Remote sensing image tiny target identification method and system under haze climate condition

Abstract

The invention provides a remote sensing image tiny target identification method and system under a haze climate condition, and the method comprises the steps: obtaining a plurality of remote sensing sample images, adding random noise to each remote sensing sample image, and obtaining a simulation sample image; performing denoising processing on the simulation sample image to obtain a restored image, and constructing restoration loss according to the difference between the restored image and the corresponding remote sensing sample image; performing data enhancement on the restored image to obtain a training sample image; calculating the centroid of each target to be recognized in the training sample image, and performing comparative learning based on the centroid to construct a comparative loss; and training a detection network based on the restoration loss and the comparison loss to obtain a target detection model, and identifying a to-be-identified target in the to-be-identified image through the target detection model to obtain an identification result. According to the method and the device, the tiny target in the haze scene can be identified, and the identification accuracy is effectively guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular, to a method and system for identifying small targets in remote sensing images under haze climate conditions. Background Art

[0002] Detecting small targets in remote sensing images facing haze climate is of great significance for environmental monitoring and disaster warning. Haze weather will reduce the visibility of remote sensing images, especially increasing the difficulty of identifying small-sized targets significantly. The related technologies have relatively low recognition accuracy for small targets in scenes with low visibility, thus affecting the application of remote sensing technology in fields such as pollution source monitoring, forest fire warning, and urban construction. Summary of the Invention

[0003] In view of the problems existing in the above prior art, the present invention proposes a method and system for identifying small targets in remote sensing images under haze climate conditions, mainly solving the problem that the related technologies cannot accurately identify small targets in a haze environment.

[0004] To achieve the above object and other objects, the technical solutions adopted by the present invention are as follows.

[0005] The present application provides a method for identifying small targets in remote sensing images under haze climate conditions. The method includes: obtaining a plurality of remote sensing sample images, and adding random noise to each of the remote sensing sample images to simulate imaging under a haze environment, obtaining simulated sample images; performing denoising processing on the simulated sample images to obtain restored images, and constructing a restoration loss based on the difference between the restored images and the corresponding remote sensing sample images to adjust the prediction parameters of the denoiser; performing data augmentation on the restored images to obtain training sample images; calculating the centroids of each target to be recognized in the training sample images, and performing contrast learning based on the centroids to construct a contrast loss; training a detection network based on the restoration loss and the contrast loss to obtain a target detection model, so as to identify the target to be recognized in the image to be recognized through the target detection model and obtain a recognition result.

[0006] In an embodiment of the present application, the step of adding random noise to each of the remote sensing sample images includes: generating a random haze coefficient, where the haze coefficient is used to characterize the concentration of haze; adjusting the scattering coefficient of the remote sensing sample image based on the haze coefficient to generate the simulated sample images.

[0007] In an embodiment of the present application, the step of performing denoising processing on the simulated sample images includes: performing denoising processing using a pre-constructed denoiser, where the denoiser includes a haze filter for filtering haze noise, a super-resolution optimizer for restoring an image with a resolution lower than a threshold to an image higher than the threshold, and a sharpening filter for filtering Gaussian noise.

[0008] In an embodiment of the present application, the steps of performing data augmentation on the restored image include: scaling the restored image within a preset scaling range to obtain a plurality of first images with different scales; randomly flipping the first image horizontally or vertically with a preset probability to obtain a second image; performing a normalization operation on the second image to scale the pixel values of the second image to a preset range to obtain a third image; and padding the third image so that the size of the padded image is divisible by 32, thereby obtaining the training sample image.

[0009] In an embodiment of the present application, the steps of calculating the centroids of each target to be recognized in the training sample image and performing contrast learning based on the centroids to construct a contrast loss include: dynamically sampling the training sample image, generating corresponding centroids according to the categories of the targets to be recognized in the sampling region, and assigning attention weights to each centroid so that the smaller the target to be recognized, the higher the corresponding attention weight; storing the centroids of different categories in a preset memory bank for regional perception contrast learning, and further obtaining the contrast loss between the centroids of different categories.

[0010] In an embodiment of the present application, the steps of training the detection network based on the restoration loss and the contrast loss include: constructing a loss function through the restoration loss, the contrast loss, and the classification regression loss; adjusting the network parameters of the detection network based on the loss function until the loss value of the loss function reaches a preset threshold to obtain the target detection model.

[0011] In an embodiment of the present application, the steps of constructing a restoration loss according to the difference between the restored image and the corresponding remote sensing sample image to adjust the prediction parameters of the denoiser include: using the restoration loss as the loss of the time-effective neural network to pre-train the time-effective neural network to obtain the denoiser.

[0012] The present application also provides a remote sensing image micro-target recognition system under haze climate conditions. The system includes: an image acquisition module, configured to obtain a plurality of remote sensing sample images, and add random noise to each of the remote sensing sample images to simulate imaging under a haze environment, so as to obtain simulated sample images; a denoising module, configured to perform denoising processing on the simulated sample images to obtain restored images, and construct a restoration loss according to the difference between the restored images and the corresponding remote sensing sample images to adjust the prediction parameters of the denoiser; a data enhancement module, configured to perform data enhancement on the restored images to obtain training sample images; a contrast learning module, configured to calculate the centroids of each target to be recognized in the training sample images, and perform contrast learning based on the centroids to construct a contrast loss; a model training and recognition module, configured to train a detection network based on the restoration loss and the contrast loss to obtain a target detection model, so as to recognize the target to be recognized in the image to be recognized through the target detection model to obtain a recognition result.

[0013] As described above, a remote sensing image micro-target recognition method and system provided by the present application have the following beneficial effects.

[0014] Images under various different haze concentration scenarios can be simulated by randomly adding noise, improving the diversity of sample images; the prediction parameters of the denoiser can be guaranteed by optimizing the denoiser based on the restoration loss; the model can learn the differences between centroids of different categories and enhance the model classification prediction effect by training the model based on the restoration loss and the contrast loss. The present application can recognize micro-targets in a haze scenario, effectively guaranteeing the accuracy of recognition. Description of the Drawings

[0015] Figure 1 It is a schematic flowchart of a remote sensing image micro-target recognition method under haze climate conditions in an embodiment of the present application.

[0016] Figure 2 It is a schematic architecture diagram of optimizing the parameters of a denoiser based on a time-efficient neural network in an embodiment of the present application.

[0017] Figure 3 It is a schematic overall architecture diagram of a detection network in an embodiment of the present application.

[0018] Figure 4 It is a module diagram of a remote sensing image micro-target recognition system under haze climate conditions in an embodiment of the present application. Detailed Embodiments

[0019] The following describes the implementation modes of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation modes. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0020] It should be noted that the illustrations provided in the following embodiments only schematically illustrate the basic concept of the present invention. Therefore, only the components related to the present invention are shown in the drawings, rather than being drawn according to the number, shape, and size of the components in actual implementation. The types, quantities, and proportions of the components in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0021] The inventor has found through research that: with the development of deep learning technology, small targets can improve the recognition accuracy of small targets in the environment by detecting small targets in remote sensing images under haze conditions, ensuring that key targets such as pollution sources, buildings, and transportation facilities can still be accurately identified even in low visibility environments, thereby providing reliable data support for environmental monitoring and disaster warning. Remote sensing target detection can not only provide the recognition of static targets, but also monitor dynamic targets, further enhancing the response ability to emergencies. In intelligent monitoring and environmental protection systems, the small target detection technology under haze conditions has become one of the key technologies to improve data accuracy, real-time performance, and system reliability. With the continuous development of deep learning and artificial intelligence technologies, these detection technologies can be combined with systems such as real-time data analysis and drone patrols to promote the development of remote sensing monitoring towards intelligence and automation, further enhancing environmental management and emergency response capabilities. Although the object detection method based on deep learning has achieved gratifying results in remote sensing images, it is a very challenging task to accurately identify small targets from low-quality images under the haze weather conditions in complex climates. Although the existing object detection methods based on deep learning have made progress on standard datasets, they still face the balance problem in the co-optimization of image denoising and object detection.

[0022] In addition, these methods often fail to fully utilize the potential information that helps improve detection accuracy. To alleviate this problem, the present invention proposes a method for identifying tiny targets in remote sensing images under haze climate conditions, providing technical support for the identification of tiny targets in remote sensing images and environmental detection security under haze climate conditions. This method proposes an adaptive remote sensing image denoising technique. Through defogging processing and super-resolution reconstruction, a clear and high-resolution remote sensing image is generated. At the same time, the time-effective memory network is used to adaptively adjust the parameters of the denoiser to further improve the denoising effect. The enhanced image is scaled and randomly inverted to obtain more diverse training images. These images contain multi-level feature information of the same remote sensing image at different scales and orientations, enhancing the generalization ability of the model. Then, a dynamic region perception module is proposed. This module can generate centroids for the same type of tiny targets and focus more attention on difficult samples through a dynamic sampling strategy. By generating centroids for different categories, the module uses the region perception contrast learning method to increase the inter-class interval and effectively reduce the intra-class difference, thereby optimizing the classification performance of the targets. Finally, by combining the restoration loss, dynamic contrast loss, and classification regression loss functions, the entire detection network is trained to further improve the detection accuracy and ensure the stability and reliability of the model in the haze environment.

[0023] The technical solution of the present application will be elaborated in detail below in conjunction with specific embodiments.

[0024] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of the method for identifying tiny targets in remote sensing images under haze climate conditions in an embodiment of the present application. The method provided by the embodiment of the present application includes:

[0025] Step S100, obtain a plurality of remote sensing sample images, and add random noise to each of the remote sensing sample images to simulate imaging under a haze environment, obtaining simulated sample images.

[0026] In one embodiment, high-altitude images of the corresponding scene can be obtained through mobile shooting devices such as drones. The scene here can be an airport, a school, a forest, etc. The target to be identified and the shooting height included in the specific scene can be determined according to actual needs and are not limited here. The initially obtained remote sensing sample images are clear images of the corresponding scene. Based on the obtained clear images, random noise is added to simulate the imaging effect of the corresponding scene under a haze environment. The specific number of remote sensing sample images can also be set and adjusted according to the actual needs of model training and is not limited here.

[0027] In one embodiment, the step of adding random noise to each of the remote sensing sample images includes: generating a random haze coefficient, where the haze coefficient is used to characterize the concentration of haze; and adjusting the scattering coefficient of the remote sensing sample image based on the haze coefficient to generate the simulated sample image. Specifically, the haze coefficient can be set in advance, or a random number within a set range can be generated through a random number generation algorithm as the set value of the haze coefficient. The formulas for calculating the scattering coefficient based on the haze coefficient and adjusting the clear image based on the scattering coefficient can be expressed as follows:

[0028]

[0029] I(x) = kJ(x)t0 + A(1 - kt0(x)) (2)

[0030] Where: β0 is the initial atmospheric scattering coefficient; t0(x) is the initial transmittance; α is the haze coefficient, and the fog concentration is controlled by randomly setting the size of the haze coefficient. I(x) is the foggy image (i.e., the simulated sample image); J(x) is the fog-free image (i.e., the remote sensing sample image); A is the atmospheric light intensity; d(x) represents the scene depth, and k is a hyperparameter used to further adjust the degree of haze.

[0031] Adding random noise based on the above formula can make different simulated sample images have different degrees of noise, thereby simulating the imaging effects under different haze concentrations.

[0032] Step S110: Denoise the simulated sample image to obtain a restored image, and construct a restoration loss based on the difference between the restored image and the corresponding remote sensing sample image to adjust the prediction parameters of the denoiser.

[0033] In one embodiment, the simulated sample image can be denoised by a denoiser. The denoiser can be pre-trained based on a time-effective memory neural network. The denoiser can include a haze filter for filtering haze noise, a super-resolution optimizer for restoring an image with a resolution lower than a threshold to be higher than the threshold, and a sharpening filter for filtering Gaussian noise. Specifically, the dehazing filter can be expressed as:

[0034]

[0035] Where: ω is a hyperparameter used to adjust the dehazing filter; C represents the color channel, Ω(x) represents the local neighborhood window (local patch) centered on pixel x, which can also be called the neighborhood range or local receptive field, A c represents the value of the atmospheric light intensity on channel C. I c (y) represents the pixel value of the image at position y and channel C.

[0036] The super-resolution optimizer can be expressed as:

[0037]

[0038] Wherein: To restore the high-resolution image, H represents the downsampling operation, and R(I HR ) is the regularization operation on I HR , I HR is the input low-resolution image, α is the regularization parameter; I LR represents the observed low-resolution image, and the low-resolution image is an image with a resolution lower than a threshold. The specific threshold can be set and adjusted according to actual application requirements and is not limited here.

[0039] The sharpening filter can be expressed as:

[0040] F(x,λ) = I(x) + λ(I(x) - Gau(I(x))) (5)

[0041] Wherein: I(x) is the input image, Gau(I(x)) represents the Gaussian filter, and λ is a positive scaling factor.

[0042] Please refer to Figure 2 , Figure 2 , which is a schematic diagram of the architecture for optimizing the parameters of the denoiser based on the aging neural network in an embodiment of the present application. The parameters of each component in the denoiser can be adjusted through the aging neural network, so that the image processed by the denoiser has a sufficiently high quality. Exemplarily, a restoration loss can be constructed based on the difference between the restored image after denoising and the corresponding remote sensing sample image before denoising, and the aging memory neural network can be trained based on the restoration loss, and then the parameters of each component in the denoiser can be adjusted to obtain a denoiser that can achieve the expected effect. The specific architecture and training process of the aging neural network are well-known in the art and will not be elaborated here. The specific calculation method of the restoration loss can be expressed as:

[0043]

[0044] Wherein: I ij is the original image, and C ij is the restored image.

[0045] Step S120, perform data augmentation on the restored image to obtain a training sample image.

[0046] In one embodiment, data augmentation can be performed on the restored image in multiple dimensions such as scale, quantity, orientation, etc., thereby increasing the diversity of training samples. The steps of performing data augmentation on the restored image include: scaling the restored image within a preset scaling range to obtain first images with multiple different scales; randomly flipping the first images horizontally or vertically according to a preset probability to obtain second images; performing a normalization operation on the second images to scale the pixel values of the second images to a preset range to obtain third images; padding the third images to make the size of the padded images divisible by 32, thereby obtaining the training sample images. Specifically, perform a scaling operation on high-quality image patches, with the scaling range from 0.6 to 1.6, thereby obtaining first images with different scales; perform a random flip on the images with different scales horizontally or vertically with a probability of 30% to obtain second images with different scales that have been randomly flipped; then perform a normalization operation on the second images after flipping with different scales to scale the pixel values of the images to a specific range for better subsequent model training or inference. The foregoing scaling range and flip probability are only examples and can be adjusted according to actual application requirements, which are not limited here. The calculation formula for normalization can be expressed as follows:

[0047]

[0048] where: μ represents the image mean, X represents the image matrix, σ represents the standard deviation, and N represents the number of pixels of the image X. After completing the normalization, pad the images that do not meet the requirements so that the size of the padded images is divisible by 32 for subsequent calculations.

[0049] Step S130, calculate the centroids of each target to be recognized in the training sample image, and perform contrastive learning based on the centroids to construct a contrastive loss.

[0050] In one embodiment, the steps of calculating the centroids of each target to be recognized in the training sample image and performing contrastive learning based on the centroids to construct a contrastive loss include: dynamically sampling the training sample image, generating corresponding centroids according to the categories of the targets to be recognized in the sampling region, and assigning attention weights to each centroid such that the smaller the target to be recognized, the higher the corresponding attention weight; storing the centroids of different categories in a preset memory bank for region-aware contrastive learning, thereby obtaining the contrastive loss between the centroids of different categories.

[0051] In one embodiment, a dynamic region awareness module can be set up. This dynamic region awareness module performs dynamic contrast learning on the enhanced image, calculates the corresponding contrast loss, and adds the obtained inter-class pair loss to the model training process. The contrast loss can prompt the model to learn more discriminative feature representations, which specifically includes the following steps: By means of dynamic sampling, more attention is allocated to the tiny and difficult-to-recognize targets, and the same centroid representation is generated for the to-be-recognized targets of the same class, and different centroid representations are generated for the to-be-recognized targets of different classes. The centroids of different classes are pushed into the memory bank for region awareness contrast learning, thereby constructing the contrast loss.

[0052]

[0053] Among them, q represents the query vector, and k + represents the positive sample, and k - represents the negative sample, τ represents the temperature coefficient, which controls the smoothness of the softmax distribution, and M i is the set of all positive samples corresponding to the current i-th sample.

[0054] Step S140, based on the restoration loss and the contrast loss, train the detection network to obtain a target detection model, so as to identify the to-be-recognized target in the to-be-recognized image through the target detection model and obtain the recognition result.

[0055] In one embodiment, a loss function can be constructed by combining the restoration loss, the contrast loss, and the classification regression loss. Exemplarily, the loss function can be the sum of the three, and the classification regression loss can adopt the cross-entropy loss function. Train the detection network based on the loss function to obtain the corresponding target detection model. The recognition accuracy of the target detection model obtained through the foregoing process for tiny targets in images under a haze environment can be effectively improved.

[0056] Please refer to Figure 3 , Figure 3 which is the overall architecture diagram of the detection network in an embodiment of the present application. The detection network can extract features (feature extraction) from the training sample images through the CSPDarkNet+SPPF network, can extract the to-be-recognized targets in the images by means of dynamic sampling, allocate higher weights to the tiny targets through the attention head mechanism (detect head), and at the same time calculate the centroids and push the centroids of different classes into the memory banks (memory banks) for dynamic region awareness contrast learning, which can enable the model to better learn discriminative feature representations and improve its performance in the task of recognizing tiny targets in remote sensing images under a complex haze environment.

[0057] Based on the technical solution of the embodiment of the present application, an unmanned aerial vehicle device is used to move and capture high-altitude images to obtain images for micro-target recognition. These images may be affected by severe haze weather, or it is relatively difficult to obtain images under real climate conditions. In order for the model to further process haze weather conditions in different complex climate scenarios and then simulate the complex environment in the real scenario. In order to simulate severe haze weather under complex climate, a haze simulation algorithm is designed. This algorithm involves adjusting the scattering coefficient and haze coefficient, and adding different degrees of haze to clear images in a random manner. Randomly adjust the parameters of haze according to the requirements of a specific scenario, such as controlling the haze concentration; in order to reduce the influence of noise on the detection accuracy, input these images with noise into an adaptive denoising module. This module includes a defogging filter, a super-resolution optimizer, and a sharpening filter. By adjusting the parameters of these filters through a time-effective neural network module that predicts the specific parameters of the filters, noise can be effectively reduced and the quality of the images can be improved. Next, a comparison calculation is performed between the original image and the denoised image to obtain the restoration loss. By comparing the differences between the two, the influence degrees of the simulated noise and the denoising process on the images can be understood. Finally, according to the restoration loss, adjust the parameters of the filter predictor and the parameters of the detection model so that the model can better adapt to the remote sensing micro-target recognition task under complex haze environment; in order to enrich the diversity of data, perform a scaling operation on the obtained high-quality remote sensing image patches, and the scaling range is from 0.6 to 1.6 to obtain the images to be detected with different scales. Perform a random flip in the horizontal or vertical direction with a probability of 30% on the images with different scales to obtain different scales of images after random flipping. Then, perform a normalization operation on the images with different scales after flipping, and scale the pixel values of the images to a specific range for better subsequent model training or inference. Finally, perform a padding operation on the normalized images so that the size of the enhanced images can be divisible by 32 for subsequent calculations; the dynamic region perception module uses a dynamic sampling method to allocate more attention to difficult micro-target samples and generate centroids of the same category to ensure that the model can better learn the features of difficult samples. Then, the comparison template is the centroids of different categories so that the model can learn the feature differences between different categories from them. And generate the same centroid for the same category in the same image to ensure that the model can learn the feature representations of different regions in the same image. Push the centroids of different region categories into the memory bank for dynamic region perception contrast learning. Through this series of steps, the model can better learn discriminative feature representations and improve its performance in the remote sensing image micro-target recognition task under complex haze environment.

[0058] Please refer to Figure 4 , Figure 4It is a block diagram of a remote sensing image micro-target recognition system under haze climate conditions in an embodiment of the present application. An embodiment of the present application also provides a remote sensing image micro-target recognition system under haze climate conditions, and the system includes: an image acquisition module 40, configured to acquire a plurality of remote sensing sample images, and add random noise to each of the remote sensing sample images to simulate imaging under a haze environment, so as to obtain simulated sample images; a denoising module 41, configured to perform denoising processing on the simulated sample images to obtain restored images, and construct a restoration loss according to the difference between the restored images and the corresponding remote sensing sample images to adjust the prediction parameters of the denoiser; a data augmentation module 42, configured to perform data augmentation on the restored images to obtain training sample images; a contrast learning module 43, configured to calculate the centroids of each target to be recognized in the training sample images, and perform contrast learning based on the centroids to construct a contrast loss; a model training and recognition module 44, configured to train a detection network based on the restoration loss and the contrast loss to obtain a target detection model, so as to recognize the target to be recognized in the image to be recognized through the target detection model and obtain a recognition result.

[0059] The execution process of the specific system has been elaborated in detail in the foregoing method embodiment and will not be repeated here.

[0060] The above embodiments are only illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A method for identifying small targets in remote sensing images under haze climate conditions, characterized in that, The method includes: Obtaining a plurality of remote sensing sample images, and adding random noise to each of the remote sensing sample images to simulate imaging in a haze environment, thereby obtaining simulated sample images; Performing denoising processing on the simulated sample images to obtain restored images, and constructing a restoration loss based on the difference between the restored images and the corresponding remote sensing sample images to adjust the prediction parameters of the denoiser; Performing data augmentation on the restored images to obtain training sample images; Calculating the centroids of each target to be recognized in the training sample images, and performing contrastive learning based on the centroids to construct a contrastive loss; Training a detection network based on the restoration loss and the contrastive loss to obtain an object detection model, so as to identify the objects to be recognized in the image to be recognized through the object detection model and obtain a recognition result.

2. The method for identifying small targets in remote sensing images under haze climate conditions according to claim 1, wherein The step of adding random noise to each of the remote sensing sample images includes: Generating a random haze coefficient, where the haze coefficient is used to characterize the concentration of haze; Adjusting the scattering coefficient of the remote sensing sample image based on the haze coefficient to generate the simulated sample image.

3. The method for identifying tiny targets in remote sensing images under haze climate conditions according to claim 1, wherein, The step of performing denoising processing on the simulated sample images includes: performing denoising processing using a pre-constructed denoiser, where the denoiser includes a haze filter for filtering haze noise, a super-resolution optimizer for restoring an image with a resolution lower than a threshold to be higher than the threshold, and a sharpening filter for filtering Gaussian noise.

4. The method for identifying tiny targets in remote sensing images under haze climate conditions according to claim 1, wherein, The step of performing data augmentation on the restored images includes: Scaling the restored images within a preset scaling range to obtain a plurality of first images with different scales; Randomly flipping the first images horizontally or vertically according to a preset probability to obtain second images; Performing a normalization operation on the second images to scale the pixel values of the second images to a preset range to obtain third images; Performing image padding on the third images so that the size of the padded images can be divisible by 32, thereby obtaining the training sample images.

5. The method for identifying tiny targets in remote sensing images under haze climate conditions according to claim 1, characterized in that, The step of calculating the centroids of each target to be recognized in the training sample images and performing contrastive learning based on the centroids to construct a contrastive loss includes: Performing dynamic sampling on the training sample images, generating corresponding centroids according to the categories of the targets to be recognized in the sampling regions, and assigning attention weights to each centroid, such that the smaller the target to be recognized, the higher the corresponding attention weight; Storing the centroids of different categories in a preset memory bank for region-aware contrastive learning, and thereby obtaining the contrastive loss between the centroids of different categories.

6. The method for identifying small targets in remote sensing images under haze climate conditions according to claim 1, characterized in that, The step of training a detection network based on the restoration loss and the contrastive loss includes: Constructing a loss function through the restoration loss, the contrastive loss, and a classification regression loss; Adjusting the network parameters of the detection network based on the loss function until the loss value of the loss function reaches a preset threshold to obtain the object detection model.

7. The method for identifying tiny targets in remote sensing images under haze climate conditions according to claim 1, characterized in that, The step of constructing a restoration loss based on the difference between the restored images and the corresponding remote sensing sample images to adjust the prediction parameters of the denoiser includes: Taking the restoration loss as the loss of a temporal neural network to pre-train the temporal neural network to obtain the denoiser.

8. A remote sensing image micro-target recognition system under haze climate conditions, characterized in that, The system includes: An image acquisition module, configured to obtain a plurality of remote sensing sample images, and add random noise to each of the remote sensing sample images to simulate imaging in a haze environment, so as to obtain simulated sample images; A denoising module, configured to perform denoising processing on the simulated sample images to obtain restored images, and construct a restoration loss based on the differences between the restored images and the corresponding remote sensing sample images to adjust the prediction parameters of the denoiser; A data augmentation module, configured to perform data augmentation on the restored images to obtain training sample images; A contrastive learning module, configured to calculate the centroids of each target to be recognized in the training sample images, and perform contrastive learning based on the centroids to construct a contrastive loss; A model training and recognition module, configured to train a detection network based on the restoration loss and the contrastive loss to obtain a target detection model, so as to recognize the target to be recognized in the image to be recognized through the target detection model to obtain a recognition result.

Images