Infrared flying target image generation method, device, equipment and storage

Abstract

The embodiment of the invention relates to the technical field of image processing and the related technical field, and particularly discloses an infrared flight target image generation method and device, equipment and storage. According to the infrared flight target image generation method, device, equipment and storage, a visible light flight target real image or a physical simulation image is converted into an infrared flight target image, and the infrared flight target image generation method comprises the following steps that firstly, a training sample set is obtained, and the training sample set comprises a background and a flight target visible light source domain image; afterwards, the source domain image is input into an end-to-end network, which mainly comprises a content mapping module and a color perception space. And multi-scale extraction of target texture details is realized through content information coding. And finishing up-sampling reconstruction in combination with the pattern coding information. According to the method, the simulation generation quality of the infrared flying target can be effectively improved, and the background is not influenced by noise. Meanwhile, the generated infrared flying target can be fused with multiple backgrounds such as a mountain forest, and the coverage range of a simulation scene is expanded.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of image processing technology and related technical fields, and specifically disclose a method, device, equipment and storage for generating an infrared flying target image. Background Art

[0002] In recent years, low-altitude aircraft and other flying targets have developed rapidly, and disorderly flight phenomena have occurred frequently, seriously interfering with civil aviation passenger planes, affecting major events, and disrupting social public order. However, flying targets are usually small in size and difficult to capture by traditional radars. The development of flying target detection technology all day has become an important measure to ensure public safety. At present, multi-band optoelectronic tracking systems are the main means of all-day detection of low-altitude aircraft, but the measured targets are easily restricted by time, environment and other factors, and traversal image training and testing cannot be achieved. In addition, infrared shooting equipment is often too expensive, and due to many constraints such as the environment and the equipment on board, the acquisition of infrared images is easily restricted by factors such as cost and performance.

[0003] Image simulation technology has developed rapidly and has become one of the important means of data augmentation. Existing infrared image simulation technology usually uses the target and background fusion method to obtain high-quality infrared images. Since both are derived from measurement data, the simulation results are more realistic than traditional simulation results based on infrared radiation characteristics. However, due to the small reflection cross-sectional area of ​​UAV targets in remote imaging, affected by environmental noise and temperature, the image texture and edge information of UAV targets in the infrared perspective are often blurred. It is difficult to effectively obtain a target mask with clear edges through segmentation algorithms or manual marking, resulting in low quality of simulated images. Therefore, how to generate infrared images with clear edge textures has become an important problem that needs to be solved in the simulation tasks of flying targets such as UAVs.

[0004] With the rapid development of deep learning, the image style transfer algorithm based on generative adversarial network (GAN) has achieved good results in cross-domain image generation tasks, providing new ideas for infrared image simulation tasks. For example, the invention patent CN202311260125.3 "Infrared image generation method based on improved generative adversarial network" and the invention patent CN202410088720.1 "Infrared target image generation method, device, equipment and storage medium" have made certain progress in infrared target image generation by improving the generative adversarial network. However, due to the poor accuracy of infrared images, infrared images generated by existing deep algorithms will lead to a lack of style and texture information to varying degrees. In order to generate infrared images with clear edge textures of flying targets such as drones, reconstructing the local information of the target under the infrared background from the clear texture details of visible light images has become a new research idea. However, the generator part of the existing algorithms is mostly based on a fully convolutional network, and the size of the convolution kernel limits the receptive field. This makes it difficult for the generator to learn the relationship between long-distance spatial components.

[0005] Aiming at the problem of information distortion in the generation task, this paper uses the characteristics of infrared images of flying targets such as drones to propose an end-to-end infrared simulation algorithm for flying targets based on unsupervised adversarial learning. From the perspective of features, the multi-scale spatial compression module and hybrid attention are used as sub-components to extract information of targets of different dimensions and obtain infrared images of flying targets with clear edge textures. Summary of the invention

[0006] In view of the above problems, the present invention discloses an infrared flight target image generation method, device, equipment and storage, and the specific technical solution is as follows:

[0007] S1. Acquire a training sample set, wherein the training sample set includes flight target images obtained through computer simulation and real flight target images;

[0008] Specifically, the real flying target image is obtained by extracting frames from the flying target video stream collected by the optoelectronic tracking module.

[0009] Specifically, the image content obtained by the computer simulated flight target image consists of two parts: a simulated flight target image and a background. The simulated flight target image is randomly generated by the simulation software, while the background comes from the real sky or ocean background collected from the optoelectronic tracking module in step S1.

[0010] S2, using the acquired training sample set collected in S1 as input, generating an infrared flight target image. The infrared flight target image generation method is an improved generative adversarial network.

[0011] Specifically, in the process of encoding content information, the network adopts an encoder that combines multi-scale feature reconstruction and a hybrid attention module to achieve multi-scale extraction of target texture details.

[0012] Specifically, the network cooperates with the style encoding information of the color perception domain to complete upsampling reconstruction and alleviate the mode collapse phenomenon in the training stage.

[0013] Specifically, the network introduces a mean square error loss to limit the output of the generator to alleviate the high-dimensional feature loss caused by convolutional stacking and optimize the generator.

[0014] Specifically, the steps of introducing the mean square error loss to limit the output of the generator are as follows:

[0015] In order to narrow the distribution between the input image and the target domain to improve the authenticity of the generated image, the present invention uses the classic adversarial loss L gan and cycle consistency loss L cyc , which can be specifically expressed as:

[0016]

[0017] in, E: expected value; G, D, They represent the generator, discriminator, domain encoder, and content encoder respectively, where * indicates that it contains two style domains, A and B. A , G B : Generator A and Generator B, corresponding to the generated style A and style B pictures; Where x and y represent the results generated by generator A and B for the first time. Indicates obtaining the content information of result x in domain B. It means obtaining the domain style information of the result y in the domain A, and then using the generator GA to reconstruct the above information and perform a difference cycle consistency analysis with the real image A (the same applies later).

[0018] S222. After that, in order to ensure that the content space does not carry domain style information, the present invention uses a content discriminator to determine whether the content features of the input image are mapped in the same space, and uses the content discriminant loss L con To optimize:

[0019]

[0020] S223. Finally, in order to ensure that the generated target domain image corresponds to the original domain image in terms of feature dimension, the present invention introduces the mean square error loss L based on the traditional loss function of the generative adversarial network. fea , and use the feature loss between multiple layers of feature maps to update the generator.

[0021]

[0022] Where d[] is the loss of the input data calculated using mean square error; f(x): the true target domain image; The target domain image generated by the generator G. Therefore, the total loss function is expressed as follows:

[0023] L total =λ gan L gan +λ cyc L cyc +λ con L con +λ fea L fea

[0024] where λ gan , cyc , con and λ fea Represents the weight, the sum of the four is 1, and the specific value is obtained based on experimental experience.

[0025] S231, the real flying target image is collected by using an optoelectronic tracking module, mainly including a dual-spectrum pan-tilt camera of visible light and medium-wave infrared, and the collected visible light and infrared images correspond to each other at the same time;

[0026] S232. The computer simulated flight target image is collected using simulation software, mainly including Unreal Engine UE5, etc.

[0027] Specifically, the photoelectric tracking module collects the flying drone images frame by frame, and simultaneously records the drone latitude and longitude and posture information corresponding to each frame of the image, as well as the camera magnification information. By extracting these matching key data, a flying target with the same screen ratio and posture can be generated in the simulation software. In addition, the background of the simulated drone image must be the same as the real drone image it matches, which is obtained by applying image elimination technology to the real image to remove the real drone target. Only when the model accepts such an input image can it output a generated image with the same background.

[0028] S3. Compare and analyze the infrared flight target image generation method constructed in S2.

[0029] Specifically, the present invention takes the drone as the object and the sky and the ocean as the background to construct the flight images of the drone at different times.

[0030] Specifically, the present invention uses 1500 pairs of images for training and 300 pairs of images for testing in each scene. The initial network parameters of the model are randomly selected from a Gaussian distribution with a mean of 0 and a variance of 0.02. The batch size is set to 4. The network loss is optimized using the Adam optimizer, including setting the β1 and β2 parameters to 0.5 and 0.999 respectively, and training 600 times to ensure model convergence. The learning rate is fixed to 0.0001 in the first 400 rounds and then reduced to 0 linearly with the number of iterations.

[0031] Specifically, in order to more intuitively compare the performance differences of various models in the task of converting UAV visible light images to infrared images, the present invention selected infrared image generation results at different time periods (daytime and evening) under the background of sky and ocean for intuitive comparison.

[0032] The present invention also discloses a computer-readable storage medium, on which a computer program is stored. After the computer program is run, an infrared flight target image generation method as described in any one of the above embodiments is executed.

[0033] The present invention discloses a computer device, including a processor and a storage medium. The storage medium stores a computer program. The processor reads and runs the computer program from the storage medium to execute an infrared flying target image generation method as described in the above embodiment.

[0034] The advantages of the present invention are specifically as follows:

[0035] (1) The present invention constructs an infrared flight target image data set. The present invention obtains the real image of the flight target through an optoelectronic tracking module, etc., and reconstructs the data acquisition scene in a three-dimensional simulation environment through the drone and camera posture, magnification and other information recorded during the acquisition process. The information frame is aligned through GPS timing, and finally a simulated image that completely corresponds to the real image is generated. This method can also be used with randomly generated posture magnification data to obtain non-matching simulated images. Using these image data, the model can be trained and tested in an unsupervised image generation network.

[0036] (2) The present invention projects the source domain image into the content mapping module and the color perception space respectively. In the process of encoding content information, an encoder combining multi-scale feature reconstruction and a hybrid attention module is used to achieve multi-scale extraction of target texture details. The upsampling reconstruction is completed in conjunction with the style encoding information of the color perception domain, alleviating the mode collapse phenomenon in the training phase. Without relaxing the cycle consistency loss, the mean square error loss is introduced to constrain the output of each layer. Experiments show that these improvements have played a positive role in the generation of infrared flight target images.

[0037] (3) Through the design of comparative analysis methods, the performance of the present invention can be effectively verified.

[0038] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0040] Figure 1 A flowchart showing a method, device, equipment and storage for generating an infrared flying target image according to an embodiment of the present invention is shown;

[0041] Figure 2 A schematic diagram of the overall architecture of a model according to an embodiment of the present invention is shown.

[0042] Figure 3 A diagram showing a case of multi-scale feature reconstruction according to an embodiment of the present invention is shown.

[0043] Figure 4 The figure shows the experimental comparative analysis result according to the embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] like Figure 1 As shown, the present invention discloses a method, device, equipment and storage for generating an infrared flying target image, the method comprising the following steps:

[0046] Acquire a training sample set, wherein the training sample set includes visible light and infrared flight target images obtained through computer simulation and real visible light and infrared flight target images;

[0047] The collected training sample set is used as input to convert the visible light image into an infrared flight target image with clear texture. The infrared flight target image generation method is an improved generative adversarial network. In the process of encoding content information, the network adopts an encoder combining multi-scale feature reconstruction and a hybrid attention module to realize multi-scale extraction of target texture details; the upsampling reconstruction is completed in conjunction with the style encoding information of the color perception domain to alleviate the mode collapse phenomenon in the training stage; the mean square error loss is introduced to limit the output of the generator to reduce the high-dimensional feature loss caused by convolution superposition, and the generator is optimized.

[0048] A comparative analysis is conducted on the constructed infrared flying target image generation method.

[0049] Exemplarily, taking drone images as an example, the goal of the present invention is to convert visible light drone images into infrared drone images with a realistic style, so the data set should contain two types of images, namely computer-simulated visible light / infrared drone images and real-collected visible light / infrared drone images.

[0050] Specifically, in order to train the generation model of drones of different modes in a targeted manner, the present invention uses the four-rotor drone DJI M210 as the experimental object, and constructs the visible light image data set and the corresponding infrared image data set of the drone with the sky and the ocean as the background. Among them, the real flight target image is collected using an optoelectronic tracking module, mainly including a dual-spectrum gimbal camera, etc. The computer simulated flight target image is collected using simulation software, mainly including Unreal Engine UE5, etc.

[0051] Specifically, the acquisition process of the present invention is based on a 500mm fixed-focus mid-band infrared imaging lens and an artificial intelligence thermal imaging spectral pan-tilt camera for video and image acquisition. The DS-MDH003 engineering debugging tool and the camera host software are used to adjust the shooting angle and screen ratio of the lens in real time. After acquisition and screening, a total of 7,500 sets of visible light images with a resolution of 960×544 and infrared images with a resolution of 560×420 were obtained.

[0052] Specifically, the photoelectric tracking module collects the flying drone images frame by frame, and simultaneously records the drone latitude and longitude and posture information corresponding to each frame of the image, as well as the camera magnification information. By extracting these matching key data, a flying target with the same screen ratio and posture can be generated in the simulation software. In addition, the background of the simulated drone image must be the same as the real drone image it matches, which is obtained by applying image elimination technology to the real image to remove the real drone target. Only when the model accepts such an input image can it output a generated image with the same background.

[0053] Furthermore, the present invention proposes an improved generative adversarial network. In the process of encoding content information, the network uses an encoder that combines multi-scale feature reconstruction and a hybrid attention module to achieve multi-scale extraction of target texture details; it completes upsampling reconstruction with style coding information in the color perception domain to alleviate the mode collapse phenomenon in the training stage; it introduces a mean square error loss to limit the output of the generator to reduce the high-dimensional feature loss caused by convolution superposition, and optimizes the generator.

[0054] Furthermore, the content encoder proposed in the present invention uses parallel 3×1, 1×3 and 3×3 three-channel asymmetric convolutions to extract the deep features of the image. It can decompose the original convolution into three nonlinear activation layers and combine them together, thereby reducing the amount of parameter calculation and improving the ability to extract local key features in the horizontal and vertical directions.

[0055] Specifically, the extracted shallow features F i-1 As the input of multi-scale feature reconstruction (MFR) (such as Figure 3 As shown in Figure 3, feature maps S, P, and H are obtained. Before information fusion, feature information of different scales is gradually spliced ​​together through Concat to achieve information sharing and integration.

[0056] Specifically, in order to reduce the spatial redundant features caused by multi-scale encoding, a spatial reconstruction unit (SRU) is added after the asymmetric convolution, which compresses the redundant information by separately reconstructing it, aiming to separate the feature maps rich in information from the feature maps corresponding to less spatial content information. N×C×H×W , using trainable parameters γ∈R C To measure the variance of spatial pixels in each batch and channel, and obtain the importance of different feature maps according to the level of spatial information richness. The weight value of the reweighted feature map is mapped to the range of 0 to 1 through the sigmoid function, and the weight W1 rich in feature information and the weight W2 with less information are obtained through threshold gating. The whole process can be expressed as:

[0057] W=Gate(Sigmoid(W γ (GN(x))))

[0058] Where GN(x) is the group normalization layer, Gate is the gated unit, and W γ is the normalized correlation weight, expressed as:

[0059]

[0060] γ: trainable affine transformation parameter used to measure the variance of spatial pixels in each batch and channel; C: number of channels in the intermediate feature map X; w i: The weight of information richness under different channels; W γ : Normalized correlation weight. After re-separating the original feature map by weight W, the feature rich in information is obtained. and less informative features In order to further compress spatial redundant information, a cross-reconstruction operation is used to fully combine the two weighted information-rich features to achieve reconstruction of feature information at multiple scales.

[0061] Furthermore, in the encoder, the present invention proposes a robust and effective multi-level attention module SGA, which mainly includes two independent attention mechanisms, namely self-attention and global spatial attention, which serves as the middle part of the encoder and decoder layers, helps to capture broader and richer context representations and strengthen the connection of long-distance spatial components.

[0062] Specifically, the encoder features are embedded into a two-level attention module in parallel. In the self-attention module, the feature map combined with the positional encoding is mapped into three output query matrices Q, key K, value V, and projected into different linear embedding matrices. Similar to previous work, our self-attention mechanism TSA can be expressed as:

[0063]

[0064] K T : the transposed matrix of the key, is the dimension of the query / key-value sequence. At the same time, the SGA module also uses global spatial attention to selectively aggregate global context information into local information and encodes more extensive location information into local features, thereby optimizing feature representation. Input feature F∈R c×h×w Convolution is performed with two different sets of channels c and c / 8 respectively. At the same time, the generated feature maps are transposed and normalized to obtain two sets of attention maps R and S and W value outputs. The feature representation of each position is as follows:

[0065]

[0066] Among them, B p，q is the influence of position p on the element at position q, h*w: the pixel of the feature map. In order to make full use of the obtained contextual information and spatial relationships, at the end of the SGA module, a weighted combination scheme of the original feature embedding and the attention feature embedding is performed, which is defined as follows:

[0067] F SGA =μ1F TSA +μ2F GSA +F i

[0068] Among them, μ1 and μ2 are parameters that control the self-attention map and the spatial attention map, respectively. Their initial values ​​are 0 and gradually increase linearly to expand the weight.

[0069] Furthermore, the present invention introduces a mean square error loss to limit the output of the generator, so as to reduce the high-dimensional feature loss caused by convolution superposition and optimize the generator.

[0070] Specifically, in order to narrow the distribution between the input image and the target domain to improve the authenticity of the generated image, the present invention uses the classic adversarial loss L gan and cycle consistency loss L cyc , which can be specifically expressed as:

[0071]

[0072] in, E: expected value; G, D, They represent the generator, discriminator, domain encoder, and content encoder respectively, where * indicates that it contains two style domains, A and B. A , G B : Generator A and Generator B, corresponding to the generated style A and style B pictures; Where x and y represent the results generated by generator A and B for the first time. Indicates obtaining the content information of result x in domain B. It means obtaining the domain style information of the result y in the domain A, and then using the generator GA to reconstruct the above information and perform a difference cycle consistency analysis with the real image A (the same applies later).

[0073] Specifically, in order to ensure that the content space does not carry domain style information, the present invention uses a content discriminator to determine whether the content features of the input image are mapped in the same space, and uses the content discriminant loss L con To optimize:

[0074]

[0075] Specifically, in order to ensure that the generated target domain image corresponds to the original domain image in terms of feature dimension, the present invention introduces the mean square error loss L based on the traditional loss function of the generative adversarial network. fea , and use the feature loss between multiple layers of feature maps to update the generator.

[0076]

[0077] Where d[] is the loss of the input data calculated using mean squared error; f(x) is the real target domain image; f(G(Ec*, Ed*) is the target domain image generated by the generator G.

[0078] Therefore, the total loss function is expressed as follows:

[0079] L total =λ gan L gan +λ cyc L cyc +λ con L con +λ fea L fea

[0080] where λ gan , cyc , con and λ fea Represents the weight, the sum of the four is 1, and the specific value is obtained based on experimental experience.

[0081] Furthermore, the present invention conducts comparative analysis on the constructed infrared flight target image generation method.

[0082] Specifically, the present invention takes drones as objects and the sky and the ocean as backgrounds to construct drone flight images at different time periods.

[0083] Specifically, the present invention uses 1500 pairs of images for training and 300 pairs of images for testing in each scene. The initial network parameters of the model are randomly selected from a Gaussian distribution with a mean of 0 and a variance of 0.02. The batch size is set to 4. The network loss is optimized using the Adam optimizer, including setting the β1 and β2 parameters to 0.5 and 0.999 respectively, and training 600 times to ensure model convergence. The learning rate is fixed to 0.0001 in the first 400 rounds and then reduced to 0 linearly with the number of iterations.

[0084] Specifically, in order to more intuitively compare the performance differences of each model in the task of converting UAV visible light images to infrared images, the infrared image generation results of different time periods (daytime and evening) under the background of sky and ocean are selected for intuitive comparison. The specific results of each method are as follows Figure 4 As shown. Figure 4 In the figure, the first column is the visible light image input during the test, and the subsequent columns are the test results of different methods. These methods have certain application value in infrared image simulation, but they also have some common shortcomings. Judging from the generation effects of the five examples, the UVCGAN and UGATIT methods have obvious errors in the generation of drone infrared features, while CycleGAN and CUT can correctly generate infrared features. However, there are still deficiencies in the fidelity and style information of the image. Compared with other methods, the infrared features generated by the method proposed in the present invention are more accurate, clearer and more realistic.

[0085] In another embodiment of the present invention, a computer-readable storage medium is disclosed. A computer program is stored on the medium. When the computer program is run, a method for generating an infrared flying target image as described in any one of the above embodiments is executed.

[0086] In another embodiment of the present invention, a computer device is disclosed, including a processor and a storage medium, wherein a computer program is stored on the storage medium. The processor reads and runs the computer program from the storage medium to execute an infrared flight target image generation method as described in any one of the above embodiments.

[0087] Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent substitutions for some of the technical features therein; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method, device, equipment and storage medium for generating an infrared flying target image, characterized in that: The planning method comprises the following steps: S1. Acquire a training sample set, wherein the training sample set includes flight target images obtained through computer simulation and real flight target images; S2. Using the training sample set obtained in S1 as input, generate infrared flight target images. The infrared flight target image generation method is an improved generative adversarial network. Specifically, the network projects the source domain image into the content mapping module and the color perception space to improve the generation quality, and introduces the mean square error loss to constrain the output of each layer; S3. Compare and analyze the infrared flight target image generation method constructed in S2.

2. The infrared flight target image generation method, device, equipment and storage medium according to claim 1, characterized in that: In step S1, the real flying target image is obtained by extracting frames from the flying target video stream collected by the optoelectronic tracking module.

3. The infrared flying target image generation method, device, equipment and storage medium according to claim 1, characterized in that: The picture content obtained by the computer simulated flight target image in step S1 consists of two parts: a simulated flight target image and a background. The simulated flight target image is randomly generated by the simulation software, while the background comes from the real sky background collected from the optoelectronic tracking module in step S1.

4. The infrared flying target image generation method, device, equipment and storage medium according to claim 1, characterized in that: The infrared flight target image generation method is an improved generative adversarial network, and the specific steps are as follows: S211. In the process of encoding content information, an encoder combining multi-scale feature reconstruction and hybrid attention module is used to achieve multi-scale extraction of target texture details; S212. Upsampling reconstruction is completed in conjunction with the style encoding information of the color perception domain to alleviate the mode collapse phenomenon in the training stage.

5. The infrared flying target image generation method, device, equipment and storage medium according to claim 1, characterized in that: The mean square error loss is introduced to limit the output of the generator to reduce the high-dimensional feature loss caused by convolution stacking and optimize the generator. The specific steps are as follows: S221. In order to reduce the distribution between the input image and the target domain and improve the authenticity of the generated image, the present invention uses the classic adversarial loss L gan and cycle consistency loss L cyc , which can be specifically expressed as: in, E: expected value; G, D, They represent the generator, discriminator, domain encoder, and content encoder respectively, where * indicates that it contains two style domains, A and B. A , G B : Generator A and Generator B, corresponding to the generated style A and style B pictures; Among them, x and y represent the results generated by generator A and B for the first time. Indicates obtaining the content information of result x in domain B. Indicates obtaining the domain style information of the result y in domain A, and then using the generator G A Reconstruct the above information and perform difference cycle consistency analysis with the real image A (the same applies later); S222. After that, in order to ensure that the content space does not carry domain style information, the present invention uses a content discriminator to determine whether the content features of the input image are mapped in the same space, and uses the content discriminant loss L con To optimize: S223. Finally, in order to ensure that the generated target domain image corresponds to the source domain image in terms of feature dimension, the present invention introduces the mean square error loss L based on the traditional loss function of the generative adversarial network. fea , and use the feature loss between multiple layers of feature maps to update the generator; Where d[] is the loss of the input data calculated using mean square error; f(x): the true target domain image; The target domain image generated by the generator G. Therefore, the total loss function is expressed as follows: L total =λ gan L gan +λ cyc L cyc +λ con L con +λ fea L fea where λ gan , cyc , con and λ fea Represents the weight, the sum of the four is 1, and the specific value is obtained based on experimental experience.

6. A method, device, equipment and storage medium for generating an infrared flying target image according to any one of claims 1 to 5, characterized in that: include: S231, the real flying target image is collected by using an optoelectronic tracking module, mainly including a dual-spectrum pan-tilt camera of visible light and medium-wave infrared, and the collected visible light and infrared images correspond to each other at the same time; S232, the computer simulated flight target image is collected by using simulation software, mainly including Unreal Engine UE5, etc.; S233, the flying drone images are collected frame by frame through the optoelectronic tracking module, and the latitude and longitude and posture information of the drone corresponding to each frame of the image and the camera magnification information are recorded at the same time. By extracting these matching key data, a flying target with the same screen ratio and posture can be generated in the simulation software. In addition, the background of the simulated drone image must be the same as the real drone image it matches, which is obtained by applying image elimination technology to the real image to remove the real drone target. Only when the model accepts such an input image can it output a generated image with the same background.

7. The infrared flying target image generation method, device, equipment and storage medium according to claim 1, characterized in that: The comparative analysis specifically comprises the following steps: S31. The present invention takes the drone as the object and the sky and the ocean as the background to construct the drone flight images at different time periods; S32. The present invention uses 1500 pairs of images for training in each scene and 300 pairs of images for testing. The initial network parameters of the model are randomly selected from a Gaussian distribution with a mean of 0 and a variance of 0.

02. The batch size is set to 4. The network loss is optimized using the Adam optimizer, including setting the β1 and β2 parameters to 0.5 and 0.999, respectively, and training 600 times to ensure model convergence. The learning rate is fixed at 0.0001 in the first 400 rounds and then reduced linearly to the number of iterations of 0; S33. In order to more intuitively compare the performance differences of various models in the task of converting UAV visible light images to infrared images, the present invention selected infrared image generation results at different time periods (daytime and evening) under the background of sky and ocean for intuitive comparison.

8. A computer-readable storage medium, characterized in that: A computer program is stored on the medium, and after the computer program is run, an infrared flying target image generation method as described in any one of claims 1 to 7 is executed.

9. A computer device, characterized in that: It comprises a processor and a storage medium, wherein a computer program is stored in the storage medium, and the processor reads and runs the computer program from the storage medium to execute an infrared flight target image generation method as claimed in any one of claims 1 to 7.

Images