Infrared small target segmentation detection method based on texture enhancement network

Abstract

The present invention is a method for detecting infrared small targets segmentation based on a texture enhancement network, comprising step 1: obtaining a picture with an infrared small target from an original video data set, dividing the picture into a training set and a test set, and then marking the picture to obtain a data picture and a corresponding label image data set; step 2: inputting the obtained data picture and the label image data set into a texture enhancement network to obtain three sub-outputs; step 3: constructing a difference loss to measure the obtained three sub-output feature maps to obtain a difference loss, and calculating an average output based on the three sub-outputs; step 4: converting the obtained average output into a single-channel segmentation prediction image to obtain the output of the final prediction image. The present invention constructs a difference loss based on a texture enhancement network, approximates the network in a more efficient and accurate manner, improves the detection probability, and generates a concise and clear prediction segmentation picture without redundant clutter.

Description

Technical Field

[0001] The invention belongs to the field of computer vision, and specifically relates to an infrared small target segmentation detection method based on a texture enhancement network. Background Art

[0002] In recent years, with the continuous advancement of computer technology and the continuous improvement of computer performance, it has become a trend to use computers to replace heavy manual work. Computer vision, as an important basic component in the field of artificial intelligence research, has a wide range of applications, such as target detection, face recognition, scene reconstruction, and target tracking. Infrared small target detection is one of the key technologies of infrared search and follow, and is widely used in marine monitoring systems and early warning systems.

[0003] Infrared small target detection has attracted more and more attention in recent years. However, due to the long detection distance, the small infrared target image has a small imaging area and weak intensity, and the interference of atmospheric clouds makes it easy for infrared small targets to be submerged in the background. In addition, the imaging process will generate noise, which can easily be misdetected as a target during detection, resulting in a "high false alarm" phenomenon. In recent years, the research on infrared small target detection technology has developed rapidly, and many methods have been proposed one after another, such as the RIPT model proposed by Dai et al. that simultaneously utilizes local and non-local prior information (Dai Y, Wu Y. Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection [J]. IEEE journal of selected topics in applied earth observations and remote sensing, 2017, 10 (8): 3752-3767.), the infrared small target detection network new-IPT proposed by Kong et al. (Kong X, Yang C, Cao S, et al. Infrared small target detection via nonconvex tensor fibered rank approximation [J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 60: 1-21.), and the weighted enhanced local contrast measure algorithm wslcm proposed by Han et al. for infrared small target detection (Han J, Moradi S, Faramarzi I, et al. Infrared small target detection based on the weighted strengthened local contrast measure [J]. IEEE Geoscience and Remote Sensing Letters, 2020, 18(9): 1670-1674.). These traditional infrared small target detection methods rely heavily on manual feature settings, which makes the model unstable and difficult to adapt to scene changes. With the development of deep learning, various types of neural networks have emerged and have been applied to many computer vision tasks, such as face recognition, image classification, text detection and recognition, etc.Compared with traditional algorithms, deep learning methods can greatly improve the accuracy of many tasks and improve the robustness of algorithms. Many researchers have applied the idea of ​​deep learning to infrared small target detection. Compared with classical methods, target detection algorithms based on deep learning do not require artificial feature setting, are convenient and flexible to model, and can accurately distinguish the categories of various targets. They can accurately detect various targets while meeting real-time performance, giving full play to their advantages. The deep adversarial learning network MDvsFA_cGAN (Wang H, Zhou L, Wang L. Miss detection vs. false alarm: Adversarial learning for small objects segmentation in infrared images [C] / / Proceedings of the IEEE / CVF International Conference on Computer Vision. 2019: 8509-8518.) proposed by Wang et al. puts the infrared small target detection task into the field of deep learning. However, the existing small target detection models based on deep learning generally have the problem of low detection accuracy and easy loss of detailed texture features. Summary of the invention

[0004] In order to solve the above problems and improve the accuracy of infrared small target detection, the present invention provides an infrared small target segmentation detection method based on a texture enhancement network. The method effectively extracts the target features of the image without increasing the complexity of network calculations and has higher recognition accuracy.

[0005] In order to achieve the above object, the present invention is achieved through the following technical solutions:

[0006] The present invention is an infrared small target segmentation detection method based on texture enhancement network, comprising the following steps:

[0007] Step 1: Get pictures with small infrared targets from the original video dataset and divide them into training set and test set. Then use the annotation tool to annotate these pictures to get the data pictures and the corresponding ground_truth (label image).

[0008] Step 2: Input the data image and label image dataset obtained in step 1 into the texture enhancement network to obtain three sub-outputs.

[0009] The texture enhancement network is composed of an encoding network and a decoding network. The encoding network consists of 4 convolutional modules and 2 prior modules. The 4 convolutional modules are represented as conv1, conv2, conv3 and conv4, and the 2 prior modules are represented as p1 and p2, respectively. Among them, conv1, conv2 and conv3 each contain two identical convolutional layers and a maximum pooling layer, and conv4 contains two identical convolutional layers. The convolution kernel size of the convolutional layers contained in conv1, conv2, conv3 and conv4 is 3, and the step size is 1. The number of convolution kernels of the convolutional layers contained in conv1, conv2, conv3 and conv4 is 32, 64, 128 and 256, respectively. The pooling size and step size of the maximum pooling layer are both 2. The prior module consists of a bilinear downsampling, a convolutional layer and a Relu activation function, and the downsampling is specifically as follows: p1 is sampled to 64*64, and p2 is sampled to 32*32. The convolution kernel size of the p1 and p2 convolution layers is 3, the stride is 1, and the number of convolution kernels is 64 and 128 respectively.

[0010] The decoding network is composed of three texture enhancement modules and three classification modules to obtain some deeper features. Each classification module outputs a dual-channel feature map. The three texture enhancement modules are represented as G1, G2 and G3 respectively, and the classification module is represented as F. The texture enhancement module is composed of an upsampling operation and a neural network layer based on texture enhancement. The output sizes of the upsampling operations of the texture enhancement modules G1, G2 and G3 are 32*32, 64*64 and 128*128 respectively. The neural network layer is composed of a convolution layer. The convolution kernel size of the convolution layer corresponding to the texture enhancement module G1, the texture enhancement module G2 and the texture enhancement module G3 is 3, the step size is 1, and the number of convolution kernels is 128, 64 and 32 respectively. The classification module F is composed of an upsampling operation and a convolution layer with a convolution kernel size of 1, a step size of 1 and a number of convolution kernels of 2. The upsampling operation is specifically: selecting the output size as w*h, which is to restore the image to the original size. After three classification modules, the following are obtained: sub-output 1, sub-output 2 and sub-output 3.

[0011] The neural network layer based on texture enhancement is implemented by using anisotropic Gabor operator. It is known that Gabor operator has 5 parameters: θ, γ, λ, σ, its implementation includes the following steps:

[0012] Step 2-1: Set the parameters θ, γ, λ, σ is initialized to a random matrix of size (outchannel, inchannel), where outchannel represents the number of channels of the output feature map of this layer (that is, the number of convolution kernels corresponding to G1, G2 and G3), and inchannel represents the number of channels of the input feature map of this layer (G1, G2 and G3 are 256, 128 and 64 respectively);

[0013] Step 2-2: Set a horizontal convolution kernel x and initialize it into a two-dimensional matrix

[0014] [[-1 0 1]

[0015] [-1 0 1]

[0016] [-1 0 1]];

[0017] Step 2-3: Set a vertical convolution kernel y and initialize it into a two-dimensional matrix

[0018] [[-1 -1 -1] [0 0 0] [1 1 1]];

[0021] Step 2-4: Calculate the horizontal and vertical convolution kernels with the transformed coordinates of size (outchannel, inchannel, 3, 3) according to the following formulas (1) and (2): and

[0022]

[0023]

[0024] Step 2-5: Convolution kernel and Input formula (3) to calculate the anisotropic texture enhancement convolution kernel K(x,y) of size (outchannel, inchannel, 3, 3)

[0025]

[0026] Step 2-6: Specify K(x,y) as the convolution kernel in the texture enhancement layer to perform convolution operation to implement a neural network layer based on texture enhancement.

[0027] Step 3: Construct a difference loss to measure the three sub-output feature maps obtained in step 2, obtain the difference loss, and calculate the average output based on the three sub-outputs.

[0028] The three sub-network outputs in step 2 are represented as: out1, out2 and out3, and out4 is calculated according to formula (4):

[0029] out4=(out1+out2+out3) / 3 (4)

[0030] Step 3-1: Use formula (5) to measure the difference between the output out1, out2, out3 and the label image, and get Loss1, Loss2 and Loss3. First, calculate the mean and variance of the output and the label image, respectively, denoted as δ 1 , δ 2 , γ 1 and γ 2

[0031]

[0032] where λ 1 and λ 2 They are set to 1 and 0.01 respectively, p(x) represents the label image, and q(x) represents the probability distribution of the output; this probability distribution is calculated by the sigmoid function (binary classification function).

[0033] Step 3-2: Use formula (6) to measure the difference between out4 and the label image and get Loss4

[0034] Loss = -∑p(x)logq(x) (6)

[0035] Step 3-3: Combining steps 3-1 and 3-2, we get the difference loss of the texture enhancement network, which can be expressed by formula (7):

[0036] TotalLoss=Loss1+Loss2+Loss3+Loss4 (7).

[0037] Step 4: Convert the average output obtained in step 3 into a single-channel segmentation prediction image to obtain the output of the final prediction image. Specifically, the infrared image is subjected to the difference loss structure in step 3 to obtain the output out4, and then out4 is input into the np.argmax function (the maximum return index function in python) to obtain the final output result - a single-channel segmentation prediction image.

[0038] The beneficial effects of the present invention are: based on the backbone structure of the U-Ne network, the present invention designs a texture-enhanced infrared small target segmentation network, which is also divided into two parts, encoding and decoding. Taking into account the characteristics of the infrared small target image with little target pixel information, there is no need to add jump connections to link the encoding features and the decoding features. The simple serial structure can already meet the small target task.

[0039] Different from the ordinary U-Net network, the texture enhancement network in the present invention timely supplements the image prior information in the early encoding process to retain the key target features in the image as much as possible so that they are not lost in the layers of the network. The decoding network is mainly composed of a neural network layer based on texture enhancement, which is designed under the inspiration of Gabor filtering.

[0040] The present invention combines a convolutional neural network layer based on texture enhancement to fully extract key information of small target depth features and retain the texture features in the image to the greatest extent;

[0041] In addition, considering that small targets are small in size and easily submerged in clutter, the present invention constructs a difference loss based on the texture enhancement network, combines the information of overall features and local features, approximates the network in a more efficient and accurate way, effectively supervises the predicted segmented image, comprehensively and effectively measures the performance of the network, and further improves the network detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The invention discloses an infrared small target segmentation detection method based on a texture enhancement network.

[0043] Figure 2 It is the difference loss structure diagram.

[0044] Figure 3 These are some pictures of the data set and the corresponding comparison method results. DETAILED DESCRIPTION

[0045] The following will disclose the embodiments of the present invention with drawings. For the purpose of clear description, many practical details will be described together in the following description. However, it should be understood that these practical details should not be used to limit the present invention. That is to say, in some embodiments of the present invention, these practical details are not necessary.

[0046] like Figure 1-3 As shown, a small infrared target segmentation detection method based on texture enhancement network includes the following steps:

[0047] Step 1

[0048] Get pictures with small infrared targets from the original video dataset, divide these pictures into training set and test set, and then use the annotation tool to annotate these pictures to obtain data pictures and their corresponding ground_truth, i.e., label image dataset.

[0049] Step 2

[0050] The data image and the label image dataset obtained in step 1 are input into the texture enhancement network to obtain three sub-outputs. The texture enhancement network here is composed of an encoding network and a decoding network, in which the prior information output by the prior module is fused with the output of the convolution module by feature splicing. The specific implementation method is as follows:

[0051] For the convenience of representation, the four convolutional modules are represented as conv1, conv2, conv3 and conv4 respectively. The two prior modules are represented as p1 and p2 respectively.

[0052] The input images read are all of uniform size 128*128 and the original image length and width (w, h) are saved, and then they are input into the convolution module conv1, and the extracted features are combined with the image prior information passed through the prior module p1; the obtained feature information is input into conv2 again and spliced ​​with the prior information again; the feature information passes through 4 convolution modules: conv1, conv2, conv3 and conv4 to complete the encoding process.

[0053] Conv1, Conv2 and Conv3 all include two identical convolutional layers and one max pooling layer, while Conv4 includes two identical convolutional layers. The convolution kernel size of these convolutional layers is 3 and the stride is 1. The difference lies in the number of convolution kernels, which are 32, 64, 128 and 256 respectively. The parameters of the max pooling layer involved are consistent: the pooling size and stride are both 2.

[0054] The prior module p is composed of a bilinear downsampling layer, a convolution layer, and a Relu activation function. The downsampling operation is to facilitate concatenation with the subsequent feature maps. p1 is sampled to 64*64 and p2 is sampled to 32*32. The convolution kernel size of the p1 and p2 convolution layers is 3, the stride is 1, and the number of convolution kernels is 64 and 128 respectively.

[0055] The decoding network consists of three texture enhancement modules and a classification module to obtain some deeper features. Each classification module outputs a dual-channel feature map.

[0056] For the convenience of representation, the three texture enhancement modules are represented as G1, G2 and G3 respectively, and the classification module is represented as F.

[0057] The texture enhancement module consists of an upsampling operation and a neural network layer based on texture enhancement. The output sizes of the upsampling operations of G1, G2 and G3 are 32*32, 64*64 and 128*128 respectively. The neural network layer based on texture enhancement is composed of a convolutional layer. The convolution kernel size of the corresponding convolutional layer in G1, G2 and G3 is 3, the stride is 1, and the number of convolution kernels is 128, 64 and 32 respectively.

[0058] The classification module F is composed of an upsampling operation and a convolution layer with a convolution kernel size of 1, a step size of 1, and a convolution kernel number of 2. The upsampling operation selects an output size of w*h, which is to restore the image to the original size (w, h). After three classification modules, we get: sub-output 1, sub-output 2, and sub-output 3.

[0059] The specific implementation of the neural network layer based on texture enhancement is as follows:

[0060] Different from ordinary convolution operators, the neural network layer based on texture enhancement is implemented by using anisotropic Gabor operators. The texture enhancement layer combined with Gabor operators strengthens the scale and direction decomposition of the image, and the features in the image are fully preserved.

[0061] 1) It is known that the Gabor operator has 5 parameters: θ, γ, λ, σ. First, they are initialized to random matrices of size (outchannel, inchannel), where outchannel represents the number of channels of the output feature map of this layer (that is, the number of convolution kernels corresponding to G1, G2 and G3), and inchannel represents the number of channels of the input feature map of this layer (G1, G2 and G3 are 256, 128 and 64 respectively).

[0062] 2) Set a horizontal convolution kernel x and initialize it into a two-dimensional matrix.

[0063] [[-1 0 1]

[0064] [-1 0 1]

[0065] [-1 0 1]]

[0066] 3) Set a vertical convolution kernel y and initialize it into a two-dimensional matrix.

[0067] [[-1 -1 -1] [0 0 0] [1 1 1]]

[0070] 4) According to the following formulas (1) and (2), the horizontal and vertical convolution kernels with the transformed coordinates of size (outchannel, inchannel, 3, 3) are calculated. and

[0071]

[0072]

[0073] 5) Then the convolution kernel and Input formula (3) to calculate the anisotropic texture enhancement convolution kernel K(x,y) of size (outchannel, inchannel, 3, 3).

[0074]

[0075] 6) Specify K(x,y) as the convolution kernel in the texture enhancement layer to perform convolution operation to implement a neural network layer based on texture enhancement.

[0076] Step 3

[0077] Construct a difference loss to measure the three sub-output feature maps obtained in step 2, obtain the difference loss, and calculate the average output based on the three sub-outputs. The difference loss structure is as follows: Figure 2 As shown, the specific implementation steps are as follows:

[0078] To simplify the representation, the three sub-network outputs in step 2 are represented as: out1, out2 and out3; then out4 is calculated according to formula (4).

[0079] out4=(out1+out2+out3) / 3 (4)

[0080] 1) Use formula (5) to measure the difference between the output out1, out2, out3 and ground_truth, and get Loss1, Loss2 and Loss3. First, calculate the mean and variance of the output and ground_truth, respectively, denoted as δ 1 , δ 2 , γ 1 and γ 2 .

[0081]

[0082] where λ 1 and λ 2 Set to 1 and 0.01 respectively. p(x) represents ground_truth, q(x) represents the probability distribution of the output, which is calculated by the sigmoid function (binary classification function).

[0083] 2) Use formula (6) to measure the difference between out4 and ground_truth and get Loss4.

[0084] Loss = -∑p(x)logq(x) (6)

[0085] 3) Combining the above two parts, we get the difference loss of the texture enhancement network, which can be expressed by formula (7).

[0086] TotalLoss=Loss1+Loss2+Loss3+Loss4 (7)

[0087] Step 4

[0088] The average output obtained in step 3 is converted into a single-channel segmentation prediction image to obtain the output of the final prediction image. The output is converted into a single-channel segmentation prediction image. The specific implementation steps are as follows: the infrared image is subjected to the difference loss structure in step 3 to obtain the output out4, and then out4 is input into the np.argmax function (the return maximum index function in python) to obtain the final output result - a single-channel segmentation prediction image.

[0089] In order to verify the effectiveness of the present invention, we used multiple current mainstream methods to test on the same data set and compared the experimental results with those of the present invention.

[0090] The methods used in this paper include IPI, RIPT, srws, MDvsFA-cGAN, new-IPT and wslcm. The U-Net network is also applied to the dataset for comparison with the present invention. The experimental comparison results are shown in Figure 2. Figure 3 The experimental results are shown in Table 1.

[0091] Table 1

[0092] Recall <![CDATA[ Prec ]]> F1 <![CDATA[ mIOU ]]> Pd The present invention 0.678 0.578 0.593 0.722 0.726 <![CDATA[ MDvsFA-cGAN ]]> 0.68 0.46 0.458 0.666 0.551 IPI 0.425 0.301 0.314 0.594 0.036 RIPT 0.271 0.164 0.161 0.544 0.023 New-IPT 0.613 0.515 0.538 0.694 0.067 <![CDATA[ srws ]]> 0.165 0.944 0.281 0.163 0.345 <![CDATA[ wslcm ]]> 0.126 0.352 0.163 0.545 0.013 U-Net 0.418 0.461 0.473 0.675 0.277

[0093] From the analysis of Table 1, it can be seen that compared with the traditional small target detection method and other methods based on deep learning, the method disclosed in the present invention has different degrees of superiority in various indicators, among which the improvement of indicator Pd is particularly obvious. Figure 3 It can be found that the predicted results of the method of the present invention can produce accurate target positioning and shape segmentation outputs without redundant clutter, and the false alarm rate is very low. When facing various clutter backgrounds, target shapes and target sizes, more stable prediction outputs can be maintained.

[0094] The training set size of the dataset used is 2904, and the test set size is 100. After comparing with the infrared small target detection methods in recent years on five indicators, good results were achieved. The explanations of each indicator are as follows:

[0095] First, assume that the target is a positive sample and the background is a negative sample.

[0096] Recall: Measures how many positive samples are predicted correctly.

[0097] Prec (precision): measures how many of the samples predicted to be true are correct.

[0098] F1: The harmonic mean of Prec and Recall (2*Prec*Recall / (Prec+Recall)).

[0099] mIOU (mean intersection over union): The background and target are taken as two categories, and their intersection over union ratios are calculated separately, and then the average is calculated.

[0100] Pd (target level accuracy): measures how many predicted targets are real targets in units of targets.

[0101] From the comparison results, it can be seen that the effect of the present invention is better than that of the compared algorithms in terms of both background suppression and target enhancement.

[0102] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A small infrared target segmentation detection method based on texture enhancement network, characterized by: The infrared small target segmentation detection method comprises the following steps: Step 1: Get pictures with small infrared targets from the original video dataset, divide these pictures into training set and test set, and then use the annotation tool to annotate these pictures to obtain data pictures and the corresponding label image dataset; Step 2: Input the data image and label image dataset obtained in step 1 into the texture enhancement network to obtain three sub-outputs; Step 3: construct a difference loss to measure the three sub-output feature maps obtained in step 2, obtain the difference loss, and calculate the average output based on the three sub-outputs; Step 4: Convert the average output obtained in step 3 into a single-channel segmented prediction image to obtain the output of the final prediction image, where: In step 3, the difference loss implementation step includes: Step 3-1: The three sub-network outputs in step 2 are expressed as: out1, out2 and out3, and out4 is calculated according to formula (4) out4=(out1+out2+out3) / 3 (4) Step 3-2: Use formula (5) to measure the difference between the output out1, out2, out3 and the label image, and get Loss1, Loss2 and Loss3. First, calculate the mean and variance of the output and the label image, denoted as δ1, δ2, γ1 and γ2 respectively. Where λ1 and λ2 are set to 1 and 0.01 respectively, p(x) represents the label image, and q(x) represents the probability distribution of the output; Step 3-3: Use formula (6) to measure the difference between out4 and the label image and get Loss4 Loss = -∑p(x)logq(x) (6) Step 3-4: Combining steps 3-2 and 3-3, we get the difference loss of the texture enhancement network, which can be expressed by formula (7): TotalLoss=Loss1+Loss2+Loss3+Loss4 (7).

2. The infrared small target segmentation detection method based on texture enhancement network according to claim 1 is characterized by: The texture enhancement network in step 2 is composed of a coding network and a decoding network.

3. The infrared small target segmentation detection method based on texture enhancement network according to claim 2 is characterized by: The encoding network consists of 4 convolutional modules and 2 prior modules, the 4 convolutional modules are represented as conv1, conv2, conv3 and conv4, and the 2 prior modules are represented as p1 and p2, wherein conv1, conv2 and conv3 each contain two identical convolutional layers and a maximum pooling layer, and conv4 contains two identical convolutional layers. The convolution kernel size of the convolutional layers contained in conv1, conv2, conv3 and conv4 is 3, and the step size is 1. The number of convolution kernels of the convolutional layers contained in conv1, conv2, conv3 and conv4 is 32, 64, 128 and 256 respectively, the pooling size and step size of the maximum pooling layer are both 2, and the prior module is composed of a bilinear downsampling, a convolutional layer and a Relu activation function.

4. The infrared small target segmentation detection method based on texture enhancement network according to claim 3 is characterized by: The downsampling is specifically performed as follows: p1 is sampled to 64*64, p2 is sampled to 32*32, the convolution kernel sizes of the p1 and p2 convolution layers are both 3, the step sizes are both 1, and the numbers of convolution kernels are 64 and 128 respectively.

5. The infrared small target segmentation detection method based on texture enhancement network according to claim 2 is characterized by: The decoding network consists of three texture enhancement modules and three classification modules to obtain some deeper features. Each classification module outputs a dual-channel feature map. The three texture enhancement modules are represented as G1, G2 and G3 respectively, and the classification module is represented as F. The texture enhancement module is composed of an upsampling operation and a neural network layer based on texture enhancement. The output sizes of the upsampling operations of the texture enhancement modules G1, G2 and G3 are 32*32, 64*64 and 128*128 respectively. The neural network layer based on texture enhancement is composed of a convolution layer. The convolution kernel size of the corresponding convolution layer in the texture enhancement modules G1, G2 and G3 is 3, the step size is 1, and the number of convolution kernels is 128, 64 and 32 respectively.

6. The infrared small target segmentation detection method based on texture enhancement network according to claim 5 is characterized by: The classification module F is composed of an upsampling operation and a convolution layer with a convolution kernel size of 1, a step size of 1, and a number of convolution kernels of 2. The upsampling operation is specifically: selecting the output size as w*h, restoring the image to the original size, and passing through three classification modules to obtain: sub-output 1, sub-output 2 and sub-output 3.

7. The infrared small target segmentation detection method based on texture enhancement network according to claim 5 is characterized by: The neural network layer based on texture enhancement is implemented by using anisotropic Gabor operator. It is known that Gabor operator has 5 parameters: θ, γ, λ, σ, its implementation includes the following steps: Step 2-1: Set the parameters θ, γ, λ, σ is initialized to a random matrix of size (outchannel, inchannel), where outchannel represents the number of output feature map channels of this layer, and inchannel represents the number of input feature map channels of this layer; Step 2-2: Set a horizontal convolution kernel x and initialize it into a two-dimensional matrix [[-1 0 1] [-1 0 1] [-1 0 1]]； Step 2-3: Set a vertical convolution kernel y and initialize it into a two-dimensional matrix [[-1-1-1] [0 0 0] [1 1 1]]； Step 2-4: Calculate the horizontal and vertical convolution kernels with the transformed coordinates of size (outchannel, inchannel, 3, 3) according to the following formulas (1) and (2): and Step 2-5: Convolution kernel and lose Enter formula (3) to calculate the anisotropic texture enhancement convolution kernel K(x,y) of size (outchannel, inchannel, 3, 3) Step 2-6: Specify K(x,y) as the convolution kernel in the texture enhancement layer to perform convolution operation to implement a neural network layer based on texture enhancement.

8. The infrared small target segmentation detection method based on texture enhancement network according to claim 1 is characterized by: In step 7, the obtained average output is converted into a single-channel segmentation prediction image as follows: the infrared image is subjected to the difference loss structure in step 3 to obtain output out4, and then out4 is input into the np.argmax function to obtain the final output result - a single-channel segmentation prediction image.

Images