Forest fire smoke image generation method and system based on improved CycleGAN

Abstract

The invention relates to an improved CycleGAN-based forest fire smoke image generation method and system, and belongs to the crossing field of a forest fire safety technology, a deep learning technology and a computer vision identification technology. According to the method, the potential distribution rule of fog can be learned under the condition that there is no paired foggy image-fogless image, and the foggy image corresponding to the input fogless image is synthesized and output. A CycleGAN framework is adopted in the model, a double-branch generator is used for synthesizing a foggy image, a physical formula of an atmospheric scattering model, a dense residual module and a double-path attention module are introduced, and training schemes of feature reconstruction loss, confrontation loss and feature matching loss are optimized. The pixel-pixel corresponding paired images with strong physical constraints are generated, and the precision of the generated images is improved. The method for generating the non-paired image of the forest fire smoke image provides an effective solution for solving the problem that a high-quality paired data set in an image defogging task in recent years is difficult to obtain.

Description

Technical Field

[0001] The present invention belongs to the cross - field of forest fire safety technology, deep learning technology and computer vision recognition technology, and particularly relates to a method and system for generating forest fire smoke images based on improved CycleGAN. Background Art

[0002] Forest fires are natural disasters that seriously affect the earth's environmental ecology. When forest fires have occurred and spread widely, technologies such as flame detection may not be useful. Moreover, the thick smoke generated by combustion will obscure the surface conditions, causing huge visual interference and affecting fire fighting and rescue. Datasets related to forest fires are scarce, and it has always been a difficult problem to obtain datasets for the defogging task. In the real world, it is difficult for us to obtain clear and foggy image pairs under the same background because it is difficult to ensure that background objects and lighting conditions do not change when shooting at different times. Without reference images (real ground), it is difficult to implement objective evaluation methods such as PSNR and SSIM algorithms. Even if there are examples of training defogging models using unpaired datasets, their experiments will finally use paired datasets to verify the performance of the models. A method that can effectively solve the problem of dataset shortage must be found.

[0003] Currently, the research on synthesizing paired defogging datasets mainly focuses on using atmospheric scattering models and scene depth maps for synthesis, by simulating smoke in real - world scenarios and taking pictures. However, they usually assume that the defogging scenario is indoor or outdoor, and there are few patents on generating smoke images caused by forest fires, etc. Although the research on image defogging for traditional indoor and urban scenes is important, there is still a large gap compared with the forest fire scene. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for generating forest fire smoke images based on improved CycleGAN, which can simulate clear forest graphics and use deep learning technology to generate paired high - quality forest fire smoke images from clear forest images to solve the problem of obtaining paired datasets related to forest fires.

[0005] To achieve the above object, the technical solution of the present invention is: A method and system for generating forest fire smoke images based on improved CycleGAN, including:

[0006] Step S1: Obtain a large number of random foggy - fog - free random images as input images;

[0007] Step S2: Pre - process the obtained images, scale the input pictures by 1.12 times and then randomly crop them into a size of 256*256, and then randomly horizontally flip them as a slight image enhancement;

[0008] Step S3: Modify the generator G in the CycleGAN structure of the generative adversarial network A into a dual-branch structure, including a transmission rate learning subnet and an atmospheric light intensity learning subnet;

[0009] Step S4: Introduce dense residual blocks and dual-path attention modules to optimize feature reconstruction, and jointly train using adversarial loss, cycle consistency loss, and cycle perception consistency loss; by minimizing the total loss function, optimize the parameters of the generator and discriminator, so that the generator can generate images with natural smoke effects from clear images;

[0010] Step S5: Perform unsupervised learning using the improved CycleGAN of the generative adversarial network, and train two generators and two discriminators, where:

[0011] Generator G A is used to generate images with smoke from clear images; Generator G B is used to recover clear images from images with smoke; Discriminator D A is used to judge the authenticity of the generated smoke images; Discriminator D B is used to judge the authenticity of the generated clear images;

[0012] Step S6: Use the trained improved CycleGAN network model to perform fog generation processing on the clear images to be processed.

[0013] Furthermore, in Step S3, the transmission rate learning subnet adopts an encoder-decoder architecture. The transmission rate learning subnet consists of three parts: an encoder, a decoder, and a residual dense module. The encoder consists of an initial layer and two downsampling layers. The residual dense module consists of 2 residual dense groups. The residual dense group is combined by a residual dense block and a feature attention module. The decoder uses two upsampling operations and a Tanh activation function.

[0014] Furthermore, the transmission rate learning subnet also sets an average pooling layer, multiplies the feature map output by the encoder after passing through the average pooling layer with the feature map output by the residual dense module, and uses it as the input of the decoder.

[0015] Furthermore, in Step S3, the atmospheric light intensity learning subnet consists of three convolutional layers. The input image enters the first convolutional layer to initially extract features, eliminates unnecessary information in the feature space through a global average pooling layer, and then passes through two convolutional layers with a convolutional kernel size of 1 and a sigmoid activation function to obtain the final atmospheric light value A.

[0016] Further, in step S3, a foggy image is synthesized according to the atmospheric scattering model formula I(x) = J(x)t(x) + A(1 - t(x)), where t(x) is the transmission rate map output by the transmission rate learning subnet, A is the atmospheric light intensity output by the atmospheric light intensity learning subnet, J(x) is the input image, and I(x) is the output foggy image.

[0017] Further, in step S4, the adversarial loss and the cycle consistency loss are as follows:

[0018] For the mapping function G A : X → Y and its discriminator D B , the objective is expressed as:

[0019] L GAN (G A , D B ) = E X～Pdata(X) [log(1 - D B (G A (X))))] + E Y～Pdata (Y)[logD B (Y)]

[0020] L GAN (G A , D B ) is the adversarial loss; E X～Pdata(X) represents the expected value of the cycle consistency loss sampled in the X domain (smokeless image); E Y～Pdata(Y) represents the expected value of the cycle consistency loss sampled in the Y domain (smoky image).

[0021] For each generated fake image G A (X) → Y', it must be able to be restored to the original image through the mapping function G B : Y → X, that is, G B (G A (X)) = X r ≈ X, where X r represents the image restored after being mapped twice by the generators G A and G B , that is, the smokeless image restored by G B (G A (X)); the constraint between them is called the cycle consistency loss, and is expressed by the formula:

[0022] L cycle (G A , D B ) = E X～pdata(X) [||G B (GA(X)) - X||1] + E Y～Pdata (Y)[||GA (G B (Y)) - Y || 1]

[0023] L cycle (G A , D B ) represents the cycle consistency loss; E X～Pdata(X) represents the expected value of the cycle consistency loss sampled on the X domain (smokeless images), which makes CycleGAN not destroy smokeless images and ensures that the original image can be restored after dehazing; E Y～Pdata(Y) represents the expected value of the cycle consistency loss sampled on the Y domain (smoky images), which makes CycleGAN not destroy smoky images and ensures that the original image can be restored after adding back the smoke.

[0024] Furthermore, in step S4, the cycle-aware consistency loss is specifically as follows:

[0025] Assume that after passing through the cycle G B (G A (X)) = X r , G A (G B (Y)) = Y r ; X, X r and Y, Y r should have structural similarity in features at high and low levels, where X r represents the image restored after being mapped twice by the generators G A and G B , that is, the smokeless image restored by G B (G A (X)); Y r represents the image restored after being mapped twice by the generators G B and G A , that is, the smoky image restored by G A (G B (Y)); The cycle-aware consistency loss L Perceptual uses the pre-trained VGG19 model to output the features extracted from the second pooling layer and the fifth pooling layer, and the loss is compared through the L2 norm. The formula is as follows:

[0026]

[0027] where represents the features extracted by the VGG19 model from the second pooling layer and the fifth pooling layer; represents the features obtained from the second pooling layer and the fifth pooling layer of the original smokeless image; represents the features obtained from the second pooling layer and the fifth pooling layer of the image restored after being mapped twice; Represents the features obtained from the second pooling layer and the fifth pooling layer of the original smoky image; Represents the features obtained from the second pooling layer and the fifth pooling layer of the image restored through two mappings.

[0028] Furthermore, the total loss function is expressed as follows:

[0029] L = λ GAN *L GAN +λ cycle *L cycle +λ P *L Perceptual

[0030] Where L GAN , L cycle and L Perceptual are the adversarial loss, the cycle consistency loss, and the cycle perception consistency loss respectively, and the parameters λ GAN , λ cycle and λ P are the weights that control the relative importance of each objective.

[0031] Furthermore, in step S3, the Adam optimization algorithm is used during the training process to minimize the total loss function to optimize the parameters of the generator and the discriminator.

[0032] The present invention also provides a forest fire smoke image generation system based on an improved CycleGAN, including a memory, a processor, and computer program instructions stored on the memory and capable of being run by the processor. When the processor runs the computer program instructions, the method steps as described above can be implemented.

[0033] Compared with the prior art, the present invention has the following beneficial effects: The present invention can simulate a clear forest graph, can learn the potential distribution law of fog without paired foggy - fog - free graphs, synthesize and output a foggy image corresponding to the input fog - free image, and the output foggy image must correspond pixel - by - pixel to the content of the original image (fog - free image) to ensure that the image pair can meet the model training and image restoration evaluation indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is the overall flowchart of the present invention;

[0035] Figure 2 are some typical cases in the smoke dataset;

[0036] Figure 3 is the network structure diagram of CycleGAN;

[0037] Figure 4It is the double-branch structure diagram of the improved generator A;

[0038] Figure 5 It is the foggy-fogless image result generated by the present invention. Specific implementation manners

[0039] The technical solution of the present invention will be specifically described below in conjunction with the accompanying drawings.

[0040] The present invention provides a method and system for generating forest fire smoke images based on an improved CycleGAN, including:

[0041] Step S1, obtaining a large number of random foggy-fogless random images as input images;

[0042] Step S2, preprocessing the obtained images, scaling the input pictures to 1.12 times and then randomly cropping them into a size of 256*256, and then randomly horizontally flipping them as slight image enhancement;

[0043] Step S3, changing the generator G in the generative adversarial network CycleGAN structure A to a double-branch structure, including a transmission rate learning subnet and an atmospheric light intensity learning subnet;

[0044] Step S4, introducing a dense residual block and a dual-path attention module to optimize feature reconstruction, and jointly training with adversarial loss, cycle consistency loss, and cycle perception consistency loss; by minimizing the total loss function, optimizing the parameters of the generator and the discriminator, so that the generator can generate images with a natural smoke effect from clear images;

[0045] Step S5, performing unsupervised learning using the improved generative adversarial network CycleGAN, training two generators and two discriminators, where:

[0046] Generator G A is used to generate images with smoke from clear images; Generator G B is used to recover clear images from images with smoke; Discriminator D A is used to judge the authenticity of the generated smoke images; Discriminator D B is used to judge the authenticity of the generated clear images;

[0047] Step S6, using the trained improved CycleGAN network model to perform fog generation processing on the clear images to be processed.

[0048] The following is the specific implementation process of the present invention.

[0049] Please refer to Figure 1, the present invention provides a method for generating forest fire smoke images based on improved CycleGAN, which generally includes obtaining and preprocessing dehazed images, constructing a training dataset, improving the CycleGAN generative adversarial network, training the model, and generating a paired hazy - haze - free image set.

[0050] Step S1: Obtain a large number of random hazy - haze - free images as input images. We used two datasets. The first is the smoke dataset we proposed, which includes 2088 unhazed pictures and 2303 hazy pictures for the forest scene, unpaired.

[0051] The second dataset is sourced from the NTIRE dehazing challenges over the years. We adopted O - Haze from the NTIRE2018 dehazing challenge, Dense - Haze from the NTIRE2019 dehazing challenge, NH - Haze from the NTIRE2020 dehazing challenge, and NH - Haze2 from the NTIRE2021 dehazing challenge.

[0052] Step S2: Preprocess the obtained images. Scale the input pictures by 1.12 times and then randomly crop them to a size of 256*256, and then randomly flip them horizontally as a slight image enhancement. The specific situation of dataset division is as follows: Randomly extract 200 pictures from the two image domains of the smoke dataset as our test set; O - Haze contains 35 pairs of outdoor haze images and ground truth images for training, and 10 pairs of images for testing; Dense - Haze and NH - Haze each contain 45 dense fog images and non - uniform fog images and their paired ground truth, and 10 pairs of images for testing; NH - Haze2 uses the first 20 pairs of images for training and the last 5 pairs for testing. We integrate these datasets and train them in a random sampling manner (unpaired), which includes 145 pairs of images for training and 35 pairs of images for testing.

[0053] Step S3: Change the generator GA in the CycleGAN network structure to a dual - branch structure, including a transmission rate learning subnet and an atmospheric light intensity learning subnet.

[0054] a) The first branch is the transmission rate learning subnet, inspired by the classic encoder - decoder structure. The transmission rate learning subnet predicts the transmission rate map by inputting the unhazed picture. The atmospheric scattering model formula can be expressed as:

[0055] I(x) = J(x)t(x)+A(1 - t(x))

[0056] t(x) = e -βd(x)

[0057] Where I and J represent the hazy image and the haze-free image respectively, and A represents the global atmospheric light intensity. t(x) represents the transmittance map, where β and d(x) are the atmospheric scattering parameter and the scene depth respectively.

[0058] It can be seen that, generally speaking, synthesizing a hazy image using the atmospheric scattering model requires predicting the depth information, that is, the depth map D, and a random atmospheric scattering parameter β to synthesize the transmittance map. However, our network can directly predict t(x) through learning for subsequent image synthesis.

[0059] The transmittance learning subnet consists of three parts: an encoder, a decoder, and a residual dense module. The encoding module consists of an initial layer and two downsampling layers. Therefore, the feature information of the input image can be encoded into the feature map. The residual dense module consists of 2 residual dense groups, and each residual dense group is composed of a residual dense block and a feature attention module, which can further extract more complex and different-scale features from the input features. In particular, since the purpose of this network is to generate a hazy image from a haze-free image, we can consider this as a process from simple to complex, and it should be relatively more concise compared to the dehazing network. So we only placed two residual dense groups. After feature extraction, the decoding module uses two upsampling operations and a Tanh activation function to output the final dehazed image. In addition, we set an average pooling layer, and the feature map output by the encoder is multiplied by the feature map output by the residual dense module after passing through the average pooling layer as the input of the decoder. We believe that feature extraction based on CNN will focus on the edges and shapes of the image, and such features are not conducive to the generation of non-uniform haze. Adding an average pooling layer can effectively solve this problem.

[0060] b) The second branch is the atmospheric light intensity learning subnet. The atmospheric light intensity learning subnet consists of three convolutional layers. The original image enters the first convolutional layer to initially extract features, and an unnecessary information in the feature space is eliminated through a global average pooling layer. Finally, after two convolutional layers with a kernel size of 1 and a sigmoid activation function, the final atmospheric light value A is obtained.

[0061] Finally, the hazy image is synthesized through the formula I(x) = J(x)t(x) + A(1 - t(x)), where t(x) is the transmittance map output by the transmittance learning subnet, A is the atmospheric light intensity output by the atmospheric light intensity learning subnet, J(x) is the original image (haze-free), and I(x) is the output hazy image. Since the output haze-free is synthesized through a physical formula and there is a strong physical connection with the original image, compared with directly generating an image through a neural network, the image generated by our model will not change the shape and details of the background objects.

[0062] Step S4: During the training process, a dense residual block and a dual-path attention module are introduced to optimize feature reconstruction, and adversarial loss, cycle consistency loss, and cycle-aware consistency loss are used for joint training; by minimizing the total loss function, the parameters of the generator and discriminator are optimized, enabling the generator to generate images with natural smoke effects from clear images.

[0063] a) Adversarial loss and cycle consistency loss

[0064] We use adversarial loss in the model. For the mapping function G A : X → Y and its discriminator D B , we express the objective as

[0065] L GAN (G A , D B ) = E X～Pdata(X) [log(1 - D B (G A (X)))] + E Y～Pdata(Y) [logD B (Y)]

[0066] L GAN (G A , D B ) is the adversarial loss; E X～Pdata(X) represents the expected value of the cycle consistency loss sampled in the X domain (smokeless images); E Y～Pdata(Y) represents the expected value of the cycle consistency loss sampled in the Y domain (smoky images).

[0067] For each generated fake image G A (X) → Y’, it must be able to be restored to the original image through the mapping function G B : Y → X, i.e., G B (G A (X)) = X r ≈ X, where X r represents the image restored after two mappings by the generators G A and G B , i.e., the smokeless image restored by G B (G A (X)); the constraint between them is called the cycle consistency loss, expressed by the formula:

[0068] L cycle (G A , D B ) = E X～Pdata(X) [||G B (G A (X)) - X||1] + E Y～Pdata(Y) [||GA (G B (Y)) - Y || 1]

[0069] L cycle (G A , D B ) represents the cycle consistency loss; E X～Pdata(X) represents the expected value of the cycle consistency loss sampled on the X domain (smokeless images), so that CycleGAN will not damage the smokeless images and ensure that the original image can be restored after dehazing; E Y～Pdata(Y) represents the expected value of the cycle consistency loss sampled on the Y domain (smoky images), so that CycleGAN will not damage the smoky images and ensure that the original image can be restored after adding back the smoke.

[0070] b) Cyclic perceptual - consistency loss

[0071] We used Cyclic perceptual - consistency loss in the model. Assume that after the cycle G B (G A (X)) = X r , G A (G B (Y)) = Y r . X, X r and Y, Y r should have structural similarity in features at high and low levels, where X r represents the image restored after being mapped twice by the generators G A and G B , that is, the smokeless image restored by G B (G A (X)); Y r represents the image restored after being mapped twice by the generators G B and G A , that is, the smoky image restored by G A (G B (Y)); The cyclic perceptual - consistency loss L Perceptual uses a pre - trained VGG19 model, outputs the features extracted from the second pooling layer and the fifth pooling layer, and through the L2 - norm contrast loss, the formula is as follows:

[0072]

[0073] where represents the features extracted by the VGG19 model from the second pooling layer and the fifth pooling layer; Denote the features obtained from the second pooling layer and the fifth pooling layer of the original smokeless image; Denote the features obtained from the second pooling layer and the fifth pooling layer of the image restored through two mappings; Denote the features obtained from the second pooling layer and the fifth pooling layer of the original smoky image; Denote the features obtained from the second pooling layer and the fifth pooling layer of the image restored through two mappings.

[0074] The objective of Cycle-Dehaze can be expressed as follows:

[0075] L = λ GAN *L GAN + λ cycle *L cycle + λ P *L Perceptual

[0076] where L GAN 、L cycle and L Perceptual are the adversarial loss, the cycle consistency loss, and the cycle perception consistency loss respectively. The parameters λ GAN 、λ cycle and λ P are the weights that control the relative importance of each objective respectively.

[0077] Step S5: Use the improved generative adversarial network (CycleGAN) for unsupervised learning to train two generators and two discriminators, where:

[0078] By default, β1 = 0.5, β2 = 0.999, and the initial learning rate is 0.0002. The hyperparameters λ GAN , λ cycle , λ P of the loss function are 0.5, 5, and 2.5 respectively. All experiments are conducted on two Nvidia 3090 24G GPUs. The model is trained for 60 epochs on the smoke dataset and linearly reduces the learning rate to 0 starting from the 40th epoch. Since the number of NTIRE images is small, we set the epoch to 100 on the NTIRE dataset and linearly reduce the learning rate to 0 starting from the 60th epoch.

[0079] Step S6: Use the trained improved CycleGAN network model to perform fog generation processing on the clear images to be processed.

[0080] The present invention also provides a forest fire smoke image generation system based on an improved CycleGAN, which includes a memory, a processor, and computer program instructions stored on the memory and capable of being run by the processor. When the processor runs the computer program instructions, the method steps described in any one of the above can be implemented.

[0081] The above are the preferred embodiments of the present invention. All changes made according to the technical solution of the present invention, when the functional effects produced do not exceed the scope of the technical solution of the present invention, fall within the protection scope of the present invention.

Claims

1. A method and system for generating forest fire smoke images based on an improved CycleGAN, characterized in that Including: Step S1: Obtain a large number of random foggy - fog - free random images as input images; Step S2: Pre - process the obtained images. Scale the input pictures by 1.12 times and then randomly crop them into the size of 256*256, and then randomly horizontally flip them as a slight image enhancement; Step S3: Modify the generator G in the CycleGAN structure of the generative adversarial network A into a dual-branch structure, including a transmission rate learning subnet and an atmospheric light intensity learning subnet; Step S4: Introduce a dense residual block and a dual - path attention module to optimize feature reconstruction, and jointly train with adversarial loss, cycle - consistency loss, and cycle - perceptual consistency loss; By minimizing the total loss function, optimize the parameters of the generator and discriminator, so that the generator can generate images with natural smoke effects from clear images; Step S5: Use the improved generative adversarial network CycleGAN for unsupervised learning, and train two generators and two discriminators, where: Generator G A used to generate images with smoke from clear images; Generator G B used to recover a clear image from an image with smoke; Discriminator D A used to judge the authenticity of the generated smoke image; Discriminator D B used to judge the authenticity of the generated clear image; Step S6: Use the trained improved CycleGAN network model to perform fog generation processing on the clear images to be processed.

2. The method for generating forest fire smoke images based on an improved CycleGAN according to claim 1, characterized in that, In step S3, the transmission rate learning subnet adopts an encoder - decoder architecture. The transmission rate learning subnet consists of three parts: an encoder, a decoder, and a residual dense module. The encoder consists of an initial layer and two down - sampling layers. The residual dense module consists of 2 residual dense groups. The residual dense group is combined by a residual dense block and a feature attention module. The decoder uses two up - sampling operations and a Tanh activation function.

3. A method for generating forest fire smoke images based on an improved CycleGAN according to claim 2, wherein, The transmission rate learning subnet also sets an average pooling layer. Multiply the feature map output by the encoder after passing through the average pooling layer with the feature map output by the residual dense module as the input of the decoder.

4. A method for generating forest fire smoke images based on an improved CycleGAN according to claim 1, characterized in that In step S3, the atmospheric light intensity learning subnet consists of three convolutional layers. The input image enters the first convolutional layer to initially extract features, eliminates unnecessary information in the feature space through a global average pooling layer, and then passes through two convolutional layers with a convolutional kernel size of 1 and a sigmoid activation function to obtain the final atmospheric light value A.

5. A method for generating forest fire smoke images based on an improved CycleGAN according to claim 1, characterized in that, In step S3, according to the atmospheric scattering model formula I(x) = J(x)t(x)+A(1 - t(x)), synthesize a foggy image, where t(x) is the transmission rate map output by the transmission rate learning subnet, A is the atmospheric light intensity output by the atmospheric light intensity learning subnet, J(x) is the input image, and I(x) is the output foggy image.

6. A method for generating forest fire smoke images based on an improved CycleGAN according to claim 2, characterized in that In step S4, the adversarial loss and cycle - consistency loss are as follows: For the mapping function G A : X → Y and its discriminator D B , the objective is expressed as: L GAN (G A ,D B ) = E X～Pdata(X) [log(1 - D B (G A (X))))] + E Y～Pdata(Y) [logD B (Y)] L GAN (G A ,D B ) i.e., the adversarial loss; E X～Pdata(X) represents the expected value of the cycle-consistency loss for upsampling in the X domain, i.e., non-smoky images; E Y～Pdata(Y) represents the expected value of the cycle-consistency loss for upsampling in the Y domain, i.e., smoky images; For each generated fake image G A (X) → Y’, it must be possible to restore it to the original image through the mapping function G B : Y → X, i.e., G B (G A (X)) = X r ≈ X, where X r represents the image restored after two mappings by the generators G A and G B , i.e., the smokeless image restored through G B (G A (X)); the constraint between them is called the cycle consistency loss, which is expressed by the formula as: L cycle (G A ,D B ) = E X～Pdata(X) [||G B (G A (X)) - X||1] + E Y～Pdata(Y) [||G A (G B (Y)) - Y||1] L cycle (G A ,D B ) represents the cyclic consistency loss; E X～Pdata(X) represents the expected value of the cyclic consistency loss for upsampling in the X domain, i.e., the smokeless images; E Y～Pdata(Y) represents the expected value of the cyclic consistency loss for upsampling in the Y domain, i.e., the smoky images.

7. A method for generating forest fire smoke images based on an improved CycleGAN according to claim 6, characterized in that In step S4, the cycle - perceptual consistency loss is as follows: Assume after cycle G B (G A (X)) = X r , G A (G B (Y)) = Y r ; X, X r and Y, Y r should have structural similarity in features at high and low levels, where X r represents the image restored after being mapped twice by generators G A and G B , that is, the smokeless image restored by G B (G A (X)); Y r represents the image restored after being mapped twice by generators G B and G A , that is, the smoky image restored by G A (G B (Y)); The cycle perception consistency loss L Perceptual uses a pre-trained VGG19 model to output the features extracted by the second pooling layer and the fifth pooling layer. Through the L2 norm comparison loss, the formula is as follows: Among them represents the features extracted by the VGG19 model at the second pooling layer and the fifth pooling layer; represents the features obtained at the second pooling layer and the fifth pooling layer of the original smokeless image; represents the features obtained at the second pooling layer and the fifth pooling layer of the image restored after two mappings; represents the features obtained at the second pooling layer and the fifth pooling layer of the original smoky image; represents the features obtained at the second pooling layer and the fifth pooling layer of the image restored after two mappings.

8. A method for generating forest fire smoke images based on improved CycleGAN according to claim 7, characterized in that The total loss function is expressed as follows: L = λ GAN *L GAN + λ cycle *L cycle + λ P *L Perceptual Where L GAN , L cycle and L Perceptual are the adversarial loss, the cycle consistency loss, and the cycle perception consistency loss respectively, and the parameters λ GAN , λ cycle and λ P are the weights that control the relative importance of each objective respectively.

9. A method for generating forest fire smoke images based on an improved CycleGAN according to claim 1, characterized in that In step S3, the Adam optimization algorithm is used during the training process to minimize the total loss function to optimize the parameters of the generator and discriminator.

10. A forest fire smoke image generation system based on improved CycleGAN, characterized in that, Including a memory, a processor, and computer program instructions stored on the memory and capable of being run by the processor. When the processor runs the computer program instructions, it can implement the method steps described in any one of claims 1 - 9.

Images