained.

[0042]

[0043] The size of each angle can be calculated by the cosine law. For example, the calculation formula of angle θ A is as follows:

[0044]

[0045] Similarly, angles θ B and θ C can also be calculated by the cosine law.

[0046] The area of a triangle can also be used as a geometric feature. Through the calculation of the area, the shape of the triangle can be further described. The area can be calculated by Heron's formula, which is as follows:

[0047]

[0048] Similarly, the area of triangle DAB, Area2, can be obtained.

[0049] These two triangles constitute the local topological relationship between star points. After extracting the geometric features of the two triangles, a unique geometric hash code is generated through the following steps for star point matching.

[0050] All geometric features (edge length, angle, area) are combined into a feature vector, which can completely describe the geometric shape of each triangle:

[0051] F = [L AB , L BC , L CA , θ A , θ B , θ C , Area1, L DA , L AB , L BD , θ D , θ A , θ B , Area2]

[0052] To generate hash codes, continuous features need to be discretized, dividing each feature value into several discrete intervals, and the values ​​in the feature vector are quantized. Assuming each feature value is quantized into k intervals, then each feature value f... i It is mapped to a discrete eigenvalue d i ,in:

[0053] d i =quantize(f i k)

[0054] The discretized feature values ​​are combined into a unique hash code. Assume there are n discretized feature values ​​d1, d2, ..., dn. n Then the final geometric hash code can be represented as:

[0055]

[0056] In this way, the hash code maps the different geometric features of the star points to a unique integer value. By matching similar hash values, the corresponding set of star points in two images can be quickly found. Assuming the hash codes in the two images are H1 and H2, the matching process can be determined by the following formula:

[0057] |H1-H2|<∈

[0058] Here, ∈ represents a set tolerance range to allow for minor errors in matching. If the condition is met, the sets of stars represented by the two hash codes are considered to match.

[0059] Step 4: Align the star chart;

[0060] After matching the star points obtained in step three, an affine transformation needs to be performed on the star point coordinates of the two images to ensure that they can be correctly aligned on the original image, and to ensure that the corresponding star points in the two sub-images can be accurately overlapped in the original coordinate system.

[0061] Suppose that a matching set of star points has been found in two images. Let the coordinates of the star points in the first image be (x1, y1), (x2, y2), ..., (x...). n y n The coordinates of the matching star points in the second image are (x′1, y′1), (x′2, y′2), ..., (x′1, y′2). n y′ n After obtaining the matched star points, the coordinates of the star points in the two images need to be transformed so that they can be correctly aligned on the original images. The calculation of the affine matrix is ​​usually based on a set of matched star point pairs. The optimal transformation matrix is ​​found by solving a least-squares problem to keep the geometric relationship between the two sets of star points consistent.

[0062] Affine transformation can handle linear transformations such as scaling, rotation, translation, etc. to ensure that corresponding star points in two subgraphs can be accurately overlapped in the original coordinate system. Affine transformation can be represented in the following matrix form:

[0063]

[0064] where (x i ,y i ) is the star point coordinate in the first graph, (x′ i, y′ i ) is the coordinate of the corresponding star point in the second graph, a, b, c, d are the parameters of the affine transformation matrix, t x and t y are the translation amounts.

[0065] In order to calculate the six parameters a, b, c, d, tx, ty in the affine matrix, at least three pairs of matched star points are needed. For each pair of matched star points, the following linear equation can be written:

[0066] x′ i = ax i + by i + t x

[0067] y′ i = cx i + dy i + t y

[0068] Substituting each pair of matched star points into the equation, a system of linear equations can be obtained. For n pairs of star point matches (n≥3), the equation system can be written as:

[0069]

[0070] X′ = AP

[0071] where X’ is the coordinate of the matched target star point, A is the known star point coordinate matrix, and P is the affine matrix parameter to be solved.

[0072] The equation system is solved by the least squares method to find the best affine matrix parameters a, b, c, d, tx, ty that minimize the matching error. The objective function to be minimized is:

[0073]

[0074] By solving this minimization problem, the affine transformation matrix P can be obtained, and all star points in the second image are transformed to map them into the coordinate system of the first image. For any star point (x, y), the coordinate after the affine matrix transformation is:

[0075]

[0076] Through the transformation, the matching regions in the star maps taken at different time periods can be well aligned.

[0077] Step five, selection of positive and negative samples of the data set;

[0078] After the star maps are aligned in step four, the pollution-free (or weak pollution) region and the strong pollution region in the star maps can be further cropped to generate the data set. The strong pollution region contains a part with excessively high brightness or excessively large background noise as the negative sample in the data set, and the pollution-free or weak pollution region saves relatively real star points and background for use as the positive sample part in the data set for subsequent denoising model training. Through the star map data taken at different time periods in the real scene, an astronomical image denoising data set covering various pollution types can be constructed to improve the generalization performance and denoising effect of the model under actual observation conditions.

[0079] Specific implementation method two, combination Figures 1 to 6 The embodiment is described, and the embodiment is an embodiment of the positive sample data set construction method for the astronomical image denoising application described in the specific implementation method one:

[0080] 1. Working conditions;

[0081] In this experiment, an intel core i7-7700K CPU@4.20ghz*8 processor is used, a PC running Windows 10 is used, one GeForce GTX 1070Ti graphics card is used, and the programming language is Python.

[0082] 2. Experimental content and result analysis;

[0083] As shown in Figure 2 , it is a flowchart for generating the topology of local star points and implementing star map alignment, which briefly summarizes the subdivision process. Further, as shown in Figure 3 , when the star map matching is implemented, the local maximum value detection algorithm is used to identify the star points in the map to better generate a suitable double triangle structure. After the double triangle structure is matched, the affine transformation matrix is used to match and align the star map, so that the star map matching efficiency is greatly improved.

[0084] In order to better compare the performance of the data set construction method proposed in this paper on star image denoising and stray light removal, a subjective visual quality comparison analysis will be carried out.

[0085] As shown in Figure 4 , by comparing the original star image, it can be seen that the denoising effect of the data set construction method of the application after training the network has better subjective visual quality than the denoising effect of the network trained by the data set constructed by simulating astronomical noise and stray light, and there is no phenomenon of missing star points, and the original image can be approached to a great extent. As shown in Figure 5 、 Figure 6 , the data set of the application and the simulated data set are respectively trained to remove noise from the star image, and by comparing with the original star image, it can be found that the method of the application compared with the simulated data set can better preserve and reconstruct the details and brightness information of the target in the star image.

[0086] As shown in Figures 4 to 6 , when denoising the star image, the application can better preserve the detail information and brightness information of the spatial target, and is superior to other methods in subjective visual quality.

[0087] The above experimental results show that for the star image to be processed, the application improves the star point recognition accuracy through the local maximum star point detection algorithm, and subsequently establishes the geometric relationship between the star points using a double-triangle structure, selects star points to construct two non-intersecting triangles, extracts edge length and angle features to generate a geometric hash code, thereby realizing efficient and accurate star point matching. The use of the double-triangle structure can fully utilize the geometric characteristics between the star points, enhance the robustness and noise resistance of the matching. Based on the matching result, the star image can be further accurately aligned and cropped, and an astronomical image denoising data set covering various pollution types can be constructed. Finally, the processed star image has better denoising effect and better star point detail preservation effect.

[0088] The above is only the preferred embodiment of the application, and it should be noted that for those skilled in the art, without departing from the principles of the application, several improvements and refinements can be made, and these improvements and refinements should also be considered as the protection scope of the application.

Claims

1. A method for constructing a positive sample dataset for astronomical image denoising applications, characterized in that: The method is realized by the following steps: Step one, obtaining star maps of different time periods, and cutting the star maps to obtain two sub-maps in the same area; Step two, star point identification and positioning; In each sub-map after cutting, star points are accurately identified and positioned by a local maximum value detection algorithm, and the coordinate information of the identified star points is retained.

5. The method for constructing a positive sample data set for astronomical image denoising applications according to claim 1, characterized in that: Similarly, let the four star points in the sub-map G2 be E, F, G, and H, the three closest points E, F, and G form a first triangle EFG, and the remaining star point H forms a second triangle HEF with the two closest points E and F; Then, a double triangle is formed by the four star points in each of the two sub-maps, and the geometric features of the length, angle, and area of the double triangle are recorded; After extracting the geometric features of the double triangle, a unique geometric hash code is generated for the matching of the star points. The specific process of step four is: Affine transformation is performed on the star point coordinates of the two sub-maps, the least squares method is used to calculate the affine matrix parameters, and affine transformation is used to align the star maps.

2. The method of claim 1, wherein the method is applied to astronomical image denoising.

7. The method for constructing a positive sample data set for astronomical image denoising applications according to claim 6, characterized in that:

3. The method of claim 2, wherein: Each pair of matched star points is substituted into the following equation to obtain a system of linear equations; 4. The method of claim 3, wherein: The neighborhood scanning is performed on the subgraph to find local maximum value higher than surrounding pixels in brightness value, to identify potential star point positions, for each local maximum value point, the pixel coordinates thereof in the image are directly obtained as the accurate position of the star point, the star point coordinates in the subgraph G1 are , and the star point coordinates in the subgraph G2 are . For n pairs of star point matches, n ≥ 3, the equation set is written as: First, set the four star points in subgraph G1 as ; Select the three closest stars by comparing the distances between the four stars Form the first triangle, stars Automatically become one vertex of the second triangle, stars The distance to each vertex of the first triangle, select the two closest vertices, set And Two stars, then the second triangle is ; ​ ​ ​ Finally, the corresponding star point sets in the two subgraphs are found by matching similar hash values; let the hash codes of the two subgraphs be and respectively, and the judgment is made by the following formula: ； wherein, is a set tolerance range, and if the condition is satisfied, the star point sets represented by the two hash codes are considered to be matched.

6. The method of claim 1, wherein: ​ ​ ​ Set in two subgraphs, find a set of matching star points; set the star point coordinates of the first subgraph as , the matching star point coordinates of the second subgraph are ( , ),( , ),…,( , ); after obtaining the matching star points, transform the star point coordinates of the two subgraphs, and the affine transformation adopts the following matrix form: ； wherein, is the coordinate of a star point in the first sub-image, is the coordinate of the corresponding star point in the second sub-image, is a parameter of the affine transformation matrix, and is a translation amount; ​ ； ； ​ ； ； In the formula, X ’ is the matched target star point coordinate, A is the known star point coordinate matrix, and P is the affine matrix parameter to be solved. The system of equations is solved by least squares to find the optimal affine matrix parameters a, b, c, d, t x ,t y such that the matching error is minimized; the objective function to be minimized is: ； By solving for and minimizing the objective function, the affine transformation matrix P is obtained. This matrix is ​​then used to transform all the star points in the second sub-figure, mapping them to the coordinate system of the first sub-figure. For any star point (… The coordinates after affine matrix transformation are ( , After affine transformation, matching regions in star images taken at two different times are aligned.

Images