A method for creating an infrared small target detection dataset and a training model

Abstract

This invention provides a method for creating an infrared small target detection dataset, including S1, segmenting a satellite infrared remote sensing image to obtain at least two background images with different environments; S2, creating simulated target points; S3, adding simulated target points to the background images and fusing them with the background images; and S4, expanding the dataset of simulated target points. This invention uses space-based remote sensing satellite images as a background, providing a large data foundation for remote sensing infrared small target detection; it achieves diversity in complex backgrounds, motion characteristics of small targets, morphological diversity of small targets, and type diversity of small targets, making the simulated data closer to the required real data and greatly enriching the content of the dataset.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing technology, specifically relating to a method for creating a multi-state small target detection dataset and a training model based on infrared remote sensing. Background Technology

[0002] With continuous breakthroughs in the infrared field and the growing demand for space surveillance and early warning, infrared detection has gradually received significant attention. Infrared target detection technology is a key technology for applications in the military, aviation, and aerospace fields. Infrared detection systems detect targets through thermal radiation; they are passive detection systems, not easily affected by interference. Detecting infrared targets in complex backgrounds is of great significance for the development of infrared early warning, precision guidance, and other fields.

[0003] However, detecting and identifying small infrared targets presents significant challenges, a long-standing and important problem. Small targets in infrared images occupy very few pixels, lacking detailed object features and extractable texture features, posing a major challenge for deep learning. Furthermore, during infrared imaging, factors such as background clouds, atmospheric radiation, and random noise directly affect image quality, significantly reducing target intensity; sometimes, the target's radiation intensity is even lower than that of non-target areas in the infrared image. These issues lead to target loss and high false alarm rates when using infrared-based target detection. Therefore, processing images of small infrared targets against complex backgrounds is a highly practical research direction, and the creation of large-scale infrared moving small target datasets is crucial.

[0004] In summary, infrared moving small target image datasets are the foundation for important problems such as detection, recognition, and situational awareness. However, in practice, it is difficult to obtain real images due to various difficulties. Therefore, in order to better develop the infrared small target detection problem, the creation of diverse small target simulation datasets based on infrared remote sensing is a very important task. Summary of the Invention

[0005] To address the shortcomings of the prior art, this invention proposes a method for creating an infrared small target detection dataset and a training model. This method simulates infrared moving small targets in diverse ways, including background diversity, morphological diversity, and motion diversity, to automatically obtain a large dataset of simulated infrared moving small targets, serving as the foundational data for subsequent algorithm improvement experiments. To achieve the above objectives, this invention employs the following specific technical solutions:

[0006] A method for creating an infrared small target detection dataset includes the following steps:

[0007] S1. Segment the satellite infrared remote sensing image to obtain at least two background images of different environments;

[0008] S2. Create simulated target points;

[0009] S3. Add simulated target points to the background image and perform fusion processing with the background image;

[0010] S4. Expand the dataset of simulation target points.

[0011] Preferably, step S1 includes the following steps:

[0012] S101. Acquire satellite infrared remote sensing images, including near-infrared band remote sensing images and mid-infrared band remote sensing images;

[0013] S102. Segment the near-infrared remote sensing image and the mid-infrared remote sensing image respectively to obtain at least two background images of different environments. The background images of different environments are the same size.

[0014] Preferably, step S2 includes the following steps:

[0015] S201. Set the simulation target point to a Gaussian kernel shape, and use a two-dimensional Gaussian kernel distribution model to simulate the existence shape and radiation characteristics of the simulation target point under the background image. The two-dimensional Gaussian kernel distribution model is as follows:

[0016]

[0017] Where A is the grayscale value of the center point of the two-dimensional Gaussian;

[0018] σ is the variance;

[0019] g(x,y) is the gray value at coordinates (x,y);

[0020] S202. Fill the Gaussian kernel shape of the simulated target point that does not complete a full pixel to achieve diversity in the Gaussian kernel shape of the simulated target point.

[0021] Preferably, step S3 includes the fusion processing of the simulated target point and the background image under two environments: the simulated target point is not occluded and the simulated target point is occluded.

[0022] Preferably, the fusion processing method for simulated target points in an environment where they are not occluded includes the following steps:

[0023] S311. Compare the radiation characteristics of the simulated target point with the background image where the simulated target point is located. Select the pixel value of the simulated target point or the background image where the simulated target point is located as the base value, and select the pixel value of the simulated target point or the background image where the simulated target point is located as the auxiliary value.

[0024] S312. Set the weight coefficients for the positions of the simulation target point and the background image where the simulation target point is located based on the base value and auxiliary value, and perform weighted processing on the gray values ​​of the positions of the simulation target point and the background image where the simulation target point is located respectively; then add the weighted gray value of the simulation target point to the weighted gray value of the background image.

[0025] Preferably, the fusion method in an environment where the simulated target point is occluded includes the following steps:

[0026] S321. Compare the radiation characteristics of the simulated target point with the background image where the simulated target point is located;

[0027] S322. Select the gray value of the simulation target point with high radiation characteristics or the gray value of the background image where the simulation target point is located as the gray value of the current position.

[0028] Preferably, step S4 includes the following steps:

[0029] S401. Set different inflection point positions, movement speeds, and movement angles in the x and y directions for the motion trajectory of the simulated target point;

[0030] S402. Set different initial positions and initial velocities for the simulation target point.

[0031] An infrared moving small target training model is provided, which is trained using the infrared small target detection dataset obtained by the above-mentioned method for creating infrared small target detection datasets.

[0032] The present invention can achieve the following technical effects:

[0033] 1. This invention uses space-based remote sensing satellite images as a background, providing a large data foundation for the detection of small targets in remote sensing infrared.

[0034] 2. This invention achieves diversity in complex backgrounds, diversity in the motion characteristics of small targets, diversity in the morphology of small targets, and diversity in the types of small targets, greatly enriching the content of the dataset.

[0035] 3. It has high accuracy and broad applicability. By adjusting the image diversity parameters, the simulated data can be made closer to the required real data, providing reliable data support for subsequent algorithm research. Attached Figure Description

[0036] Figure 1 This is a flowchart of a method for creating an infrared small target detection dataset according to an embodiment of the present invention;

[0037] Figure 2 This is a flowchart of a method for creating an infrared small target detection dataset according to another embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof.

[0039] The purpose of this invention is to provide a method for creating an infrared small target detection dataset and a training model. The method and training model for creating an infrared small target detection dataset provided by this invention will be described in detail below through specific embodiments.

[0040] Figure 1 The present invention illustrates the process of creating an infrared small target detection dataset according to an embodiment of the present invention, including the following steps:

[0041] S1. Segment the satellite infrared remote sensing image to obtain at least two background images of different environments;

[0042] S2. Create simulated target points;

[0043] S3. Add simulated target points to the background image and fuse them with the background image;

[0044] S4. Expand the dataset of simulation target points.

[0045] This invention achieves background image diversity by segmenting satellite infrared remote sensing images; it also achieves morphological diversity of simulated target points by setting the simulated target points to Gaussian kernel shape and considering sub-pixel offset; after fusing the prepared simulated target points with the segmented background image, it achieves motion and morphological diversity of the simulated target points by adjusting parameters such as inflection point position, initial position, motion speed, motion angle, and initial velocity, thereby greatly enriching the content of the infrared small target detection dataset and making the simulated data closer to the required real data.

[0046] S1. Segment the satellite infrared remote sensing image to obtain at least two background images of different environments.

[0047] In a preferred embodiment of the present invention, images in the near-infrared and mid-infrared bands are obtained from a satellite infrared remote sensing image database provided by an open-source satellite data platform; according to actual needs, the remote sensing images in the near-infrared and mid-infrared bands are segmented into multiple background images of the same size.

[0048] When selecting background images from the satellite infrared remote sensing image database, images with different amounts and thicknesses of clouds, or images with different levels of clutter complexity, or images with different overall brightness and contrast can be selected according to actual needs to achieve diversity in background images.

[0049] On the other hand, since the brightness of the same image varies in different bands, images in both the near-infrared and mid-infrared bands are selected to increase the diversity of the background image.

[0050] Finally, by setting up the program, the system can specify the amount, thickness, or amount of clutter in the background images to achieve automatic background image replacement.

[0051] S2. Create simulated target points.

[0052] The simulated target point is set to a Gaussian kernel shape, and its outward scattering radiation characteristics are used to simulate the target point.

[0053] In a preferred embodiment of the present invention, a two-dimensional Gaussian kernel distribution model is used to simulate the target point. The two-dimensional Gaussian kernel distribution model simulates the scattering pattern of the target point diffusion as follows:

[0054]

[0055] Where A is the grayscale value of the center point of the two-dimensional Gaussian. A two-dimensional Gaussian kernel surface model can be established on the horizontal plane with the center point of the two-dimensional Gaussian as the origin, presenting a shape that diffuses from the center to the surrounding areas.

[0056] σ is the variance;

[0057] g(x,y) is the gray value at coordinates (x,y) in the two-dimensional Gaussian surface model.

[0058] Considering sub-pixel offset, when the motion of the simulated target point does not satisfy a single pixel, the Gaussian kernel is truncated and filled. Specifically:

[0059] When the Gaussian kernel of n*n pixels moves and produces sub-pixel shift, the simulated target point is no longer completely stored in n*n pixels and shifts to the left, right or other positions. At this time, the gray values ​​that exceed the range of the pixel grid are truncated and the original symmetrical distribution is no longer retained.

[0060] At the same time, using the highest gray value existing in the original n*n pixel grid after offset as the reference value, Gaussian distribution is extended to fill the empty positions in the n*n pixel grid again. The gray value is inversely proportional to the distance, so that the pixels in the empty positions after filling are uniformly transitioned with the pixels in the surrounding positions.

[0061] In another embodiment of the present invention, step S2 can be performed first, that is, simulated target points are created first, and then step S1 is performed to obtain the background image.

[0062] S3. Add simulated target points to the background image and fuse them with the background image.

[0063] Step S3 includes the fusion processing of the simulated target point and the background image under two environments: the simulated target point is not occluded and the simulated target point is occluded.

[0064] In a preferred embodiment of the present invention, when the simulated target point to be fused is not obscured by clouds, the radiation characteristics of the simulated target point and its location in the background image are compared. If the radiation characteristics of the simulated target point are higher than those of its location in the background image, the gray value of the simulated target point is selected as the base value, and a larger weight coefficient is assigned to the simulated target point. The pixel value of its location in the background image is used as the auxiliary value, and a smaller weight coefficient is assigned to the background image. The gray values ​​of the simulated target point and its location in the background image are weighted respectively. The weighted gray values ​​of the simulated target point and its corresponding location in the background image are added together to complete the fusion of the simulated target point with the background image.

[0065] In another preferred embodiment of the present invention, when the simulated target point to be fused is obscured by clouds, the radiation characteristics of the simulated target point and the background image where the simulated target point is located are compared. If the radiation characteristics of the background image where the simulated target point is located are higher than the radiation characteristics of the simulated target point, then the gray value of the background image where the simulated target point is located is selected as the gray value of the current position, i.e., the following expression applies:

[0066]

[0067] Where s represents the radiation characteristics of the coordinates of the simulated target point;

[0068] S i,j The radiation characteristics at the location of the simulated target point in the background image after adding clutter noise;

[0069] S represents the radiation characteristics of the simulated target point.

[0070] S4. Expand the dataset of simulation target points.

[0071] In a preferred embodiment of the present invention, a custom program is used to set different motion speeds and angles in the x and y directions for the motion trajectory of the simulated target point, simulating different motion states of a small target. By setting different inflection point positions, speeds, and angles for the simulated target point, data of different simulated target points with diverse motion and types can be obtained.

[0072] On the other hand, the initial position and initial velocity of the simulated target point can be adjusted in various ways to increase the scale of the data and further ensure its realism.

[0073] Figure 2 The flowchart of another embodiment of the present invention for creating an infrared small target detection dataset is shown. See [link to documentation]. Figure 2 :

[0074] Using satellite imagery data provided by an open-source satellite data platform, background images with no cloud cover, many clouds, few clouds, thick clouds, a lot of clutter, high contrast, low contrast, low overall brightness, and high overall brightness were selected from a large database. These background images were obtained from satellite remote sensing data in the near-infrared and mid-infrared bands as background images required to create an infrared small target detection dataset.

[0075] The required background image is segmented using existing technology to obtain multiple background images of the required size. Then, through program settings, the corresponding background images are assigned to achieve automated replacement of diverse background images.

[0076] The simulation target point is set to a Gaussian kernel shape, and the sub-pixel offset is considered. That is, when the motion does not meet the requirement of a single pixel, the Gaussian kernel is truncated and filled to achieve the diversity of Gaussian kernel shape, that is, to achieve the diversity of simulation target point shape.

[0077] The simulated target point is fused with the background image in two cases: when the simulated target point is obscured by clouds and when it is not obscured by clouds.

[0078] After fusion, the motion trajectory of the simulated target point is set with different motion speeds and angles in the x and y directions through a custom program to simulate different motion states of the small target. By setting different inflection points, different speeds and angles, as well as the initial position and initial velocity of the simulated target point, the scale of the data is increased, which further ensures the realism and makes the simulation data closer to the required real data, providing reliable data support for subsequent algorithm research.

[0079] After processing through steps S1-S4, a large number of simulated target points with diverse backgrounds, motions, and types are output as the basic training set for subsequent detection, recognition, early warning, and situational awareness. The data in the labels are the coordinates of the current small target and the size of the Gaussian kernel of the small target.

[0080] An infrared moving small target training model is provided, which is trained using the infrared small target detection dataset obtained by the above-mentioned method for creating infrared small target detection datasets.

[0081] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0082] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

[0083] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for creating an infrared small target detection dataset, characterized in that, Includes the following steps: S1. Segment the satellite infrared remote sensing image to obtain at least two background images of different environments; S2. Create simulated target points; Step S2 includes the following steps: S201. Set the simulated target point to a Gaussian kernel shape, and use a two-dimensional Gaussian kernel distribution model to simulate the existence shape and radiation characteristics of the simulated target point under the background image. The two-dimensional Gaussian kernel distribution model is as follows: （1） in, A The grayscale value at the center point of the two-dimensional Gaussian sphere; For variance; coordinates ( x, y The grayscale value under ( ) S202. Fill the Gaussian kernel shape of the simulated target point whose motion does not fill a full pixel to achieve diversity in the Gaussian kernel shape of the simulated target point; when n n When sub-pixel offsets are generated by Gaussian kernel motion across individual pixels, grayscale values ​​exceeding the pixel's range are truncated so that the offset values ​​remain within the original range. n n The highest grayscale value in the pixel grid is the baseline value, and the values ​​are directed towards... n n The empty spaces in each pixel are filled with a Gaussian distribution, so that the gray value is inversely proportional to the distance. S3. Add the simulated target point to the background image and fuse it with the background image; the fusion processing method for the simulated target point in an environment where it is not occluded includes the following steps: S311. Compare the radiation characteristics of the simulated target point with the radiation characteristics of the background image where the simulated target point is located. If the radiation characteristics of the simulated target point are higher than the radiation characteristics of the background image where it is located, then select the gray value of the simulated target point as the base value and the pixel value of the background image where it is located as the auxiliary value. S312. Set the weight coefficients of the simulated target point and the background image where the simulated target point is located according to the base value and the auxiliary value, and perform weighted processing on the gray values ​​of the simulated target point and the background image where the simulated target point is located respectively; then add the weighted gray value of the simulated target point to the weighted gray value of the background image. The fusion method for simulated target points in occluded environments includes the following steps: S321. Compare the radiation characteristics of the simulated target point with the location of the background image where the simulated target point is located; S322. Select the gray value of the simulated target point with high corresponding radiation characteristics or the gray value of the background image where the simulated target point is located as the gray value after fusion at the current position. S4. Expand the dataset of the simulation target points.

2. The method for creating an infrared small target detection dataset according to claim 1, characterized in that, Step S1 includes the following steps: S101. Acquire satellite infrared remote sensing images, including near-infrared band remote sensing images and mid-infrared band remote sensing images. S102. The near-infrared remote sensing image and the mid-infrared remote sensing image are segmented respectively to obtain at least two background images of different environments, wherein the background images of different environments are of the same size.

3. The method for creating an infrared small target detection dataset according to claim 1, characterized in that, Step S3 includes the fusion processing of the simulated target point and the background image under two environments: when the simulated target point is not occluded and when the simulated target point is occluded.

4. The method for creating an infrared small target detection dataset according to claim 1, characterized in that, Step S4 includes the following steps: S401, for the simulation target point at... x direction and y Different inflection point positions, movement speeds, and movement angles are set for the movement trajectory in different directions; S402. Set different initial positions and initial velocities for the simulated target point.

5. A training method for an infrared moving small target recognition model, characterized in that, The infrared small target detection dataset obtained by the method for creating the infrared small target detection dataset as described in any one of claims 1-4 is used for training.

Images