An infrared small target detection method and system based on an adaptive trade-off structure tensor

Abstract

The application provides an infrared small target detection method and system based on an adaptive trade-off structure tensor, and relates to the technical field of infrared image target detection, wherein the method comprises the following steps: obtaining a to-be-processed infrared image containing a to-be-detected target, and extracting a corresponding structure tensor after pre-processing the to-be-processed infrared image; constructing an improved edge perceiver and an angle point perception factor according to the structure tensor, and weighting the edge interference and the angle point by using a structure similarity contrast metric mode of the improved edge perceiver and the angle point perception factor, so that the recognizability of the target is better increased; simple threshold segmentation is performed on the weighted result, and an accurate detection result can be obtained; the application significantly improves the target detection capability while suppressing the background clutter, maintains the saliency of the target, and has a high detection rate and a low false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of infrared image target detection technology, and more specifically to an infrared small target detection method and system based on adaptive trade-off structure tensor. Background Technology

[0002] Currently, infrared small target detection has become a research hotspot in recent years, with broad application prospects in remote sensing, early warning devices, and search systems. Compared to ordinary visible light, infrared imaging has a greater advantage in penetration. And compared to radar imaging, infrared imaging is more convenient and faster. However, its two characteristics of being "weak" and "small" also make it a difficult area for research. "Weak" refers to the low signal-to-noise ratio of the target, poor contrast with the background, and weak infrared radiation intensity; while "small" refers to the small number of target pixels, making it difficult to obtain texture information during detection, with only grayscale and position information available. Background clutter interference is a major challenge in detection; therefore, the research on infrared small target detection against complex backgrounds has been gaining momentum in recent years.

[0003] However, existing infrared small target detection methods are not stable enough when faced with complex backgrounds containing a lot of clutter and clear high-contrast edges. The significant background edges also have sparse features that are difficult to suppress, and the threshold is usually set to a fixed value, which makes it difficult to flexibly separate the target from the complex background.

[0004] Therefore, how to provide an infrared small target detection method that can solve the above problems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides an infrared small target detection method and system based on adaptive trade-off structure tensor, which significantly improves the target detection capability while suppressing background clutter, maintains the salience of the target, and has a high detection rate and a low false alarm rate.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An infrared small target detection method based on adaptive tradeoff structure tensor includes the following steps:

[0008] Acquire an infrared image containing the target to be tested, and extract the corresponding structure tensor after preprocessing the infrared image;

[0009] An improved edge perceptron and a corner sensing factor are constructed based on the structure tensor. An adaptive edge suppression and target point enhancement model is constructed based on the improved edge perceptron and the corner sensing factor. The adaptive edge suppression and target point enhancement model is used to process the infrared image to be processed, so as to simultaneously enhance the target by adaptively constraining the background interference components according to different scenes.

[0010] The structural similarity metric is used to measure the weighting of background clutter edge components and target components in the processed infrared image to be processed;

[0011] Threshold segmentation is applied to the weighed results to obtain the detection results.

[0012] Preferred options also include:

[0013] Calculate adjustable adaptive parameters based on structural similarity metrics, and use the adjustable adaptive parameters to balance the weights of the improved edge perceptron and the corner perception factor.

[0014] Preferred options also include:

[0015] The adjustable adaptive parameters are designed to determine the optimal measurement method.

[0016] Preferably, the specific process of constructing an improved edge perceptron includes:

[0017] The first and second feature values ​​of the infrared image to be processed are obtained through the structure tensor.

[0018] An edge-aware index is constructed based on the first feature value and the second feature value;

[0019] By introducing a nonlinear function to improve the edge perception index, the final improved edge perceptron is obtained.

[0020] Preferably, the specific process for constructing corner perception factors includes:

[0021] A corner perception factor is constructed based on the first feature value and the second feature value.

[0022] Preferably, the specific process for determining the optimal adjustable adaptive parameters includes:

[0023] Determine the value range and step size of the adjustable adaptive parameter, and determine the number of adjustable adaptive parameters to be tested based on the step size;

[0024] A structural similarity measurement index is constructed, and each of the adjustable adaptive parameter input values ​​within the range is calculated in the structural similarity measurement index. The adjustable adaptive parameter with the smallest calculation result is selected as the optimal adjustable adaptive parameter.

[0025] This invention also provides an infrared small target detection system based on an adaptive tradeoff structure tensor, comprising:

[0026] The acquisition module is used to acquire an infrared image containing the target to be tested, and to extract the corresponding structure tensor after preprocessing the infrared image.

[0027] The first construction module is used to construct an improved edge perceptron and a corner sensing factor based on the structure tensor, construct an adaptive edge suppression and target point enhancement model based on the improved edge perceptron and the corner sensing factor, and process the infrared image to be processed using the adaptive edge suppression and target point enhancement model.

[0028] The second construction module uses a structural similarity metric to measure the weighted distribution of background clutter edge components and target components in the processed infrared image.

[0029] The detection module is used to perform threshold segmentation on the weighed results to obtain the detection results.

[0030] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an infrared small target detection method and system based on adaptive tradeoff structure tensor, which has the following beneficial effects:

[0031] (1) An adaptive weighted infrared small target detection method based on structure tensor is proposed, which can use structure tensor alone to characterize the target and background edges, and can accurately detect weak targets with high efficiency in complex backgrounds.

[0032] (2) An improved edge perceptron (EO) is proposed, which can effectively target various clutter edges in the background and amplify the differences between them and the target. This enhancement helps the subsequent separation process.

[0033] (3) An adaptive parameter α with an important factor that can be adjusted is set, which can adaptively select the value according to different types of background clutter, thereby enhancing the matching degree of the algorithm for edge and high-brightness clutter as well as corner clutter.

[0034] The invention was subsequently tested on six infrared sequences with different backgrounds. The results showed that the method provided by this invention consistently outperformed other methods in SCRG and BSF scores, especially in densely cloud-occupied scenarios. Experimental results demonstrate that the method provided by this invention significantly improves target detection capability while suppressing background clutter, maintaining target saliency, and exhibiting high detection rate and low false alarm rate. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0036] Figure 1 This invention provides an overall flowchart of an infrared small target detection method based on an adaptive trade-off structure tensor.

[0037] Figure 2 The present invention provides a comparison of the cases where α is 0 and the case where α is adaptively selected. The red box represents the target, the blue box represents clutter, (a1)-(a2) are the original images, (b1)-(b2) show the target and clutter effects when α is 0, and (c1)-(c2) show the processing results when α is adaptively selected.

[0038] Figure 3 This invention provides a structural principle block diagram of an infrared small target detection system based on an adaptive trade-off structure tensor.

[0039] Figure 4 This is a schematic diagram of the detection window and a schematic diagram of the background inhibition factor results provided in an embodiment of the present invention. Detailed Implementation

[0040] 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.

[0041] See Figure 1 As shown, this invention discloses an infrared small target detection method based on an adaptive tradeoff structure tensor, comprising the following steps:

[0042] Acquire an infrared image containing the target to be tested, and extract the corresponding structure tensor after preprocessing the infrared image. The preprocessing process can be Gaussian filtering of the infrared image.

[0043] An improved edge perceptron and corner sensing factor are constructed based on the structure tensor. An adaptive edge suppression and target point enhancement model is then constructed based on the improved edge perceptron and corner sensing factor. The adaptive edge suppression and target point enhancement model is then used to process the infrared image to be processed, so as to enhance the target while adaptively constraining the background interference components according to different scenarios.

[0044] The structural similarity metric is used to weigh the background clutter edge components and target components in the processed infrared image.

[0045] Threshold segmentation is applied to the weighed results to obtain the detection results.

[0046] In one specific embodiment, it also includes:

[0047] Calculate adjustable adaptive parameters based on structural similarity metrics, and use the adjustable adaptive parameters to balance and improve the weights of the edge perceptron and corner perception factors.

[0048] In one specific embodiment, the process of constructing an improved edge sensor includes:

[0049] The first and second eigenvalues ​​of the infrared image to be processed are obtained using the structure tensor.

[0050] An edge perception index is constructed based on the first and second feature values;

[0051] By introducing a nonlinear function to improve the edge perception index, the final improved edge perceptron is obtained.

[0052] Specifically, the concrete expression for the structure tensor is:

[0053]

[0054] In the formula, G β Represents the Gaussian kernel, used to enhance the robustness of geometric orientation estimation, with horizontal and vertical gradients respectively; J β Let f represent a symmetric positive semi-definite matrix with two eigenvalues ​​λ1 and λ2 (i.e., the first eigenvalue λ1 and the second eigenvalue λ2). σ This indicates that the infrared image to be processed is subjected to Gaussian smoothing filtering with variance σ, f σx f σy Let J represent the gradients of the Gaussian filter in the x and y directions, respectively. 11 J 12 J 21 J 22 Represents the elements of a matrix.

[0055] In the calculation of the structure tensor, the first eigenvalue λ1 and the second eigenvalue λ2 contain the structural information of the local region centered on the pixel. The first eigenvalue λ1 and the second eigenvalue λ2 show great feature differences in different regions of the infrared image. In flat regions, λ1≈λ2≈0, in edge regions such as clouds, λ1>>λ2≈0, and in corner regions, λ1≥λ2>>0.

[0056] Since pixel values ​​at edges are much larger than those in flat and corner areas, an edge-aware index (EI) is constructed. This index is defined as the difference between feature values ​​to ensure that the grayscale values ​​obtained in edge regions are significantly larger than those in the other two regions. The specific expression is as follows:

[0057] EI(x,y)=λ1-λ2

[0058] Experiments on edge perception metrics show that background edges processed by the edge perceptron (EI) mostly exhibit complete and continuous features, making it suitable for situations where the target and background edges do not overlap.

[0059] However, a major challenge in weak infrared target detection is separating the complex background from the target. Complex backgrounds often contain a large amount of background clutter. When numerous clutter edges appear, using an edge-aware index (EI) to highlight them all is somewhat redundant. Furthermore, these numerous edges are filtered out in the next step of highlighting small targets. In this case, detailed and complete background edges increase the complexity of the detection process. For images with many low-contrast edges and targets surrounding those edges, using an edge-aware index (EI) actually weakens the target, a corner-like region, resulting in reduced contrast between the target and its surrounding area, making it difficult to extract.

[0060] Therefore, in order to solve the above problems, the embodiments of the present invention introduce a nonlinear function e x The edge perception metric is improved to obtain the final improved edge perceptron, the specific expression of which is:

[0061]

[0062] nonlinear function e x It has an important property that e increases as the independent variable x increases. x The growth rate of λ1-λ2 will also accelerate, meaning that for larger λ1-λ2 values ​​(i.e., high-contrast edges), the improved edge perceptron will give a larger value, thus emphasizing these edges more strongly. The nonlinear function exhibits an "explosive growth" pattern, so for smaller λ1-λ2 values ​​(i.e., low-contrast edges), the grayscale value will not be significantly enhanced. The low-contrast edges processed by this formula can be weakened in the image, and the final image retains the target and strong edges that are difficult to remove with high contrast, reducing the workload for subsequent data processing. Furthermore, in some scenes, the low-contrast edges around the target in the processed image will also be reduced.

[0063] In a specific embodiment, the process of constructing the corner sensing factor includes:

[0064] Corner sensing factors are constructed based on the first eigenvalue λ1 and the second eigenvalue λ2.

[0065] Specifically, in weak infrared target detection, edge regions and corner regions coexist in the image. This characteristic is also reflected in the target, where the target region simultaneously contains features of both edges and corners. Simply using an improved edge perceptron can only effectively process images with clear background edges, but it cannot identify weak targets in images with numerous corners and blurred edges with low contrast. Therefore, this embodiment of the invention constructs a corner-perceptron indicator, the specific expression of which is:

[0066]

[0067] In this embodiment of the invention, the corner perception index constructed using the concept of harmonic mean can suppress outliers such as high-brightness noise points (PNHB) in the background by taking the reciprocal, and can highlight targets with low contrast.

[0068] The target region processed using the improved edge perceptron loses corner features. When the target occupies a large number of pixels in the image, only an outer edge remains; conversely, when the target occupies a small proportion of pixels, it may be lost, thus reducing the detection rate. Different proportions of edge and corner sensing should be used for infrared weak target images in different scenarios. For cases with relatively flat background clutter and a small proportion of target pixels, using a corner sensing indicator can achieve a higher detection rate and faster detection speed.

[0069] In one specific embodiment, it also includes:

[0070] Design the adjustable adaptive parameter α and determine the optimal measurement method.

[0071] In one specific embodiment, the process of determining the optimal adjustable adaptive parameter α includes:

[0072] Determine the range of values ​​and step size of the adjustable adaptive parameters, and determine the number of adjustable adaptive parameters to be tested based on the step size;

[0073] A structural similarity measurement index is constructed, and the structural similarity measurement index is calculated for each adjustable adaptive parameter input value within the range of values. The adjustable adaptive parameter with the smallest calculation result is selected as the optimal adjustable adaptive parameter.

[0074] In one specific embodiment, the process of determining the optimal adjustable adaptive parameters includes:

[0075] Determine the range of values ​​and step size of the adjustable adaptive parameters, and determine the number of adjustable adaptive parameters to be tested based on the step size;

[0076] A structural similarity measurement index is constructed, and the structural similarity measurement index is calculated for each adjustable adaptive parameter input value within the range of values. The adjustable adaptive parameter with the smallest calculation result is selected as the optimal adjustable adaptive parameter.

[0077] Specifically, this embodiment of the invention proposes an adjustable adaptive parameter α to control the weights of the improved edge perceiver and corner perception indicator, and the corresponding specific expression is as follows:

[0078] ATST(x,y)=α*EO+(1-α)*CI

[0079] In the formula, CI represents the corner sensing indicator, EO represents the improved edge sensing instrument, and α is the adaptive weight value, which ranges from [0,1]. The implementation process is shown in Table 1.

[0080] Table 1 shows the implementation process of α.

[0081] Input: Original image Output: ATST format <![CDATA[1. Calculate the image feature values λ1 and λ2 according to Formulas (3) and (4)]]> 2. Calculate the improved edge perceptron value EO according to formula (4). 3. Calculate the corner sensing indicator value CI according to formula (5). 4. Input the values ​​obtained in 1-3 into formula (6) to obtain the ATST formula for the undetermined coefficient α. 5. Input the ATST formula of the undetermined coefficient α into Algorithm 2.

[0082] The adaptive parameter α is selected by iterating through 100 values ​​in the range of 0-1 and choosing the α value that yields the best processing effect. The best effect is achieved by substituting the currently selected α value into the structural similarity index measure (SSIM) method proposed in this embodiment of the invention, and then determining which indicator and required weight the image is more suitable for based on the processing result, ultimately deciding on the optimal α value for the current image.

[0083] The structural similarity index measure (SSIM) is typically expressed as three factors: correlation loss, brightness distortion, and contrast distortion of an image. The specific expression is as follows:

[0084] SSIM(f,g)=l(f,g)c(f,g)s(f,g)

[0085]

[0086] In the formula, the first term of l(f,g) is the brightness comparison function, used to measure the average brightness (μ) of the two images. f and μ g The degree of closeness of the factor to μ is only determined by the degree of closeness of the factor to μ. f =μ gThe first term is the maximum and equal to 1; the second term, c(f,g), is the contrast comparison function, used to measure the closeness of the contrast between two images; the contrast is based on the standard deviation σ. f and σ g It is measured by σ. Only when σ f =σ g This term is only maximum and equal to 1 when the condition is met. The third term, s(f,g), is the structural comparison function, used to represent the correlation coefficient between two images f and g, where σ fg Let f be the covariance between f and g. The positive values ​​of the SSIM exponent are in [0,1], where 0 indicates that there is no correlation between the images and 1 indicates that f = g. C1, C2 and C3 are constants.

[0087] Using structural similarity as a metric, the weight value that minimizes the structural similarity between the current image and the original image is the value we need. The evaluation process is shown in Table 2.

[0088] Table 2 shows the specific implementation process for evaluating α.

[0089]

[0090] This invention presents two scenarios for comparison: α is 0 and α is adaptively selected based on the algorithm. Images from dataset e are used in this comparison. Figure 2 As can be seen from (b1)-(b2), when the weight α is 0, background clutter (the area within the blue box) still exists in large quantities in the image, and the clutter portion remains significant when observing its 3D image. According to... Figure 2 In (c1)-(c2), it can be observed that the background clutter is well removed. Although the target range is reduced compared to (b1)-(b2), it can be observed from (c1)-(c2) that the salience of the target is not weakened and it can still be accurately distinguished from the background, and the clutter can be removed.

[0091] See Figure 3 As shown, the present invention also provides an infrared small target detection system based on an adaptive tradeoff structure tensor, comprising:

[0092] The acquisition module is used to acquire the infrared image to be processed containing the target to be tested, and extract the corresponding structure tensor after preprocessing the infrared image to be processed.

[0093] The first construction module is used to construct an improved edge perceptron and corner sensing factor based on the structure tensor, and to construct an adaptive edge suppression and target point enhancement model based on the improved edge perceptron and corner sensing factor, and to process the infrared image to be processed using the adaptive edge suppression and target point enhancement model.

[0094] The second construction module is based on the weighted average of background clutter edge components and target components in the processed infrared image using structural similarity measurement.

[0095] The detection module is used to perform threshold segmentation on the weighed results to obtain the detection results.

[0096] To evaluate the performance of the method provided in this embodiment of the invention, four evaluation metrics widely used in the field of infrared small target detection are used: PD, FA, SCR, and BSF. Among them, the most important metrics for evaluating the target detection success rate are PD and FA, and their definitions are as follows:

[0097]

[0098] The curve plotted with Fa as the x-axis and Pd as the y-axis is called the ROC curve. The area under the curve (AUC) can quantify the detection performance. The larger the area under the ROC curve, the better the detection performance. The detection rate (Pd) reflects the detection capability of the algorithm, while the false alarm rate (Fa) represents the false detection rate of the algorithm.

[0099] The signal-to-clutter ratio (SCR) reflects the difficulty of detecting weak targets, while the signal-to-clutter gain (SCRG) evaluates the algorithm's ability to enhance the signal and describes the significance of the target enhancement after processing. The formulas used are shown below:

[0100]

[0101] In the formula, μ t μ is the average pixel value of the target. b and σ b These represent the average pixel value and standard deviation of the pixels in the surrounding neighborhood of the target, respectively. A detected target is considered a real target if the following two conditions are met: 1) The detected target overlaps with the real target in the image. 2) The center pixel position of the detected target differs from the center pixel position of the real target by no more than 5 pixels. A schematic diagram of the method used for target detection within the target area is shown below. Figure 4 As shown.

[0102] Background suppression factor (BSF) is used to represent the background suppression ability of the algorithm after processing, and its definition is as follows:

[0103]

[0104] In the formula, σ in and σ outThese represent the background standard deviation of the input image and the background standard deviation of the output image, respectively. The BSF value reflects the algorithm's ability to suppress the background; the larger the value, the stronger the suppression ability. Similarly, the larger the SCRG value, the more effective the algorithm is in enhancing the target.

[0105] This invention uses six infrared small target detection sequences under different backgrounds for comparative evaluation. Sequence a has a dark background with low contrast, and the small targets are hidden in dense grass with some overlap. Sequence b has many striped cloud layers in the background, and some small targets are hidden in the clouds. Sequence c has a high background contrast, and the clouds are dense and undulating. Sequence d has a sparse background, but the contrast between the small targets and the surrounding area is very low, and there are large dark areas in the background. Sequence e has clouds with a similar gray level to the small targets, which can easily lead to false detections. Sequence f has a lot of noise in the background, and the noise intensity is close to that of the small targets. Specific experimental conditions are shown in Table 3.

[0106] Table 3 Infrared sequence characteristics

[0107]

[0108]

[0109] This invention evaluated the effectiveness of the algorithm and compared it with four different baseline methods. Four different types of infrared small target detection algorithms were selected for comparative experiments: LR, LEF, MPCM, and TLLCM. Specific results are shown in Tables 4 and 5. Table 4 shows that the method proposed in this invention achieves the highest SCRG score in five scenarios, significantly exceeding the scores of other baseline methods in some scenarios. In sequence (c), the small target has low brightness and poor imaging; the method provided by this invention scores slightly lower than local contrast-based methods. However, in complex, occluded backgrounds, the method proposed in this invention can significantly improve the intensity of small targets. Table 5 shows that the method of this invention performs comparably to most algorithms in some scenarios and significantly outperforms other algorithms in half of the scenarios.

[0110] Table 4 Comparison of SCRG scores using different methods

[0111]

[0112] Table 5 compares BSF scores using different methods. The best scores are marked in bold.

[0113]

[0114] In summary, the method proposed in this embodiment of the invention can be adapted to most scenarios, especially in scenarios with dense cloud cover, where the SCRG and BSF scores are higher than those of the comparison algorithm, and can suppress complex backgrounds while enhancing the salience of the target.

[0115] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0116] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting small infrared targets based on adaptive tradeoff structure tensors, characterized in that, Includes the following steps: Acquire an infrared image containing the target to be tested, and extract the corresponding structure tensor after preprocessing the infrared image; Based on the aforementioned structural tensor, an improved edge perceptron and corner sensing factor are constructed. The specific process includes: The first and second feature values ​​of the infrared image to be processed are obtained through the structure tensor. An edge-aware index is constructed based on the first feature value and the second feature value; A nonlinear function is introduced to improve the edge perception index, resulting in the final improved edge perceptron, the specific expression of which is: In the formula, Represents the first eigenvalue. Indicates the second eigenvalue; A corner perception factor is constructed based on the first feature value and the second feature value, and its specific expression is as follows: An adaptive edge suppression and target point enhancement model is constructed based on the improved edge perceptron and corner sensing factor, and the adaptive edge suppression and target point enhancement model is used to process the infrared image to be processed. The structural similarity measurement index is used to measure the weighting of background clutter edge components and target components in the processed infrared image to be processed. The structural similarity measurement index includes three factors: image correlation loss, brightness distortion, and contrast distortion. The results of the trade-offs are then segmented using a threshold to obtain the detection results; Also includes: Calculate adjustable adaptive parameters based on the structural similarity measurement index, and use the adjustable adaptive parameters to balance the weights of the improved edge perceptron and the corner sensing factor; The specific process for determining the optimal adjustable adaptive parameters includes: Determine the value range and step size of the adjustable adaptive parameter, and determine the number of adjustable adaptive parameters to be tested based on the step size; The adjustable adaptive parameter is then calculated in the structural similarity measurement index for each input value within its range, and the adjustable adaptive parameter with the smallest calculation result is selected as the optimal adjustable adaptive parameter.

2. A system utilizing the infrared small target detection method based on adaptive tradeoff structure tensor as described in claim 1, characterized in that, include: The acquisition module is used to acquire an infrared image containing the target to be tested, and to extract the corresponding structure tensor after preprocessing the infrared image. The first construction module is used to construct an improved edge perceptron and a corner sensing factor based on the structure tensor, construct an adaptive edge suppression and target point enhancement model based on the improved edge perceptron and the corner sensing factor, and process the infrared image to be processed using the adaptive edge suppression and target point enhancement model. The second construction module uses a structural similarity measurement index to measure the weighting of background clutter edge components and target components included in the processed infrared image to be processed. The detection module is used to perform threshold segmentation on the weighed results to obtain the detection results.

Images