A method for infrared small target detection supervision enhancement

Abstract

The application relates to an infrared small target detection supervision enhancement method and belongs to the technical field of target detection supervision enhancement; the method comprises the following steps: using an infrared small target detection model to perform small target detection on an infrared image to be detected to generate a prediction graph; the prediction graph and a true value graph are supervised through scale self-adaptation based on a Gaussian mixture model and a double-coordinate position-sensitive loss function to enhance the detection performance of the infrared small target detection model; and the preprocessed image to be detected is input into the enhanced infrared small target detection model to perform infrared small target detection; the scale self-adaptation based on the Gaussian mixture model and the double-coordinate position-sensitive loss are constructed, the sensitivity of the model to the target scale and position is enhanced, and the detection performance of the infrared small target detection (IRSTD) is enhanced.

Description

TECHNICAL FIELD

[0001] The application relates to an infrared small target detection supervision enhancement method and belongs to the technical field of target detection supervision enhancement. BACKGROUND

[0002] Infrared small target detection (IRSTD) aims to segment weak small targets from complex infrared image backgrounds. It is a crucial task in computer vision and has attracted extensive attention in the past few decades and has been widely applied in military and civilian fields. Existing IRSTD methods can be generally divided into two categories: traditional methods developed in the early stage, such as filter-based methods, local contrast-based methods, and low-rank-based methods; and data-driven deep learning methods developed in recent years, including convolutional neural network (CNN)-based methods and Transformer-based methods, which utilize model design to better adapt to challenging scenarios.

[0003] In recent years, with the development of deep learning, especially convolutional neural network (CNN), end-to-end networks such as convolutional neural network are used to directly learn features from data without complex hand-crafted feature design. Although these deep learning methods have achieved satisfactory performance in most scenarios using different deep learning frameworks, it is still crucial to guide the model prediction through a loss function. Recent deep learning methods usually use IoU loss and Dice loss to learn useful features, which guide the model to focus on the difference between the predicted target and the ground truth. However, when the predicted target and the ground truth differ in scale and location, both IoU and Dice loss produce the same value. This insensitivity to scale and location limits the model's ability to distinguish targets of different scales and locations, ultimately limiting detection performance.

[0004] To address the insensitivity of IoU and Dice loss to target scale and location, MSHNet proposes a scale and location sensitivity (SLS) loss. When the predicted target becomes smaller or larger, the IoU loss and the Dice loss have the same value, while the SLS loss changes accordingly. However, as the scale of the predicted target approaches the ground truth, the loss value should decrease accordingly. However, the SLS loss instead shows an unexpected increasing trend. This failure of sensitivity can lead to inaccurate predictions by the model, causing false positives and missed detections. Therefore, accurately identifying different sizes of infrared small targets in complex backgrounds and reducing false positives and missed detections remains a challenging task that has not been thoroughly explored. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application proposes a supervised enhancement method of a GMM-based scale adaptive and Dual-coordinate location sensitive (GD) loss function to overcome the limitations of the existing loss function.

[0006] In the present application, the GD loss function uses a Gaussian mixture model to explicitly model the difference between the predicted target scale and the true scale, and accordingly assigns adaptive weights to the IoU loss. In addition, the present application combines the Euclidean distance with the angle difference in polar coordinates through a dual-coordinate mechanism to enhance the location sensitivity. The GD loss function enhances the scale and location sensitivity of the model, thereby reducing false positives and false negatives. Finally, the present application uses four mainstream IRSTD methods to increase the GD loss function of the present application for model supervised enhancement, thereby completely detecting small targets in infrared images on three benchmark datasets (IRSTD-1K, NUAA-SIRST, and NUDT-SIRST).

[0007] The technical solutions of the present application are as follows:

[0008] An infrared small target detection supervised enhancement method, comprising:

[0009] using an infrared small target detection model to perform small target detection on the infrared image to be detected to generate a prediction map;

[0010] wherein the prediction map and the ground truth map are supervised by a loss function based on scale adaptation and dual-coordinate location sensitivity of a Gaussian mixture model to enhance the detection performance of the infrared small target detection model;

[0011] The image to be detected is preprocessed and input into the enhanced infrared small target detection model for infrared small target detection.

[0012] As a further preferred scheme, using an infrared small target detection model to perform small target detection on the infrared image to be detected to generate a prediction map comprises:

[0013] Given an input infrared image , using an infrared small target detection model to obtain a prediction map :

[0014] ;

[0015] wherein, IRSTD refers to an infrared small target detection model.

[0016] As a further preferred solution, the infrared small target detection model IRSTD is DANA Net, ISNet, PBT or MSHNet.

[0017] As a further preferred solution, in the scale adaptive loss based on the Gaussian mixture model, the target scale adaptation is modeled by the Gaussian mixture model, including:

[0018] Two Gaussian models are fused to focus on multi-scale targets and adaptively adjust the loss result;

[0019] One of the Gaussian models has a more dramatic change when the predicted target scale is smaller than the true value target scale, and has high sensitivity to small scale targets, and the other Gaussian model focuses on the case where the scale is greater than the true value target scale, and has higher sensitivity to large scale targets;

[0020] The two Gaussian models are fused by adaptively adjusting the coefficient to obtain the scale adaptive loss method of the Gaussian mixture model.

[0021] As a further preferred solution, in the scale adaptive loss based on the Gaussian mixture model, the target scale adaptation is modeled by the Gaussian mixture model, and the specific steps are:

[0022] S21 : The sensitivity of the Gaussian mixture model to targets of different scales is enhanced by assigning adaptive weights to the IoU loss;

[0023] The calculation formula of the IoU loss is as follows:

[0024] ;

[0025] wherein, and represent the predicted graph and the true value graph, represents the count of the pixel set; is the IoU loss;

[0026] S22 : Scale adaptive loss based on Gaussian mixture model Two Gaussian models are fused to focus on different scale targets:

[0027] ;

[0028] ;

[0029] wherein, and represent two Gaussian fitting functions; and are and Change weights, It is and The fitting function, i.e., the Gaussian mixture model;

[0030] S23 : Parameters when the predicted target scale is smaller and larger than the true target scale The trend of change is shown in the following formula:

[0031] ;

[0032] ;

[0033] in, It is a parameter used to determine the range of variation of the balance function.

[0034] As a further preferred approach, in the dual-coordinate position-sensitive loss, the sensitivity of the small target detection model to target localization is enhanced by combining the Euclidean distance in Cartesian coordinates and the angular difference in polar coordinate space, including:

[0035] The Euclidean distance between the center points of the predicted target and the true target in Cartesian coordinates is calculated as the distance loss in the dual-coordinate position-sensitive loss.

[0036] The angular difference between the center points of the predicted target and the true target in polar coordinate space is calculated as the angular loss in the dual-coordinate position-sensitive loss.

[0037] By combining a scale-adaptive loss based on a Gaussian mixture model with a dual-coordinate position-sensitive loss, the predicted map generated by the infrared small target detection model is enhanced under supervision.

[0038] As a further preferred approach, in the dual-coordinate position-sensitive loss, the sensitivity of the small target detection model to target localization is enhanced by combining the Euclidean distance in Cartesian coordinates and the angular difference in polar coordinate space. The specific steps are as follows:

[0039] Given a predicted pixel set and a true pixel set, the center point of the predicted target is obtained by averaging the coordinates of all pixels in the predicted pixel set and the true pixel set, respectively. and the center point of the real target Among them, the predicted pixel set and the true pixel set refer to the pixel sets that are the targets of the predicted image and the true image, respectively.

[0040] The calculation takes into account both the Euclidean distance loss in the Cartesian coordinate system and the angular difference loss in the polar coordinate system; dual-coordinate position-sensitive loss. The definition is as follows:

[0041] ;

[0042] wherein represents a maximum operation, represents the angle difference between the center point of the prediction target and the center point of the real target when using the first coordinate origin to predict.

[0043] As a further preferred scheme, the scale adaptive loss based on the Gaussian mixture model is combined with the coordinate position sensitive loss As the final loss function of the infrared small target detection model:

[0044] .

[0045] A computer device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the above-mentioned infrared small target detection supervised enhancement method when executing the computer program.

[0046] A computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the above-mentioned infrared small target detection supervised enhancement method.

[0047] Compared with the prior art, the beneficial effects of the present application are:

[0048] 1. The scale adaptive loss based on the Gaussian mixture model and the double coordinate position sensitive loss constructed by the present application enhance the detection performance of IRSTD by enhancing the sensitivity of the model to the target scale and position.

[0049] 2. The scale adaptive loss based on the Gaussian mixture model proposed by the present application improves the perception ability of the model to different scale targets by designing two different Gaussian models.

[0050] 3. The double coordinate position sensitive loss proposed by the present application enhances the positioning sensitivity of the model to the target by combining the Euclidean distance in the Cartesian coordinate and the angle difference in the polar coordinate space, effectively locates and identifies the target, and reduces false positives and missed detections. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is the calculation principle diagram of the scale adaptive loss based on the Gaussian mixture model of the present application;

[0052] Figure 2 is the calculation principle diagram of the double coordinate position sensitive loss of the present application;

[0053] Figure 3 is the effect diagram comparison schematic diagram of the four mainstream models and other losses of the present application; ​

[0054] Figure 4 is the calculation principle diagram based on the IoU loss of the present application. DETAILED DESCRIPTION

[0055] The present application is further limited by the following description and examples with reference to the accompanying drawings, but is not limited thereto.

[0056] Terminology:

[0057] 1、IRSTD, Infrared Small Target Detection, refers to detecting small, fuzzy, and low-contrast targets (such as long-range aircraft, drones, etc.) in infrared images. Such targets usually only occupy a small part of the image (such as 1% or less of the pixels), and the background is complex, making it easy to confuse the target with clouds, ground, etc., thus posing a high challenge to the detection method. The typical IRSTD deep learning model architecture mainly includes the following modules: feature extraction module, attention mechanism enhancement module. The feature extraction module usually uses CNN (such as ResNet, U-Net) or Transformer structure to extract multi-level image features, and the attention mechanism enhancement module uses spatial attention and channel attention to enhance the perception ability of small target regions. IRSTD has very important application value in target recognition, traffic inspection, sea monitoring, disaster rescue, etc.

[0058] 2、DNANet (Dense Nested Attention Network): an infrared small target detection network based on dense nested attention mechanism, which enhances the feature expression ability of small targets by cascading densely connected multi-scale feature extraction modules and channel-spatial dual attention modules. Its core lies in solving the problem of target scale change by using cross-layer feature fusion, and suppressing complex background interference through attention mechanism.

[0059] 3、ISNet (infrared shape network): a method for extracting weak small targets from blurred background by designing Taylor finite difference (TFD) inspired edge block and bidirectional attention aggregation (TOAA) block. The TFD inspired edge block aggregates and enhances the integrated edge information from different levels to improve the contrast between the target and the background. The TOAA module uses attention mechanism to calculate the low-level information in the target row and column directions, and fuses it with the high-level information to realize the capture of target shape features and noise suppression.

[0060] 4、PBT(Progressive Background-aware Transformer): A new deep learning framework for pixel-level segmentation detection based on Transformer, which adopts an asymmetric encoder-decoder architecture. First, the encoder with specific task frequency domain prior extracts candidate target response and background context features from shallow and deep blocks, respectively. Then, the hierarchical decoder gradually refines the candidate target response under the guidance of rich background context, thereby obtaining more accurate prediction maps.

[0061] 5、MSHNet(Multi-Scale Head to the plain U-Net): A simple and effective infrared small target detection framework, which introduces a simple but effective multi-scale head into the ordinary U-Net to achieve it. The commonly used U-Net is used as the backbone network. In the decoder, feature maps with different scales are fed into different prediction heads to obtain predictions of different scales. Finally, all predictions from different feature maps are concatenated to obtain the final prediction. In the training stage, SLS loss is applied to each prediction.

[0062] 6、Cartesian coordinates, the most commonly used two-dimensional or three-dimensional coordinate system, is used to represent the position of a point. In this invention, the Cartesian coordinate system is used in image processing to represent each pixel point of the image as a point in the two-dimensional Cartesian coordinate system .

[0063] 7、Polar coordinates, a two-dimensional coordinate system, uses distance and angle to represent the position of a point. In this invention, taking the lower left corner of the image as the polar coordinate origin, the position of each pixel point in the image can be represented using distance and angle, a pixel point , where represents the distance from the origin to the point , represents the counterclockwise angle from the polar axis (usually the positive direction of the x-axis) to the line segment on which the point lies.

[0064] Example 1

[0065] An infrared small target detection supervision enhancement method, comprising:

[0066] Using a mainstream infrared small target detection model to detect small targets in the infrared image to be detected, generating a prediction map;

[0067] Wherein, the prediction map and the true value map are supervised by a scale adaptive and double coordinate position sensitive loss function based on a Gaussian mixture model to enhance the detection performance of the infrared small target detection model;

[0068] The image to be detected is preprocessed and input to the enhanced infrared small target detection model for infrared small target detection.

[0069] Embodiment 2

[0070] According to the infrared small target detection supervision enhancement method of embodiment 1, the difference lies in:

[0071] Using a mainstream infrared small target detection model, the infrared image to be detected is detected for small targets to generate a prediction map; comprising:

[0072] The improvement of the loss function of the existing model can be applied to most mainstream IRSTD frameworks at present, such as Figure 3 and Figure 4 As shown in the four mainstream methods, given the input infrared image , the infrared small target detection model is used to obtain the prediction map :

[0073] ;

[0074] Among them, refers to the infrared small target detection model IRSTD. It aims to obtain a prediction map.

[0075] The infrared small target detection model IRSTD is DNANet, ISNet, PBT or MSHNet.

[0076] In the scale adaptive loss based on Gaussian mixture model, the target scale adaptation is modeled by Gaussian mixture model, including:

[0077] Two specially designed Gaussian models are used for fusion, which respectively pay attention to multi-scale targets and adaptively adjust the loss result;

[0078] Among them, one Gaussian model has a more dramatic change when the predicted target scale is smaller than the true value target scale, and has high sensitivity to small scale targets, and the other Gaussian model pays attention to the case where the scale is larger than the true value target scale, and has higher sensitivity to large scale targets;

[0079] The two Gaussian models are fused by adaptively adjusting the coefficient to obtain the scale adaptive loss method of Gaussian mixture model.

[0080] In the scale adaptive loss based on Gaussian mixture model, the target scale adaptation is modeled by Gaussian mixture model, and the specific steps are:

[0081] S21 : As shown in Figure 1 , the sensitivity of the Gaussian mixture model to targets of different scales is enhanced by assigning adaptive weights to the IoU loss.

[0082] The Gaussian mixture model includes two Gaussian models, Figure 1 respectively The prediction target scale size changes with the range. The loss function part of the small target detection model based on the Gaussian mixture model is optimized. The GD loss function of the application uses the Gaussian mixture model to model the difference between the predicted target scale and the real scale, and accordingly assigns adaptive weights to the IoU loss. Specifically, the traditional loss function is , which optimizes the model by calculating the difference between the predicted graph and the true value graph, but the loss function has the case that the predicted scale is close to the true value scale but the loss value is unchanged, which makes it difficult for the model to optimize. The application gives adaptive weights to make the loss values of different scale targets different, thereby enhancing the sensitivity of the Gaussian mixture model to different scale targets. As shown in Figure 4 , and are two different prediction cases, the true value graph, different prediction cases, corresponding to the same loss result, will affect the optimization of the model, making the model lack sensitivity to target scale.

[0083] The calculation formula of IoU loss is as follows:

[0084] ;

[0085] wherein, and represent the predicted graph and the true value graph, represent the count of the pixel set; is the IoU loss; Intersection over Union, Intersection over Union loss, is a kind of supervision loss function commonly used in target detection and semantic segmentation tasks, which aims to more directly optimize the overlap between the predicted target and the real target.

[0086] S22 : scale adaptive loss based on Gaussian mixture model Fusion of two Gaussian models respectively paying attention to different scale targets:

[0087] ;

[0088] ;

[0089] wherein, and represent two Gaussian fitting functions; and are respectively and varying weights, is the fitting function, i.e., the Gaussian mixture model; and

[0090] S23 : Figure 1 Figures (a) and (b) respectively describe the change trend of the parameter when the predicted target size is smaller and larger than the true value target size as shown in the following formula:

[0091] ;

[0092] ;

[0093] wherein, is a parameter for balancing the range of function change. The empirical value is set to 0.2.

[0094] As shown in Figure 1 , the change trend of , and with the change of the predicted target size is visualized. When the predicted target size is smaller than the true value (as shown in (a) of , Figure 1 , shows a more significant change in relative to , which makes it more sensitive to small targets. At the same time, in order to further enhance the attention to small targets, is given a greater weight ( ) to enhance the sensitivity of small targets. Conversely, when the predicted target size is larger than the true value (as shown in (b) of , Figure 1 , shows a greater range of loss change than , indicating greater sensitivity to large-scale targets. Here is set to increase the contribution of , thereby improving the sensitivity to large targets. The design principle (as shown in (c) of Figure 1 and ) is to enhance the attention of the detector to multi-scale targets.

[0095] In the dual coordinate position-sensitive loss, the positioning sensitivity of the small target detection model to the target is enhanced by combining the Euclidean distance in Cartesian coordinates and the angle difference in polar coordinate space, including:

[0096] ​​The Euclidean distance between the center points of the predicted target and the true value target in the Cartesian coordinate is calculated as the distance loss in the dual coordinate position-sensitive loss.

[0097] The angle difference between the center points of the predicted target and the true value target in the polar coordinate space is calculated as the angle loss in the dual coordinate position-sensitive loss.

[0098] The scale adaptive loss based on the Gaussian mixture model is combined with the dual coordinate position-sensitive loss to supervise and enhance the predicted map generated by the infrared small target detection model, thereby improving the performance of the IRSTD.

[0099] In the dual coordinate position-sensitive loss, the positioning sensitivity of the small target detection model to the target is enhanced by combining the Euclidean distance in the Cartesian coordinate and the angle difference in the polar coordinate space, and the specific steps are as follows:

[0100] Given a predicted pixel set and a true pixel set, the center point of the predicted target and the center point of the true target are obtained by averaging the coordinates of all pixels in the predicted pixel set and the true pixel set respectively; wherein the predicted pixel set and the true pixel set refer to the pixel set of the predicted map and the true value map; the acquisition method is the predicted map and the true value map.

[0101] The Euclidean distance loss in the Cartesian coordinate system and the angle difference loss in the polar coordinate system are considered comprehensively to calculate; as shown in Figure 2 , the dual coordinate position-sensitive loss is defined as follows:

[0102] ;

[0103] wherein represents the maximum value operation, represents the angle difference between the center point of the predicted target and the center point of the true target when using the coordinate origin (four possible coordinate origins in the polar coordinate system: lower left corner, lower right corner, upper right corner and upper left corner). Taking the lower left corner as an example, the angle difference loss in the polar coordinate system can be obtained in the following way:

[0104] ;

[0105] As shown in Figure 2 (a), the Euclidean distance loss in is calculated by measuring the Euclidean distance between the center point of the predicted target and the center point of the true target, thereby providing a more accurate estimate of the actual spatial distance between the targets. Figure 2 (b) shows the angle difference loss in the polar coordinate system through multiple coordinate origins. Even when When the angle difference between the target in the predicted map and the target in the ground truth map is 0°, the angle difference between the targets can still be calculated by multi-view. Figure 2 As shown in (c), The intuitive principle of the design is that as the difference between the prediction center and the real center increases, The value of the loss function also increases, thereby increasing the position sensitivity of the detector and more accurately positioning the target.

[0106] It is composed of two parts: where is the Euclidean distance between the Cartesian coordinates of the target in the predicted map and the target in the ground truth map; is the angle difference between the polar coordinate space of the target in the predicted map and the target in the ground truth map. These two parts will make the model constantly fit until the result of is minimized, when the position difference of the target increases, The result of the loss function increases, which further adjusts the optimization direction and step size of the model until the result of is constantly reduced, thereby indirectly enhancing the sensitivity of the model to target positioning.

[0107] Scale adaptive loss based on Gaussian mixture model Combined with coordinate position-sensitive loss As the final loss function of the infrared small target detection model, it is a supervised enhancement method for existing mainstream models:

[0108] .

[0109] Apply to existing models to calculate the loss result of the predicted map and the ground truth map generated by the existing model and further enhance the supervision of the model to improve the performance of IRSTD.

[0110] The present application uses widely recognized evaluation indicators to report all results: the pixel-level evaluation indicators select the intersection over union (IoU) and the false detection rate (Fa), and the target-level evaluation indicators select the detection probability (Pd). Table 1 shows the quantitative comparison results of the GD loss function proposed in the present application and the widely used IoU loss, Dice loss and SLS loss on four advanced infrared small target detection methods and three data sets.

[0111] Table 1 Comparison table of detection accuracy;

[0112]

[0113] ​The experimental results show that the GD loss has a significant advantage compared with these commonly used loss functions. On the IoU index of four methods and three data sets, the GD loss is superior to other loss functions. In each of the 9 evaluation indexes of the models listed in Table 1, the GD loss achieves 8 optimal performances on the DNANet, ISNet and MSHNet models, 7 optimal and 1 sub-optimal performances on the PBT model, which fully proves the performance advantage of the GD loss.

[0114] Figure 3 For the visualization results compared with three widely used loss functions (IoU, Dice and SLS loss) in the embodiment, the correctly detected targets, the missed targets and the false detected targets are respectively framed by red, blue and yellow boxes, and the enlarged view of the target is displayed in the corner of the image; it can be seen that the existing loss function (rows 2-4) has obvious false detection and missed detection, resulting in higher Fa, however, the method of the embodiment can accurately identify and segment these small targets; therefore, the embodiment can effectively enhance the detection performance of the existing model on the infrared small target image.

[0115] Embodiment 3

[0116] A computer device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the infrared small target detection supervision enhancement method of embodiments 1 or 2 when executing the computer program.

[0117] Embodiment 4

[0118] A computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the infrared small target detection supervision enhancement method of embodiments 1 or 2.

Claims

1. An infrared small target detection supervised enhancement method, characterized in that, The method comprises the following steps: using an infrared small target detection model to perform small target detection on an infrared image to be detected to generate a prediction map; wherein the prediction map and the true value map are supervised by a scale adaptive loss based on a Gaussian mixture model and a double-coordinate position-sensitive loss function to enhance the detection performance of the infrared small target detection model; the image to be detected is preprocessed and input into the enhanced infrared small target detection model for infrared small target detection; in the scale adaptive loss based on the Gaussian mixture model, the target scale adaptation of the Gaussian mixture model is displayed by modeling, comprising: two Gaussian models are fused to focus on multi-scale targets and adaptively adjust the loss result; wherein one Gaussian model has a more drastic change when the predicted target scale is smaller than the true value target scale, and has high sensitivity to small-scale targets, and the other Gaussian model focuses on the case where the scale is larger than the true value target scale, and has higher sensitivity to large-scale targets; the two Gaussian models are fused by adaptively adjusting the coefficient to obtain a scale adaptive loss method of the Gaussian mixture model; in the double-coordinate position-sensitive loss, the positioning sensitivity of the small target detection model to the target is enhanced by combining the Euclidean distance in the Cartesian coordinate and the angle difference in the polar coordinate space, comprising: calculating the Euclidean distance between the center points of the predicted target and the true value target in the Cartesian coordinate as the distance loss in the double-coordinate position-sensitive loss; calculating the angle difference between the center points of the predicted target and the true value target in the polar coordinate space as the angle loss in the double-coordinate position-sensitive loss; combining the scale adaptive loss based on the Gaussian mixture model and the double-coordinate position-sensitive loss to supervise and enhance the prediction map generated by the infrared small target detection model.

2. The method of claim 1, wherein the method further comprises: using an infrared small target detection model to perform small target detection on an infrared image to be detected to generate a prediction map; comprising: given an input infrared image IRI, using an infrared small target detection model to obtain a prediction map P: P=Model(IRI); wherein Model refers to the infrared small target detection model IRSTD.

3. The method of claim 2, wherein the method further comprises: The infrared small target detection model IRSTD is DNANet, ISNet, PBT or MSHNet.

4. The method of claim 1, wherein the method further comprises: In the scale adaptive loss based on the Gaussian mixture model, the target scale adaptation of the Gaussian mixture model is displayed by modeling, and the specific steps are as follows: S21: enhance the sensitivity of the Gaussian mixture model to targets of different scales by assigning adaptive weights to the IoU loss; the calculation formula of the IoU loss is as follows: where P and G represent the prediction map and the ground truth map, respectively, and | · | denotes the count of pixel sets; is the IoU Loss; S22: scale adaptive loss based on gaussian mixture model Fusing two Gaussian models to focus on different scale targets respectively: s.t.w=a1×w1+a2×w2; wherein w1 and w2 represent two Gaussian fitting functions; a1 and a2 are the weights of the changes of w1 and w2 respectively, and w is the fitting function of w1 and w2, i.e. the Gaussian mixture model; S23: the change trend of the parameter w when the predicted target scale is smaller than and larger than the true value target scale is as shown in the following formula: wherein λ is a parameter for balancing the change range of the function.

5. The method of claim 1, wherein the method further comprises: In the double-coordinate position-sensitive loss, the positioning sensitivity of the small target detection model to the target is enhanced by combining the Euclidean distance in the Cartesian coordinate and the angle difference in the polar coordinate space, and the specific steps are as follows: Given the set of predicted pixels and the set of real pixels, the center point c p =(x p ,y p ) of the predicted target and the center point c gt =(x gt ,y gt ) of the real target are obtained by averaging the coordinates of all pixels in the set of predicted pixels and the set of real pixels, respectively; wherein the set of predicted pixels and the set of real pixels refer to the pixel set of the predicted image and the ground truth image for the target; The double coordinate position-sensitive loss is calculated by comprehensively considering the Euclidean distance loss in the Cartesian coordinate system and the angle difference loss in the polar coordinate system is defined as follows: where max( ) denotes a maximum value operation, Δθ i denotes the angle difference between the center point of the prediction target and the center point of the real target when using the i-th coordinate origin.

6. The method of claim 1, wherein the method further comprises: Scale adaptive loss based on gaussian mixture model With coordinate position sensitive loss Combination As the final loss function of the infrared small target detection model: 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program. The computer program is executed by the processor to realize the steps of the infrared small target detection supervised enhancement method in any one of claims 1-6.

8. A computer readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to realize the steps of the infrared small target detection supervised enhancement method in any one of claims 1-6.

Images