A method for detecting infrared dim and weak small targets in time and space domain based on three-dimensional tensor filtering

Abstract

The application discloses a kind of infrared dim small target space-time detection methods based on three-dimensional tensor filtering, comprising: constructing three-dimensional image tensor;Three-dimensional image tensor is carried out sliding window filtering operation, and target saliency image and three-dimensional dynamic clutter suppression image are obtained;Three-dimensional image tensor is carried out pixel-level highlight noise discrimination, and noise discrimination factor weighting diagram is obtained;Based on target saliency image, three-dimensional dynamic clutter suppression image and noise discrimination factor weighting diagram, enhanced target saliency image is obtained;Through adaptive threshold, the real target of enhanced target saliency image is segmented and extracted, and infrared dim small target space-time detection is completed.The application enhances target saliency by three-dimensional hollow mean filter template;Based on the design of dynamic clutter suppression method of regional motion complexity, the complex background area of irregular motion is inhibited;Based on the design of noise discrimination feature of local energy mutation characteristics, pixel-level high-intensity noise is effectively inhibited.

Description

Technical Field

[0001] This invention belongs to the field of infrared faint and small target detection and processing, specifically relating to a spatiotemporal domain detection method for infrared faint and small targets based on three-dimensional tensor filtering. Background Technology

[0002] Infrared imaging technology, due to its advantages such as strong anti-interference capability and long range, has been widely used in monitoring low-altitude infrared small targets, such as drones. However, infrared small targets have low energy and few pixels, lacking structural and textural information in infrared images, which brings great difficulties to small target detection. In addition, complex background interference in the scene, such as swaying trees and pixel-level bright noise, easily generates a large number of false alarms. Current mainstream infrared small target spatiotemporal domain detection algorithms, such as STLCF, STLDM, and STLCM, assume that there is a significant difference in grayscale values ​​between the target and the background, and that the target's movement causes fluctuations in temporal energy, ignoring the influence of moving background. However, in complex low-altitude scenes, the target's grayscale value is low, and the grayscale difference between it and the background is small, making it easy to be submerged in the background. At the same time, there are often moving background clutters in the scene, such as swaying trees, which also cause energy fluctuations in pixels. Therefore, the above methods can easily misidentify moving background pixels as target pixels, resulting in poor background suppression. In addition, pixel-level high brightness noise (PNHB) interference exists in complex low-altitude scenes, which can also lead to more false alarms in the detection results, making it difficult for the above methods to meet the detection performance requirements. Summary of the Invention

[0003] To address the detection problem of small, faint infrared targets against complex low-altitude backgrounds, this invention proposes a spatiotemporal detection method for small, faint infrared targets based on three-dimensional tensor filtering. Based on the spatiotemporal energy distribution characteristics of small infrared targets and the temporal variation characteristics of the energy in the target's neighborhood, a three-dimensional tensor filtering method is designed to enhance target saliency. Simultaneously, considering the structural distribution characteristics of temporal grayscale value changes in moving complex background regions, a dynamic clutter suppression method based on the complexity of regional motion is designed, utilizing a three-dimensional dynamic clutter suppression image to adjust background suppression performance. Furthermore, based on the characteristic that pixel-level high-brightness noise typically exhibits energy abrupt changes at a single pixel, a noise discrimination feature based on local grayscale value abrupt changes is proposed to effectively suppress pixel-level high-brightness noise.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A spatiotemporal domain detection method for faint infrared targets based on three-dimensional tensor filtering includes the following steps:

[0006] Constructing a three-dimensional image tensor for infrared sequence images against a complex low-altitude background;

[0007] A sliding window filtering operation is performed on the three-dimensional image tensor to obtain a target saliency image and a three-dimensional dynamic clutter suppression image; pixel-level highlight noise discrimination is performed on the three-dimensional image tensor to obtain a noise discrimination factor weighted image;

[0008] An enhanced target saliency image is obtained based on the target saliency image, the three-dimensional dynamic clutter suppression image, and the noise discrimination factor weighted image.

[0009] By segmenting the enhanced target saliency image using an adaptive threshold, the real target is extracted, thus completing the spatiotemporal detection of small, faint infrared targets.

[0010] Preferably, the method for constructing the three-dimensional image tensor includes:

[0011] Get the original single-frame image of the current t0 frame. Where t0≥2;

[0012] The current frame image is stacked pixel by pixel along the time dimension with the k-1 neighboring frame images to form a three-dimensional image tensor F;

[0013] The formula for calculating the 3D image tensor F is as follows:

[0014] In the formula, f t Let t be the image of the t-th frame, with a size of m×n. The size of the 3D image tensor F is m×n×k, where m, n, and k are the scales in the horizontal, vertical, and time axes, respectively.

[0015] Preferably, the method for obtaining a target saliency image by performing a sliding window filtering operation on the three-dimensional image tensor includes:

[0016] For each pixel (x, y, t0) in the t0th frame of the 3D image tensor F, the target energy distribution characteristics are extracted using the 3D trapezoidal filter template M, and the target energy distribution gain image is calculated.

[0017] For each pixel (x, y, t0) in the t0th frame of the three-dimensional image tensor F, the temporal fluctuation information of the target neighborhood energy is extracted using the three-dimensional hollow mean filter template H, and the cumulative image of the target motion energy is calculated.

[0018] Based on the target energy distribution gain image and the target motion energy accumulation image, calculate the target saliency image;

[0019] The methods for calculating the target energy distribution gain image include:

[0020]

[0021] Δf(x,y,t)=f(x,y,t)-f(x,y,t-1),

[0022]

[0023] In the formula, (x,y,t) is the pixel position, f(x,y,t) is the pixel value of the t-th frame image at (x,y), (i,j,t) is the pixel position, Δf is the inter-frame difference image, M(i,j,t) is the weight value of the t-th frame of the three-dimensional trapezoidal filter template at pixel (i,j), P(x,y) is the pixel value of the target energy distribution gain image P at position (x,y), and c1 is the number of extended pixels;

[0024] Methods for calculating the cumulative energy image of a target motion include:

[0025] E(x,y)=Q(x,y,t0)-Q(x,y,t0-k+1),

[0026]

[0027]

[0028] In the formula, (x,y,t) is the pixel position, Q(x,y,t) is the pixel value at (x,y) in the t-th frame image, N is the number of pixels in the estimated sample of pixel (x,y,t) that satisfy max(|i|,|j|)>4, (i,j,t) is the pixel position, H(i,j,t) is the weight value of the t-th frame of the three-dimensional hollow mean filter template H at (i,j), E(x,y) is the pixel value of the target motion energy accumulation image E at (x,y), c2 is the number of expanded pixels, Q(x,y,t0) is the average energy of the background region, and Q(x,y,t0-k+1) is the average energy of the superimposed target energy.

[0029] Methods for calculating target saliency images include:

[0030] D = PE,

[0031] In the formula, (x,y) is the pixel position, P(x,y) is the pixel value of the target energy distribution gain image P at (x,y), E(x,y) is the pixel value of the target motion energy accumulation image E at (x,y), and D(x,y) is the pixel value of the target saliency image D at (x,y).

[0032] Preferably, the method for performing a sliding window filtering operation on the three-dimensional image tensor to obtain a three-dimensional dynamic clutter-suppressed image includes:

[0033] For each pixel (x, y, t0) in the t0th frame of the 3D image tensor F, using the c3×c3 local gray-scale mean b of the previous k-1 frames at (x, y), r Together they form a time-series grayscale value sequence {b r}, r=1,...,k;

[0034] For the time-series grayscale value sequence {b r The phase space is reconstructed for r = 1,...,k to obtain the gray value sequence B(g);

[0035] Rearrange each grayscale value sequence B(g) in ascending order, and form a symbol sequence S(g) based on the column position index of each grayscale value;

[0036] Calculate the weighted probability P of each symbol sequence. g ;

[0037] Based on the weighted probability P g Calculate the three-dimensional dynamic clutter suppression image G(x,y);

[0038] The expression for calculating the grayscale value sequence B(g) is as follows:

[0039] B(g)=[b r (g),b r (g+T),...,b r [g+(D-1)T)], g=1,...,N

[0040]

[0041] In the formula, B(g) is {b r The g-th gray value sequence obtained after phase space reconstruction, where D is the embedding dimension, T is the hysteresis time, N = k - (D-1)T is the total number of reconstructed sequences, and M... s This represents the total number of pixels in the local spatial region at (x,y).

[0042] The expression for calculating the symbol sequence S(g) is:

[0043] S(g) = [j1,...j l ...j D ], l≤D, g≤D! ;

[0044] In the formula, j l Let D! be the column position index of the l-th element in the reconstructed sequence S(g), and let D! be the total number of different sorting structures.

[0045] The expression for calculating the three-dimensional dynamic clutter suppression image G(x,y) is as follows:

[0046]

[0047] P g =w g *p g ,

[0048]

[0049] In the formula, p g Let w be the probability of the g-th symbol sequence appearing. g Let η be the weight of the g-th symbol sequence, and η be the threshold.

[0050] Preferably, the method for performing pixel-level highlight noise discrimination on the three-dimensional image tensor to obtain a noise discrimination factor-weighted image includes:

[0051] For each pixel (x, y, t0) in the t0th frame of the three-dimensional image tensor F, noise discrimination features d are obtained based on the gray value of each pixel and the gray value distribution of the corresponding eight neighboring pixels.

[0052] Based on the noise discrimination feature d, noise is judged for the corresponding pixel (x,y,t0), and the pixel that is judged as pixel-level high brightness noise is suppressed to obtain a noise discrimination factor weighted image.

[0053] The expression for obtaining the noise discrimination feature d is:

[0054] The expression for obtaining the noise discrimination factor weighted image is:

[0055] In the formula, f(x,y,t0) is the gray value of the image at position (x,y) in frame t0, N is the eight neighboring pixels of pixel (x,y), β is the noise discrimination threshold, and NS(x,y) is the weighted value of the noise discrimination factor at (x,y).

[0056] Preferably, the method for obtaining an enhanced target saliency image based on the target saliency image, the three-dimensional dynamic clutter suppression image, and the noise discrimination factor weighted image includes:

[0057]

[0058] Where (x,y) is the pixel position, D(x,y) is the pixel value of the target saliency image D at (x,y), G(x,y) is the pixel value of the three-dimensional dynamic clutter suppression image G at (x,y), NS(x,y) is the noise discrimination factor of the noise discrimination factor weighted image NS at (x,y), and ETSM(x,y) is the pixel value of the enhanced target saliency image ETSM at (x,y).

[0059] Preferably, the method for segmenting the enhanced target saliency image using an adaptive threshold includes:

[0060] In the enhanced target saliency image ETSM, pixels with values ​​greater than the threshold Th are considered to be real targets.

[0061] The calculation method for Th is as follows:

[0062] Th=μ+q·σ

[0063] Where μ and σ are the mean and standard deviation of the enhanced target salient image, respectively, and q is a manually set value.

[0064] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0065] (1) To address the problem of enhancing the infrared faint and small target under complex low-altitude background, based on the temporal fluctuation characteristics of the target neighborhood energy caused by the continuous motion of the infrared faint and small target, a three-dimensional hollow mean filter template is designed to filter the sequential image, and the energy of the target is accumulated in time over multiple frames by differential operation, which further enhances the saliency of the infrared faint target.

[0066] (2) To address the problem of dynamic clutter suppression in complex low-altitude backgrounds, a dynamic clutter suppression method based on the structural distribution characteristics of temporal gray value changes in clutter regions was designed. The three-dimensional dynamic clutter suppression image was used to enhance the suppression capability of complex background regions with irregular motion.

[0067] (3) To address the interference problem of pixel-level high brightness noise in complex low-altitude backgrounds, based on the prior information that pixel-level high brightness noise usually has energy mutation characteristics at a single pixel and the gray distribution inside the target is relatively consistent, a noise discrimination feature based on the local gray value mutation characteristics is proposed, which can effectively suppress pixel-level high brightness noise. Attached Figure Description

[0068] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are 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.

[0069] Figure 1 This is a schematic diagram of the spatiotemporal domain detection method for small, faint infrared targets based on three-dimensional tensor filtering in an embodiment of the present invention.

[0070] Figure 2 This is a schematic diagram of the target energy distribution gain image, the target motion energy accumulation image, and the target salience image in an embodiment of the present invention;

[0071] Figure 3 This is a schematic diagram of the target saliency image, the three-dimensional dynamic clutter suppression image, the noise discrimination factor weighted image, and the enhanced target saliency image in an embodiment of the present invention;

[0072] Figure 4 This is a schematic diagram of the infrared weak target detection results and a magnified view of a local target in an embodiment of the present invention. Detailed Implementation

[0073] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] Example 1

[0076] The images were acquired using a long-wave infrared staring camera with a fixed orientation, resulting in a sequence of infrared images against a complex low-altitude background. The image size was 512×640 pixels, with a total of 60 frames. The flowchart of the method of this invention is shown below. Figure 1 As shown, the input image is for the t0th frame (t0≥2). The specific processing steps of the method are as follows:

[0077] Constructing a three-dimensional image tensor for infrared sequence images against a complex low-altitude background;

[0078] A sliding window filter operation is performed on the 3D image tensor to obtain a target saliency image and a 3D dynamic clutter-suppressed image; pixel-level highlight noise discrimination is performed on the 3D image tensor to obtain a noise discrimination factor weighted map.

[0079] An enhanced target saliency image is obtained based on the target saliency image, the three-dimensional dynamic clutter suppression image, and the noise discrimination factor weighted image.

[0080] By segmenting the enhanced target saliency image using an adaptive threshold, the real target is extracted, thus completing the spatiotemporal detection of small, faint infrared targets.

[0081] In this embodiment, a three-dimensional image tensor is constructed as input. The method for constructing the three-dimensional image tensor includes: obtaining the original single-frame image of the current t0 frame. The current frame image is then stacked pixel-by-pixel along the time dimension with the k-1 neighboring frame images to form a three-dimensional image tensor F, which is expressed by the following formula:

[0082]

[0083] Where f t Let t be the image of the t-th frame. The size of the image is m×n. The size of the three-dimensional image tensor F is m×n×k. m, n and k are the scales in the horizontal axis, vertical axis and time axis directions, respectively. In this embodiment, m = 512, n = 640 and k = 10.

[0084] In this embodiment, a sliding window filtering operation is performed on the generated 3D image tensor F to obtain the target saliency image D and the 3D dynamic clutter suppression image G of frame t0. The enhanced target saliency image ETSM consists of the target saliency image D, the 3D dynamic clutter suppression factor G, and the noise discrimination factor weighted map NS. The calculation formula for the enhanced target saliency image ETSM is as follows:

[0085]

[0086] Where (x,y) is the pixel position, D(x,y) is the pixel value of the target saliency image D at (x,y), G(x,y) is the pixel value of the three-dimensional dynamic clutter suppression image G at (x,y), NS(x,y) is the noise discrimination factor of (x,y), and ETSM(x,y) is the pixel value of the enhanced target saliency image ETSM at (x,y).

[0087] In this embodiment, the method for obtaining a target saliency image by performing a sliding window filtering operation on a three-dimensional image tensor includes:

[0088] For each pixel (x, y, t0) in the t0th frame of the 3D image tensor F, the target energy distribution characteristics are extracted using the 3D trapezoidal filter template M, and the target energy distribution gain image is calculated.

[0089] For each pixel (x, y, t0) in the t0th frame of the three-dimensional image tensor F, the temporal fluctuation information of the target neighborhood energy is extracted using the three-dimensional hollow mean filter template H, and the cumulative image of the target motion energy is calculated.

[0090] The target saliency image is calculated based on the target energy distribution gain image and the target motion energy accumulation image.

[0091] Specifically, the target energy distribution gain image is calculated as follows: For each pixel (x, y, t0) in the t0th frame of the 3D image tensor F, a 3D trapezoidal filter template M is used to extract the target energy distribution characteristics. Centered on pixel (x, y, t0), the image is expanded by c1 pixels in both the positive and negative directions of the horizontal axis, c1 pixels in both the positive and negative directions of the vertical axis, and k-1 time points in the negative direction of the time axis, for a total of (2c1+1)×(2c1+1)×k pixel values ​​as estimation samples. In this embodiment, c1 = 3. The pixel value P(x, y) at the (x, y) position of the target energy distribution gain image P is then calculated. The specific calculation method is as follows:

[0092]

[0093] Δf(x,y,t)=f(x,y,t)-f(x,y,t-1)

[0094]

[0095] Where (x,y,t) is the pixel position, f(x,y,t) is the pixel value at (x,y) in the t-th frame image, (i,j,t) is the pixel position, Δf is the inter-frame difference image, and M(i,j,t) is the weight value at pixel (i,j) in the t-th frame of the 3D trapezoidal filter template.

[0096] Specifically, since the central region of infrared small targets has high energy and the surrounding region has low energy, and the surrounding region is easily coupled with the background, leading to the introduction of background noise, a novel three-dimensional trapezoidal distribution template M was designed. Regarding the weighting of M, considering that the energy distribution in the central region of the target is similar in practical applications and tends towards maximum energy, pixels closer to the center pixel in M ​​are assigned the same higher weight, thus efficiently enhancing the target energy. Simultaneously, pixels farther from the center pixel in M ​​are assigned lower weights, and the weight of each pixel decreases linearly with increasing distance from the center point, thereby avoiding the introduction of excessive background information. Multi-frame overlay operations are also considered to weaken the interference of random noise in the background. Therefore, the three-dimensional trapezoidal filter template M designed in this scheme can effectively utilize the energy distribution characteristics of the target to enhance infrared small targets.

[0097] Specifically, the target motion energy accumulation image calculation is as follows: For each pixel (x, y, t0) in the t0th frame of the 3D image tensor F, a 3D hollow mean filter template H is used to extract the temporal fluctuation information of the target's neighborhood energy. Centered on pixel (x, y, t0), the image expands c2 pixels in both the positive and negative directions of the horizontal axis, c2 pixels in both the positive and negative directions of the vertical axis, and k time points in the negative direction of the time axis, resulting in a total of (2c2+1)×(2c2+1)×k pixel values ​​as estimation samples. In this embodiment, c2 is 10. This allows for the calculation of the pixel value E(x, y) at position (x, y) in the target motion energy accumulation image E. The specific calculation method is as follows:

[0098] E(x,y)=Q(x,y,t0)-Q(x,y,t0-k+1)

[0099]

[0100]

[0101] Where (x,y,t) is the pixel position, f(x,y,t) is the pixel value at (x,y) in the t-th frame image, N is the number of pixels in the estimated sample of pixel (x,y,t) that satisfy max(|i|,|j|)>4, (i,j,t) is the pixel position, and H(i,j,t) is the weight value of the t-th frame of the three-dimensional hollow mean filter template H at (i,j).

[0102] In the spatiotemporal domain, the motion of small infrared targets is continuous. When a target moves to its current position, the energy of its previous location changes due to the target's departure. For example, when a target moves from frame t0-k+1 to frame t0, for the central region where the target is located in frame t0, the outer neighborhood of this region only contains the background. Therefore, Q(x,y,t0) is the average energy of the background region. At the same time, the outer neighborhood of this region contains the target in frame t0-k+1. Therefore, Q(x,y,t0-k+1) is the average energy superimposed with the target's energy. So, Q(x,y,t0) < Q(x,y,t0-k+1), and E(x,y) is negative, characterizing the temporal change characteristics of the target's neighborhood energy during the continuous motion of the target.

[0103] Specifically, the target saliency image is calculated as follows: The target saliency image D is obtained by subtracting the target energy distribution gain image P and the target motion energy accumulation image E. The specific calculation formula is as follows:

[0104] D = PE

[0105] Where (x,y) represents the pixel position. P(x,y) is the pixel value of the target energy distribution gain image P at (x,y), E(x,y) is the pixel value of the target motion energy accumulation image E at (x,y), and D(x,y) is the pixel value of the target saliency image D at (x,y).

[0106] Based on the calculation formula of the enhanced target saliency image ETSM(x,y), it can be seen that D(x,y) and the enhanced target saliency image ETSM(x,y) are positively correlated. The more significant the energy distribution characteristics of the infrared small target, the larger the value of P(x,y); the more significant the temporal change characteristics of the energy in its neighborhood during the continuous movement of the infrared small target, the smaller E(x,y). Figure 2 As shown, this method fully utilizes the three-dimensional spatiotemporal energy distribution characteristics of the target during motion and the temporal variation characteristics of the energy in the target's neighborhood, effectively enhancing the target's salience.

[0107] In this embodiment, the method for obtaining a three-dimensional dynamic clutter-suppressed image by performing a sliding window filtering operation on a three-dimensional image tensor includes:

[0108] For each pixel (x, y, t0) in the t0th frame of the 3D image tensor F, using the c3×c3 local gray-scale mean b of the previous k-1 frames at (x, y), r Together they form a time-series grayscale value sequence {b r}, r=1,...,k;

[0109] For the time-series grayscale value sequence {b r The phase space is reconstructed for r = 1,...,k to obtain the gray value sequence B(g);

[0110] Rearrange each grayscale value sequence B(g) in ascending order, and form a symbol sequence S(g) based on the column position index of each grayscale value;

[0111] Calculate the weighted probability P of each symbol sequence. g ;

[0112] Based on weighted probability P g Calculate the three-dimensional dynamic clutter suppression image G(x,y).

[0113] Specifically, the calculation of 3D dynamic clutter suppression image is as follows: For each pixel (x,y,t0) in the t0th frame of the 3D image tensor F, the mean gray value b of the c3×c3 local region at (x,y) in the previous k-1 frames is used. r Together they form a time-series grayscale value sequence {b r}, r=1,...,k, in this embodiment c3 is 9, calculate its three-dimensional dynamic clutter suppression image G(x,y), the specific calculation steps are as follows:

[0114] B(g)=[b r (g),b r (g+T),...,b r [g+(D-1)T)], g=1,...,N

[0115]

[0116] Where B(g) is {b r The g-th gray value sequence obtained after phase space reconstruction, where D is the embedding dimension, T is the hysteresis time (5 and 1 in this embodiment), N = k - (D - 1)T is the total number of reconstructed sequences, and M... s This represents the total number of pixels in the local spatial region at (x,y).

[0117] Rearrange each B(g) in ascending order, and form a symbol sequence S(g) from the column position indices of each grayscale value. There are a total of D! different sorting structures, and the expression for S{i} is as follows:

[0118] S(g) = [j1,...j l ...j D ], l≤D, g≤D!

[0119] Where j l To reconstruct the column position index of the l-th element in the sequence S(g);

[0120] Gray value sequence {b r The three-dimensional dynamic clutter suppression image G(x,y) is composed of the weighted probability P of each symbol sequence. i The specific formula for calculating G(x,y) is as follows:

[0121]

[0122] P g =w g *p g

[0123]

[0124] Where p g Let w be the probability of the g-th symbol sequence appearing, w be the weight of the g-th symbol sequence, and η be the threshold, which is 1 in this embodiment. For flat regions such as deep space or inside clouds in the background, the gray value of the sequence B(g) fluctuates within a small range, making G(x,y) more sensitive to this fluctuation, which can easily interfere with the background suppression performance. Therefore, a threshold condition is set to avoid interference from this flat background region.

[0125] Based on the calculation formula of the enhanced target saliency image ETSM(x,y), it can be seen that G(x,y) and the enhanced target saliency image ETSM(x,y) are negatively correlated. The residual background noise in multi-frame images is mainly caused by moving background clutter, such as... Figure 3 As shown, there are swaying trees in the scene. When calculating G(x,y) for the constantly changing complex background clutter in the image, the time series information in the original gray-level sequence is preserved, reflecting the characteristics of the gray-level sequence in terms of structural distribution. For areas such as swaying trees in the scene, the more irregular the temporal gray-level value changes, the larger G(x,y) is, and the target saliency image ETSM(x,y) decreases. Therefore, the complex and ever-changing background clutter can be effectively suppressed.

[0126] In this embodiment, the method for performing pixel-level highlight noise discrimination on a three-dimensional image tensor to obtain a noise discrimination factor-weighted image includes:

[0127] Pixel-level high-brightness noise suppression strategy: For each pixel (x, y, t0) in the t0th frame of the 3D image tensor F, a noise discrimination feature d is designed based on the gray value of each pixel and the gray value distribution of its eight neighboring pixels to judge noise at that point, and pixels judged as pixel-level high-brightness noise are suppressed. The specific calculation formula is as follows:

[0128]

[0129]

[0130] Where f(x,y,t0) is the grayscale value of the image at position (x,y) in frame t0, N is the eight neighboring pixels of pixel (x,y), β is the noise discrimination threshold, which is 25 in this embodiment. NS(x,y) is the weighted value of the noise discrimination factor at (x,y).

[0131] In infrared images, such as Figure 4 As shown, pixel-level high-brightness noise in complex backgrounds can lead to false alarms in detection results. Because the grayscale value of pixel-level high-brightness noise exhibits abrupt changes in its surrounding neighborhood, while the grayscale value of the target gradually decreases from the center outwards, when the maximum difference in local grayscale values ​​exceeds the noise discrimination threshold β, the pixel more closely satisfies the abrupt grayscale change characteristic of the noise, resulting in a lower value in the image where the target is salient.

[0132] In this embodiment, the method for segmenting the enhanced target saliency image using an adaptive threshold includes: the object to be segmented is the enhanced target saliency image ETSM of frame t0. Pixels in the enhanced target saliency image ETSM with values ​​greater than a threshold Th are considered suspected targets. Th is calculated as follows:

[0133] Th=μ+q·σ

[0134] Where μ and σ are the mean and standard deviation of the enhanced target saliency image, respectively, and q is a manually set value, which is 3 in this embodiment.

[0135] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A spatiotemporal domain detection method for faint infrared targets based on three-dimensional tensor filtering, characterized in that, Includes the following steps: Constructing a three-dimensional image tensor for infrared sequence images against a complex low-altitude background; A sliding window filtering operation is performed on the three-dimensional image tensor to obtain a target saliency image and a three-dimensional dynamic clutter suppression image; Pixel-level highlight noise discrimination is performed on the three-dimensional image tensor to obtain a noise discrimination factor weighted image; An enhanced target saliency image is obtained based on the target saliency image, the three-dimensional dynamic clutter suppression image, and the noise discrimination factor weighted image. By segmenting the enhanced target saliency image using an adaptive threshold, the real target is extracted, thus completing the spatiotemporal detection of small, faint infrared targets. The method for obtaining a target saliency image by performing a sliding window filtering operation on the three-dimensional image tensor includes: For 3D image tensors The first in Each pixel in the frame A three-dimensional trapezoidal filter template is used. Extract the target energy distribution characteristics and calculate the target energy distribution gain image; For 3D image tensors The first in Each pixel in the frame A three-dimensional hollow mean filter template is used. Extract the temporal fluctuation information of the target's neighborhood energy and calculate the target's motion energy accumulation image; Based on the target energy distribution gain image and the target motion energy accumulation image, calculate the target saliency image; The methods for calculating the target energy distribution gain image include: ， ， ， In the formula, For pixel position, For the first Frame image in Pixel value at that location, For pixel position, For inter-frame difference images, The third trapezoidal filter template Frame in Weight value at pixel Gain image of target energy distribution exist Pixel value at the location, The number of pixels to be expanded; Methods for calculating the cumulative energy image of a target motion include: ， ， ， In the formula, For pixel position, For the first Frame Image Pixel value at that location, For pixels In the estimated sample, satisfying The number of pixels, For pixel position, Three-dimensional hollow mean filter template The Frame in The weight value at that location, Accumulated energy image of target motion exist Pixel value at the location, To expand the number of pixels, The average energy value of the background region. The average energy value superimposed with the target energy; Methods for calculating target saliency images include: ， In the formula, For pixel position, Gain image of target energy distribution exist Pixel value at that location, Accumulated energy image of target motion exist Pixel value at that location, For target saliency image exist The pixel value at that location.

2. The spatiotemporal domain detection method for faint infrared targets based on three-dimensional tensor filtering according to claim 1, characterized in that, The method for constructing the three-dimensional image tensor includes: Get the current number The original single-frame image of the frame ,in, ≥2; Compare the current frame image with its neighbors Frame images are stacked pixel by pixel along the time dimension to form a three-dimensional image tensor. ; Among them, the three-dimensional image tensor The calculation formula is: , In the formula, For the first Frame image, the size of the image is 3D image tensor The size is , , and These refer to the scale along the horizontal axis of space, the vertical axis of space, and the time axis, respectively.

3. The spatiotemporal domain detection method for faint infrared targets based on three-dimensional tensor filtering according to claim 1, characterized in that, The method for obtaining a three-dimensional dynamic clutter-suppressed image by performing a sliding window filtering operation on the three-dimensional image tensor includes: For 3D image tensors The first in Each pixel in the frame , before Frame image in place Local area gray mean Together they form a time-series grayscale value sequence ; For time-series grayscale value sequences Phase space reconstruction is performed to obtain the gray value sequence. ; Each grayscale value sequence Reordered in ascending order, the column position indices of each grayscale value form a symbol sequence. ; Calculate the weighted probability of each symbol sequence occurring. ; Based on the weighted odds Calculate the three-dimensional dynamic clutter suppression image ; Among them, the gray value sequence is calculated. The expression is: In the formula, yes The first phase space obtained after phase space reconstruction A sequence of grayscale values, For the embedding dimension, For the delay time, The total number of sequences obtained from reconstruction. for The total number of pixels in a local spatial region; Calculate symbol sequence The expression is: ； In the formula, For reconstructing the sequence The Middle The column position index of each element. The total number of different sorting structures; Calculate the three-dimensional dynamic clutter suppression image The expression is: ， In the formula, For the first The probability of a sequence of symbols appearing. For the first The weights of a sequence of symbols, The threshold value is used.

4. The spatiotemporal domain detection method for faint infrared targets based on three-dimensional tensor filtering according to claim 1, characterized in that, The method for performing pixel-level highlight noise discrimination on the three-dimensional image tensor to obtain a noise discrimination factor-weighted image includes: For 3D image tensors The first in Each pixel in the frame Based on the gray value of each pixel and the gray value distribution of its eight neighboring pixels, noise discrimination features are obtained. ; Based on the noise discrimination features For the corresponding pixel Noise is assessed, and pixels identified as pixel-level high-brightness noise are suppressed to obtain a noise discrimination factor weighted image. Among them, noise discrimination features are obtained. The expression is: , The expression for obtaining the noise discrimination factor-weighted image is: , In the formula, For the first Frame image in The grayscale value of the location, For pixels The eight neighboring pixels, For noise discrimination threshold, for The weighted value of the noise discrimination factor.

5. The spatiotemporal domain detection method for faint infrared targets based on three-dimensional tensor filtering according to claim 1, characterized in that, The method for obtaining an enhanced target saliency image based on the target saliency image, the three-dimensional dynamic clutter suppression image, and the noise discrimination factor weighted image includes: ， in, For pixel position, For target saliency image exist Pixel value at that location, Three-dimensional dynamic clutter suppression image exist Pixel value at that location, Weighted graph of noise discrimination factor exist Noise discrimination factor, To enhance the saliency of the target image exist The pixel value at that location.

6. The spatiotemporal domain detection method for faint infrared targets based on three-dimensional tensor filtering according to claim 1, characterized in that, The method for segmenting the enhanced target saliency image using an adaptive threshold includes: Enhanced target saliency image Medium greater than the threshold The pixels are considered to be the real target; The calculation method is as follows: in, and Here, represents the mean and standard deviation of the enhanced target salient image, respectively. Values ​​are set by humans.

Images