Cross-illumination-domain low-quality aircraft infrared image style migration method and system

Abstract

The invention discloses a cross-illumination-domain low-quality aircraft infrared image style migration method and system, and aims to solve the problem that domain migration is easily caused due to the fact that aircraft images acquired in acquisition scenes under different illumination conditions are different in style, so that style migration of low-quality aircraft infrared images is realized. The problem that infrared image target detection software is poor in performance and even fails under the domain offset condition is solved. The method specifically comprises the steps of image preprocessing, image style migration model building, image style migration model training, image style migration and the like. Based on a preprocessing strategy, a generative network based on multi-scale feature fusion and an attention mechanism is built in an image style migration model, and style migration is effectively achieved for a low-quality aircraft infrared image which is poor in resolution, low in contrast ratio and small in signal-to-noise ratio. The problem that the low-quality aircraft infrared image is unavailable due to target variation, background chaos, detail loss and the like in cross-illumination-domain style migration is solved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method and system for transferring the style of low-quality aircraft infrared images across illumination domains. Background Art

[0002] In the acquisition scenes with different lighting conditions, the styles of the collected aircraft images are quite different. Images of different styles are regarded as different data domains. For example, when infrared image target detection software is trained on the high-light data domain, it is easy to perform poorly or even fail in actual applications when facing night or low-light images with large domain offsets.

[0003] Image style transfer is an important means to alleviate and solve the problem of domain offset between images of different styles. The style-transferred images can be used as expanded training data to enhance the performance of infrared image target detection software, or the style-transferred images can be used as detected target images to improve the detection accuracy of infrared image target detection software.

[0004] However, the aircraft infrared images collected in practical applications are of low quality, with problems such as poor resolution, low contrast, and low signal-to-noise ratio, which brings challenges to image style transfer across illumination domains.

[0005] At present, most of the research focuses on style transfer between infrared images and visible light. There are also related technologies that realize style transfer between infrared images. However, when performing style transfer across illumination domains for poor-quality aircraft infrared images, it is unable to solve the problems of image unusability caused by target variation, background clutter, loss of details, etc. Summary of the invention

[0006] The purpose of the present invention is to provide a method and device for style transfer of low-quality aircraft infrared images across illumination domains. For aircraft infrared images collected in practical applications, an effective image preprocessing strategy is designed to enhance target information in the image, improve the signal-to-noise ratio and usability of the image, and reduce image interference factors in subsequent image processing and analysis work. Then, based on the model structure of the cyclic generative adversarial network (CycleGAN), the original generative network is improved, and a generative network based on multi-scale feature fusion and attention mechanism is designed to realize image style transfer, which can better extract infrared image features in different data domains, better solve the problem of style transfer of low-quality aircraft infrared images, and has broad application prospects.

[0007] The present invention provides a method for style transfer of low-quality aircraft infrared images across illumination domains, comprising:

[0008] Acquire original images under different lighting conditions, wherein the original images include a source domain image and a target domain image;

[0009] Preprocessing the acquired original image;

[0010] Build an image style transfer model based on the CycleGAN network;

[0011] Inputting the preprocessed image into the image style transfer model for training to obtain a trained image style transfer model;

[0012] The original image is input into the trained image style transfer model, and the image after cross-illumination domain style transfer is output.

[0013] Preferably, the preprocessing of the acquired original image includes:

[0014] Input the original grayscale images of the source domain image and the target domain image;

[0015] Filter all blank images that do not contain aircraft targets in the original images;

[0016] Crop the part showing camera parameter information to remove image noise;

[0017] The images after interference removal are resized to a uniform size.

[0018] Preferably, the image style transfer model based on the CycleGAN network is composed of a source domain image generation network, a target domain image generation network, a source domain image discrimination network and a target domain image discrimination network. The source domain image generation network is used to convert the style of the target domain image into the style of the source domain image, the target domain image generation network is used to convert the style of the source domain image into the style of the target domain image, the source domain image discrimination network is used to determine whether the image style belongs to the source domain, and the target domain image discrimination network is used to determine whether the image style belongs to the target domain; wherein, the source domain image generation network and the target domain image generation network are both generation networks based on multi-scale feature fusion and attention mechanism, and the network structures of the two are consistent.

[0019] Preferably, the building of an image style transfer model based on a CycleGAN network includes:

[0020] Build a generative network: Configure the generative network as an encoder-decoder structure, where the encoder is a CNN-based backbone network and the decoder has a fixed structure, which is used to fuse the multi-scale features of the encoder.

[0021] Input the original image and obtain the feature maps of different scales extracted by CNN through the encoder;

[0022] The decoder optimizes the deepest image feature map extracted by CNN through a connected residual structure, and uses a channel re-evaluation attention module to calculate the importance scores of feature maps of different scales and weight the original feature maps, so as to focus on all key positions of multi-scale features and effectively capture the global feature dependencies of the image; the decoder also adopts a multi-scale feature fusion method, which is used to start from the optimized deepest image feature map, sequentially fuse the weighted features extracted by CNN of the same size and upsample, and obtain the final result as the output image of the generation network;

[0023] Build a discriminant network: divide the image to be discriminated into multiple small image blocks, judge the image style of each small image block respectively, and calculate the average result of all small image blocks as the final discriminant result. Among them, the discriminant network is PatchGAN of the original CycleGAN.

[0024] Preferably, the step of inputting the preprocessed image into the image style transfer model for training further comprises:

[0025] The source domain image and the target domain image are input into the CNN backbone network of the encoder in the source domain generation network and the target domain generation network respectively, and the feature map {F1, F2, ..., F L};

[0026] For the deep feature map F L , using the residual structure to optimize, and obtain the optimized deep feature map S L ;

[0027] For the F L The external feature map F i , based on the attention mechanism, the average pooling operation is performed to obtain the compressed feature map, and the global deep convolution operation is input to obtain the attention weight, which is the importance score of each channel of the feature map. i Weighted, get the optimized F′ i ;

[0028] From the deep feature map S through multi-scale feature fusion of the decoder L Start by combining the optimized F′ of the same size i , upsampling is performed using a transposed convolution operation;

[0029] Calculate the loss function of the generation network and the discriminator network;

[0030] According to the loss function of the generated network, back propagation is performed to update the network parameters;

[0031] Repeat the training process of the above steps until the iteration of model training is completed.

[0032] Preferably, the loss functions of the computing generation network and the discriminant network include:

[0033] The total loss function of the source domain generation network and the target domain generation network is:

[0034]

[0035] in, It is an adversarial loss based on Wasserstein distance. is the cycle consistency loss, is the identity mapping loss;

[0036] Source domain discrimination network D s The loss function is:

[0037]

[0038] Target domain discrimination network D t The loss function is:

[0039]

[0040] Among them, G t2s (·) represents the image output by the source domain image generation network, G s2t (·) represents the image output by the target domain image generation network, D s (·) represents the discrimination result output by the source domain image discrimination network, D t (·) represents the discrimination result output by the target domain image discrimination network, X and Y represent the source domain and target domain image sets respectively, x and y represent the single source domain image and target domain image from X and Y respectively, P X and P Y Represent the feature distribution of target domain images and source domain images respectively.

[0041] Preferably, the updating of network parameters further comprises:

[0042] According to the loss function of the generated network, back propagation is performed to update the parameters of the generated network and keep the parameters of the discriminative network unchanged;

[0043] According to the loss function of the discriminant network, back propagation is performed to update the parameters of the discriminant network and keep the parameters of the generated network unchanged.

[0044] The present invention also provides a system for transferring the style of low-quality aircraft infrared images across illumination domains, which implements the method for transferring the style of low-quality aircraft infrared images across illumination domains as described in an embodiment of the present invention, including:

[0045] A data acquisition module, used to acquire original images under different lighting conditions, wherein the original images include source domain images and target domain images;

[0046] A preprocessing module, used for preprocessing the acquired original image;

[0047] Model building module, used to build an image style transfer model based on the CycleGAN network;

[0048] A model training module, used to input the preprocessed image into the image style transfer model for training to obtain a trained image style transfer model;

[0049] The style transfer module is used to input the original image into the trained image style transfer model and output the image after cross-illumination domain style transfer.

[0050] The present invention also provides an electronic device, comprising:

[0051] A memory, the memory being used to store a processing program;

[0052] A processor, wherein when executing the processing program, the processor implements the method for transferring the style of low-quality aircraft infrared images across illumination domains as described in an embodiment of the present invention.

[0053] The present invention also provides a readable storage medium, on which a processing program is stored. When the processing program is executed by a processor, the method for cross-illumination domain style transfer of low-quality aircraft infrared images as described in an embodiment of the present invention is implemented.

[0054] The beneficial effects of the present invention compared with the prior art are:

[0055] The present invention provides a method and system for style transfer of low-quality aircraft infrared images across illumination domains, designs an effective preprocessing strategy for low-quality aircraft infrared images, effectively reduces certain interference factors, and the generative network based on multi-scale feature fusion and attention mechanism designed in the CycleGAN model structure can more accurately extract the global and local features of aircraft infrared images, thereby significantly improving the quality of image style transfer. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flow chart of a method for style transfer of low-quality aircraft infrared images across illumination domains in an embodiment of the present invention;

[0057] Figure 2 is a schematic diagram of interference factors before image preprocessing in an embodiment of the present invention;

[0058] Figure 3 A schematic diagram of a multi-scale feature fusion module for generating a network in an embodiment of the present invention;

[0059] Figure 4 Schematic diagram of a channel re-evaluation attention module in an embodiment of the present invention;

[0060] Figure 5 The figure shows the comparison of the generated images of the style transfer of low-quality aircraft infrared images in the embodiment of the present invention. DETAILED DESCRIPTION

[0061] 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.

[0062] Embodiment 1

[0063] The present invention provides a method for style transfer of low-quality aircraft infrared images across illumination domains, including four parts: image preprocessing, building an image style transfer model, training the image style transfer model, and image style transfer. The specific steps are as follows:

[0064] S1: Acquire original images under different lighting conditions, the original images include source domain images and target domain images; the original images are images collected under different lighting conditions of aircraft take-off and landing and flight status, and the source domain images and target domain images are images of different lighting domains. For example, in the image style transfer task, the goal is to convert the style of the infrared image of the aircraft taken at night, i.e., the source domain image, to the style of the infrared image of the aircraft taken during the day, i.e., the target domain image. In this case: Source domain image: infrared image of the aircraft taken at night, characterized by a black background and a white aircraft. Target domain image: infrared image of the aircraft taken during the day, characterized by a white background and a black aircraft.

[0065] S2: preprocessing the acquired original image;

[0066] Step a: Image preprocessing, specifically including the following:

[0067] The original grayscale images of the source domain image and the target domain image are input; the blank images that do not contain the aircraft target in all the original images are filtered; the part showing the camera parameter information is cropped to remove the image interference; and the images after the interference is removed are processed to a uniform image size.

[0068] In this embodiment, the image is the original grayscale image of the input source domain image and the target domain image, see Figure 2First, remove all blank images that do not contain aircraft targets, then crop the part of the image that shows the camera parameter information, and finally unify the image size.

[0069] S3: Build an image style transfer model based on the CycleGAN network;

[0070] Step b: Build an image style transfer model

[0071] The image style transfer model is built based on the CycleGAN structure. The image style transfer model consists of a source domain image generation network, a target domain image generation network, a source domain image discrimination network, and a target domain image discrimination network. Among them, the two image generation networks have the same structure, which is a generation network designed based on multi-scale feature fusion and attention mechanism; the two image discrimination networks have the same structure, which is a discrimination network based on PatchGAN. The source domain image generation network is used to convert the style of the target domain image into the style of the source domain image, the target domain image generation network is used to convert the style of the source domain image into the style of the target domain image, the source domain image discrimination network is used to determine whether the image style belongs to the source domain, and the target domain image discrimination network is used to determine whether the image style belongs to the target domain. Specifically, it includes the following steps:

[0072] Step b1: Build the generation network

[0073] The designed generative network based on multi-scale feature fusion and attention mechanism is an encoder-decoder structure. The encoder selects a suitable CNN-based backbone network (Backbone) according to the computing resource constraints and actual needs. Generally, the network with more mature applications such as U-Net, VggNet and ResNet is selected. Figure 3 , the decoder has a fixed structure and is used to fuse the multi-scale features of the encoder. For the input image, the encoder obtains feature maps of different scales extracted by CNN. The decoder first uses the residual structure to optimize the deepest image feature map extracted by CNN, and then uses the channel re-evaluation attention module to calculate the importance scores of feature maps of different scales and weight the original feature maps. It can pay attention to all key positions of multi-scale features and effectively capture the global feature dependencies of the image. The decoder also adopts a multi-scale feature fusion method, starting from the optimized deepest image feature map, and sequentially fuses the weighted features extracted by CNN of the same size and upsamples them. The final result is used as the output image of the generated network.

[0074] Step b2: Build a discriminant network

[0075] The discriminant network is PatchGAN derived from the original CycleGAN. The image to be discriminated is divided into small blocks, and the image style of each small block is judged separately. The average result of all small blocks is calculated as the final discriminant result.

[0076] S4: Inputting the preprocessed image into the image style transfer model for training to obtain a trained image style transfer model. The specific steps include:

[0077] The source domain image and the target domain image are input into the CNN backbone network of the encoder in the source domain generation network and the target domain generation network respectively, and the feature map {F1, F2, ..., F L};

[0078] For the deep feature map F L , using the residual structure to optimize, and obtain the optimized deep feature map S L ;

[0079] For the F L The external feature map F i , based on the attention mechanism, the average pooling operation is performed to obtain the compressed feature map, and the global deep convolution operation is input to obtain the attention weight, which is the importance score of each channel of the feature map. i Weighted, get the optimized F′ i ;

[0080] From the deep feature map S through multi-scale feature fusion of the decoder L Start by combining the optimized F′ of the same size i , upsampling is performed using a transposed convolution operation;

[0081] Calculate the loss function of the generation network and the discriminator network;

[0082] According to the loss function of the generated network, back propagation is performed to update the network parameters;

[0083] Repeat the training process of the above steps until the iteration of model training is completed.

[0084] In this embodiment, it can be understood that step c: training the image style transfer model

[0085] The preprocessed images are used to train the image style transfer model, which includes the following steps:

[0086] Step c1: Generate feature maps of images using the encoder

[0087] The source domain image and the target domain image are input, and the feature map of the image is generated by the CNN backbone network of the encoder in the source domain generation network and the target domain generation network respectively.

[0088] The training process of the target domain generation network is the same as that of the source domain generation network. Taking the training process of the source domain generation network as an example, assuming that the number of layers encoded by the source domain generation network encoder is L, L feature maps {F1, F2, ..., F L}.

[0089] Step c2: Optimize deep feature maps using residual structures

[0090] For the deep feature map F L , using the residual structure to optimize, and obtain the optimized deep feature map S of the same size L , the specific calculation is as follows:

[0091] S L =F L +conv(relu(conv(F L ))),

[0092] Among them, conv(·) represents the convolution operation with a convolution kernel of 3×3, and relu(·) represents the activation function.

[0093] Step c3: Use attention mechanism to optimize the feature map of the image

[0094] See also Figure 4 , for the feature map F i (i≠L), first apply the average pooling operation to obtain the compressed feature map, and then apply the global depth convolution operation to obtain the attention weight, which can be regarded as the importance score of each channel of the feature map. i Weighted, get the optimized F′ i , the specific calculation is as follows:

[0095]

[0096] in, represents channel multiplication, GDconv(·) represents the global depthwise convolution operation, and pool(·) represents the average pooling operation.

[0097] Step c4: Multi-scale feature fusion

[0098] See also Figure 3 , the multi-scale features of the decoder are fused with S obtained from step c2 L First, combine the same size F′ obtained in step c3 i , upsampling is performed using the transposed convolution operation, which is calculated as follows:

[0099]

[0100] Among them, Tconv(·) represents the transposed convolution operation with a convolution kernel of 3×3. Represents a concatenation operation.

[0101] When i=1, S0 calculated using the above formula is the generated image output by the generative network.

[0102] Step c5: Calculate the loss function

[0103] Adversarial loss of CycleGAN combined with Wasserstein distance The optimization goals are:

[0104]

[0105] Among them, G t2s (·) represents the image output by the source domain image generation network, G s2t (·) represents the image output by the target domain image generation network, D s (·) represents the discrimination result output by the source domain image discrimination network, D t (·) represents the discrimination result output by the target domain image discrimination network, X and Y represent the source domain and target domain image sets, respectively, and x and y represent single source domain and target domain images from X and Y, respectively. and Represent the feature distributions of the generated target domain style and source domain style images respectively.

[0106] The style-transferred image G s2t (X) Input G t2s , G t2s (Y) Input G s2t , calculate the cycle consistency loss To constrain cycle consistency between generators:

[0107]

[0108] Among them, P X and P Y Represent the feature distribution of target domain and source domain images respectively.

[0109] Enter X into G t2s ,Y input G s2t , the identity mapping loss is used to ensure the robustness of the source domain and target domain generation network. The specific calculation is as follows:

[0110]

[0111] Combining the loss function of the above generated network, the total loss function of the generated network is:

[0112]

[0113] Source domain discrimination network D s The loss function is:

[0114]

[0115] Target domain discrimination network D t The loss function is:

[0116]

[0117] Step c6: Perform back propagation according to the loss function of the generated network to update G t2s and H s2t Parameters, keep D s and D t The parameters remain unchanged.

[0118] According to D s Loss function, perform back propagation, update D s Parameters, keep D t , G t2s and G s2t The parameters remain unchanged.

[0119] According to D t Loss function, perform back propagation, update D t Parameters, keep D s , G t2s and G s2t The parameters remain unchanged.

[0120] Repeat the training process of steps c1-c6 above until the iteration of model training is completed.

[0121] S5: Input the original image into the trained image style transfer model and output the image after cross-illumination domain style transfer.

[0122] Step d: Image style transfer

[0123] By using the trained image style transfer model, inputting the source domain image into the target domain image generation network can obtain the source domain image with the target domain style, and inputting the target domain image into the source domain image generation network can obtain the target domain image with the source domain style.

[0124] See also Figure 5 ,in, Figure 5(a) is the original infrared image of the aircraft. The first row is the image collected at night (characterized by a black background and a white aircraft), which is used as the source domain image; the last two rows are the images collected during the day (characterized by a white background and a black aircraft), which is used as the target domain image. Figure 5 (b) is the image style transfer result obtained using the existing Wasserstein CycleGAN (proposed by Shan et al. in MICCAI2021). Figure 5 (c) is the image style transfer result obtained by using the existing Wasserstein CycleGAN after adopting the image preprocessing strategy of the present invention. Figure 5 (d) is the image style transfer result obtained by using the present invention in Example 1.

[0125] The image style transfer result in this embodiment ( Figure 5 (d)) and other methods of image style transfer results ( Figure 5 (b) and Figure 5 (c)), it can be seen that the image preprocessing strategy of the embodiment of the present invention is effective, the image style transfer result of the present invention is closer to the expected result, and the problems of target variation, background confusion, detail loss, etc. are avoided, and the image style transfer effect is better.

[0126] Based on the same inventive concept, the present invention also provides a system for transferring the style of low-quality aircraft infrared images across illumination domains, which implements the method for transferring the style of low-quality aircraft infrared images across illumination domains as described in Embodiment 1 of the present invention, including:

[0127] A data acquisition module, used to acquire original images under different lighting conditions, wherein the original images include source domain images and target domain images;

[0128] A preprocessing module, used for preprocessing the acquired original image;

[0129] Model building module, used to build an image style transfer model based on the CycleGAN network;

[0130] A model training module, used to input the preprocessed image into the image style transfer model for training to obtain a trained image style transfer model;

[0131] The style transfer module is used to input the original image into the trained image style transfer model and output the image after cross-illumination domain style transfer. The principle thereof is referred to the above method embodiment and will not be repeated here.

[0132] The present invention also provides an electronic device, comprising:

[0133] A memory, the memory being used to store a processing program;

[0134] A processor, wherein when executing the processing program, the processor implements the method for transferring the style of low-quality aircraft infrared images across illumination domains as described in an embodiment of the present invention.

[0135] The present invention also provides a readable storage medium, on which a processing program is stored. When the processing program is executed by a processor, the method for transferring the style of low-quality aircraft infrared images across illumination domains as described in an embodiment of the present invention is implemented.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. 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 replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for style transfer of low-quality aircraft infrared images across illumination domains, characterized in that: include: Acquire original images under different lighting conditions, wherein the original images include a source domain image and a target domain image; Preprocessing the acquired original image; Build an image style transfer model based on the CycleGAN network; Inputting the preprocessed image into the image style transfer model for training to obtain a trained image style transfer model; The original image is input into the trained image style transfer model, and the image after cross-illumination domain style transfer is output.

2. The method for transferring the style of low-quality aircraft infrared images across illumination domains as claimed in claim 1, characterized in that: The preprocessing of the acquired original image includes: Input the original grayscale images of the source domain image and the target domain image; Filter all blank images that do not contain aircraft targets in the original images; Crop the part showing camera parameter information to remove image noise; The images after interference removal are resized to a uniform size.

3. The method for transferring low-quality aircraft infrared images across illumination domains as claimed in claim 1, characterized in that: The image style transfer model based on the CycleGAN network is composed of a source domain image generation network, a target domain image generation network, a source domain image discrimination network and a target domain image discrimination network. The source domain image generation network is used to convert the style of the target domain image into the style of the source domain image, the target domain image generation network is used to convert the style of the source domain image into the style of the target domain image, the source domain image discrimination network is used to judge whether the image style belongs to the source domain, and the target domain image discrimination network is used to judge whether the image style belongs to the target domain; wherein, the source domain image generation network and the target domain image generation network are both generation networks based on multi-scale feature fusion and attention mechanism, and the network structures of the two are consistent.

4. The method for transferring style of low-quality aircraft infrared images across illumination domains as claimed in claim 1, characterized in that: The image style transfer model based on the CycleGAN network includes: Build a generative network: Configure the generative network as an encoder-decoder structure, where the encoder is a CNN-based backbone network and the decoder has a fixed structure, which is used to fuse the multi-scale features of the encoder. Input the original image and obtain the feature maps of different scales extracted by CNN through the encoder; The decoder optimizes the deepest image feature map extracted by CNN through a connected residual structure, and uses a channel re-evaluation attention module to calculate the importance scores of feature maps of different scales and weight the original feature maps, so as to focus on all key positions of multi-scale features and effectively capture the global feature dependencies of the image; the decoder also adopts a multi-scale feature fusion method, which is used to start from the optimized deepest image feature map, sequentially fuse the weighted features extracted by CNN of the same size and upsample, and obtain the final result as the output image of the generation network; Build a discriminant network: divide the image to be discriminated into multiple small image blocks, judge the image style of each small image block respectively, and calculate the average result of all small image blocks as the final discriminant result. Among them, the discriminant network is PatchGAN of the original CycleGAN.

5. The method for transferring low-quality aircraft infrared images across illumination domains as claimed in claim 1, characterized in that: The inputting the preprocessed image into the image style transfer model for training further comprises: The source domain image and the target domain image are respectively input into the CNN backbone network of the encoder in the source domain generation network and the target domain generation network to generate the feature map {F1, F2, ..., F L }; For the deep feature map F L , using the residual structure to optimize, and obtain the optimized deep feature map S L ; For the F L The external feature map F i , based on the attention mechanism, the average pooling operation is performed to obtain the compressed feature map, and the global deep convolution operation is input to obtain the attention weight, which is the importance score of each channel of the feature map. i Weighted, get the optimized F′ i ; From the deep feature map S through multi-scale feature fusion of the decoder L Start by combining the optimized F′ of the same size i , upsampling is performed using a transposed convolution operation; Calculate the loss function of the generation network and the discriminator network; According to the loss function of the generated network, back propagation is performed to update the network parameters; Repeat the training process of the above steps until the iteration of model training is completed.

6. The method for transferring the style of low-quality aircraft infrared images across illumination domains as claimed in claim 5, characterized in that: The loss functions of the computing generation network and the discriminant network include: The total loss function of the source domain generation network and the target domain generation network is: in, It is an adversarial loss based on Wasserstein distance. is the cycle consistency loss, is the identity mapping loss; Source domain discrimination network D s The loss function is: Target domain discrimination network D t The loss function is: Among them, C t2s (·) represents the image output by the source domain image generation network, G s2t (·) represents the image output by the target domain image generation network, D s (·) represents the discrimination result output by the source domain image discrimination network, D t (·) represents the discrimination result output by the target domain image discrimination network, X and Y represent the source domain and target domain image sets respectively, x and y represent the single source domain image and target domain image from X and Y respectively, P X and P Y Represent the feature distribution of target domain images and source domain images respectively.

7. The method for transferring style of low-quality aircraft infrared images across illumination domains as claimed in claim 5, characterized in that: The updating of network parameters further comprises: According to the loss function of the generated network, back propagation is performed to update the parameters of the generated network and keep the parameters of the discriminative network unchanged; According to the loss function of the discriminant network, back propagation is performed to update the parameters of the discriminant network and keep the parameters of the generated network unchanged.

8. A low-quality aircraft infrared image style transfer system across illumination domains, characterized in that: Implementing the method for cross-illumination domain low-quality aircraft infrared image style transfer as claimed in any one of claims 1 to 7, comprising: A data acquisition module, used to acquire original images under different lighting conditions, wherein the original images include source domain images and target domain images; A preprocessing module, used for preprocessing the acquired original image; Model building module, used to build an image style transfer model based on the CycleGAN network; A model training module, used to input the preprocessed image into the image style transfer model for training to obtain a trained image style transfer model; The style transfer module is used to input the original image into the trained image style transfer model and output the image after cross-illumination domain style transfer.

9. An electronic device, characterized in that: include: A memory, the memory being used to store a processing program; A processor, wherein when executing the processing program, the processor implements the method for cross-illumination domain low-quality aircraft infrared image style transfer as described in any one of claims 1 to 7.

10. A readable storage medium, characterized in that: The readable storage medium stores a processing program, and when the processing program is executed by the processor, the method for cross-illumination domain low-quality aircraft infrared image style transfer is implemented as described in any one of claims 1 to 7.

Images