Self-supervised optical remote sensing image blind denoising method and system

Abstract

The present application relates to the technical field of image denoising, in particular to a self-supervised optical remote sensing image blind denoising method and system, wherein the method comprises: obtaining a remote sensing image to be denoised; performing non-local similar sampling on the remote sensing image to be denoised to obtain a first noise space-independent similar image; performing neighborhood Bernoulli sampling on the first noise space-independent similar image to obtain a first image pair set; and inputting the first image pair set into a trained denoising model to obtain a denoised remote sensing image. The present application has good processing effect on spatially correlated noise.

Description

Technical Field

[0001] This invention relates to the field of image denoising technology, and in particular to a self-supervised blind denoising method and system for optical remote sensing images. Background Technology

[0002] In the field of computer vision, digital image denoising is a crucial step in image preprocessing, and its quality plays a decisive role in the accuracy of downstream tasks. High-quality denoising provides a solid foundation for subsequent image analysis and processing. In recent years, the rapid development of deep learning has brought new solutions to the field of image denoising, and deep learning-based image denoising techniques have begun to be widely used in the field of optical remote sensing images.

[0003] Optical remote sensing images, as a crucial source of Earth's surface information, encompass rich topographic features and information on various ground objects, including mountains, forests, buildings, and roads. Downstream analysis and processing tasks, such as topographic mapping and land use classification, place extremely high demands on the image's detailed texture clarity. Compared to natural image denoising, optical remote sensing image denoising not only requires noise removal but also maximizing the restoration of the image's detailed texture to meet the needs of high-precision analysis. However, due to the complex and diverse spatial features of optical remote sensing images and their extremely rich detailed texture information, accurately restoring the original detailed texture from noise interference is an extremely challenging task.

[0004] In the development of optical remote sensing image denoising technology, deep learning-based denoising methods have gradually surpassed traditional filtering methods and become the mainstream technology in this field due to their excellent denoising effect and flexible parameter settings. However, the development and optimization of deep learning denoising models are highly dependent on large-scale training data, especially mainstream supervised learning methods, which require a large number of noise-free real images as training labels to ensure the effectiveness and generalization of the model. But in practical applications, obtaining large-scale pairs of "ground truth - noise" remote sensing images faces many difficulties, not only in terms of complex data acquisition processes but also in high costs. When training data is limited, the performance of deep learning denoising models often falls short of expectations.

[0005] To address this issue, self-supervised learning methods have emerged. These methods do not rely on real labeled data, effectively reducing data dependence. However, during training, to avoid overfitting, self-supervised learning methods typically discard some pixel information. This limits the model's ability to recover image details and textures, making it difficult to achieve ideal denoising results. Furthermore, image denoising, as a generative task, places more stringent demands on GPU memory compared to tasks like image classification and object detection. When processing high-resolution optical remote sensing images, the model's memory usage and computational load increase exponentially, severely impacting its efficiency and posing significant challenges to both hardware and algorithm design.

[0006] Furthermore, most existing self-supervised denoising methods, such as Noise2Noise and Self2Self, implicitly assume that the noise in the noisy image is spatially independent or that the noise signal has zero mean. This makes these self-supervised denoising methods excellent at denoising independently distributed noise such as additive Gaussian noise, but there is still considerable room for improvement in denoising non-spatially independent noise such as Poisson distribution noise or photon noise from optical sensors. Consequently, these methods are difficult to effectively target the real noise signals present in optical remote sensing images, thus affecting the overall image denoising effect. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a self-supervised blind denoising method and system for optical remote sensing images. For single optical remote sensing images with real mixed noise, it disrupts the spatial correlation of noise by swapping non-neighboring similar pixels to target the spatially correlated components of the real noise. It then randomly uses random Bernoulli sampling to mask a small number of pixels to construct different sets of sub-image pairs, targeting spatially independent noise components and implementing self-supervised learning constraints. Finally, the denoised image is obtained by decoding the denoised encoded feature maps and taking the average of the image set results.

[0008] On the one hand, a self-supervised blind denoising method for optical remote sensing images is provided, including: Acquire the remote sensing image to be denoised, and perform non-local similarity sampling on the remote sensing image to be denoised to obtain the first noise space-independent similar image; The first set of image pairs is obtained by performing neighborhood Bernoulli sampling on the first noise-spatial-independent similar images. The first set of image pairs is input into the trained denoising model to obtain the denoised remote sensing image.

[0009] On the other hand, a self-supervised blind denoising system for optical remote sensing images is provided, including: The acquisition module is configured to: acquire the remote sensing image to be denoised, and perform non-local similarity sampling on the remote sensing image to be denoised to obtain a first noise space-independent similar image; The sampling module is configured to perform neighborhood Bernoulli sampling on the first noise spatially independent similar image to obtain a first set of image pairs; The denoising module is configured to input the first set of image pairs into the trained denoising model to obtain the denoised remote sensing image.

[0010] The above technical solution has the following advantages or beneficial effects: This invention demonstrates excellent performance in handling spatially correlated noise. The method primarily targets single optical remote sensing images containing real-world mixed noise. For spatially correlated components in the real noise, it disrupts spatial correlation by swapping non-neighboring similar pixels. For spatially independent noise components, random Bernoulli sampling is used to mask a small number of pixels, thereby constructing different sets of sub-image pairs and achieving self-supervised learning constraints. Furthermore, the parameterless attention computation method SimAM is employed to reduce the additional GPU memory usage caused by model parameters and guide the model to pay more attention to detailed texture features in the remote sensing image. Finally, the denoised encoded feature map is decoded, and the mean of the image set results is calculated to obtain the final denoised image. Attached Figure Description

[0011] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0012] Figure 1 This is a flowchart of the method in Example 1. Detailed Implementation

[0013] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0014] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] Example 1 This embodiment provides a self-supervised blind denoising method for optical remote sensing images; like Figure 1 As shown, a self-supervised blind denoising method for optical remote sensing images includes: S101: Acquire the remote sensing image to be denoised, and perform non-local similarity sampling on the remote sensing image to be denoised to obtain the first noise space-independent similar image; S102: Perform neighborhood Bernoulli sampling on the first noisy spatially independent similar image to obtain the first image pair set; S103: Input the first set of image pairs into the trained denoising model to obtain the denoised remote sensing image.

[0016] Further, step S101: acquiring the remote sensing image to be denoised, and performing non-local similarity sampling on the remote sensing image to be denoised to obtain a first noise-space-independent similar image, specifically includes: Calculate the cosine similarity between each pixel in the remote sensing image to be denoised and all other non-neighboring pixels in the image; Sort all non-neighborhood pixels in descending order of similarity value to obtain the top m non-neighborhood pixels with the highest similarity. Each pixel in the remote sensing image to be denoised is regarded as a target pixel. From the first m non-neighbor pixels with the highest similarity, a non-neighbor pixel is randomly selected as a similar pixel of each target pixel. Then, the target pixel and the similar pixel are used as a matching pair. Finally, the set of matching pairs of all pixels corresponding to the remote sensing image to be denoised is obtained. Then, based on the set of matching pairs, the similar pixels in each matching pair are replaced with the corresponding target pixels to obtain a similar noise image; the size of the similar noise image is the same as the size of the remote sensing image to be denoised; the similar noise image is the first noise spatially independent similar image.

[0017] It should be understood that, due to the complex distribution of real noise in remote sensing images, the noise signals are usually spatially correlated. To reduce the correlation between image content and noise signals, this invention employs a non-local pixel sampling and replacement method to calculate the cosine similarity between each pixel in the input image and all other non-neighboring pixels in the image. Its formula is: (1) in, For the current target pixel, For non-neighbor pixels, after sorting the similarity from high to low, select the top m non-neighbor pixels with the highest similarity, and randomly record one of the non-neighbor pixels similar to the current pixel (randomly select the m-th non-neighbor pixel from the top m non-neighbor pixels). 1 pixel, of which (and is an integer) is ,in The target pixels are randomly selected similar non-neighbor pixels. After traversing all pixels in the input image as target pixels, a set of similar pixel matches is obtained. , This represents the number of pixels in the input image. Each similar pixel is then rearranged to its corresponding position in the original image to obtain an image of the same size as the original input image. Similar noise images .

[0018] Due to similar noise images The spatial positions of the pixels are randomly arranged from the original image, which minimizes the spatial correlation of noise signals in the new similar image. Furthermore, by replacing similar pixels, the overall features of the image are not significantly changed, ensuring the integrity of the image content. This achieves the separation of spatial correlation between image content and noise signals, laying the foundation for subsequent self-supervised denoising.

[0019] Further, step S102: performing neighborhood Bernoulli sampling on the first noisy spatially independent similar images to obtain a first set of image pairs, specifically including: For the first noise-space-independent similar image Repeated random Bernoulli sampling is performed multiple times to generate multiple pairs of images with different pixels covered. : (2-1) (2-2) in, and This represents the complementary image generated after sampling. The number of times to perform repeated sampling. This represents the element-wise multiplication operation of a matrix. The masking matrix is ​​randomly generated during Bernoulli sampling. Each element The rules for determining the value of are as follows: (3) Formula (3) means that for each element Generate a random number between 0 and 1. If the random number is less than or equal to 1, then... Then element The value is 1; if the random number is greater than 1, the random number is greater than 1. and less than or equal to Then element The value is 1.

[0020] Each element in the masking matrix has a probability. The value of y is either 1 or 0, thus forming a masking matrix B with only 1 or 0 elements. After multiplying the masking matrix B with the corresponding elements of the image y, in the new output image, the positions in B where the element is 1 retain the pixel value of y, while the others (i.e., the positions in B where the element is 0) are set to 0.

[0021] It should be understood that in order to enable the denoising model to converge with only a single noisy image as input and to avoid overfitting during training, this invention employs a neighborhood Bernoulli sampling method to randomly generate multiple sets of "complementary" sub-image pairs from a single input image as training data for the model.

[0022] In the random Bernoulli sampling process of an image, each pixel is treated as an independent Bernoulli trial. For each pixel, a probability is set. probability This represents the probability that the current pixel will remain unchanged. Accordingly, That is the probability that a pixel will change. It is used to determine whether each pixel in an image should be changed, and whether it should be kept or removed.

[0023] It means that the current pixel has The possibility remains unchanged, otherwise (i.e.) The probability of ( ) will change to 0. Specifically, in code implementation, this means generating a random number in the range [0, 1] for each pixel; if this random number is less than or equal to 0, the probability of ( ) will change to 0. If the random number is less than 1, then it remains unchanged; if the random number is less than 1, then it remains unchanged. If , then it changes to 0.

[0024] For each pixel in the input image, a decision is made based on a set probability to retain or modify the pixel. Random Bernoulli sampling introduces randomness while maintaining the basic structure of the image, in order to avoid overfitting during the self-supervised training process. Due to the masking matrix Randomly generated according to the Bernoulli distribution, therefore in the set Each pair of images in Since the input images are not entirely identical, random Bernoulli sampling constrains the denoising model to learn a mapping of the input image itself, avoiding overfitting during model training. Furthermore, during the sampling process, the probability of each pixel being occluded is equal.

[0025] When the number of repetitions When large enough, the input image All pixels are included in the set In this approach, it can be assumed that all information in the original input image is used in the training, thus minimizing the loss of image information due to Bernoulli sampling.

[0026] Further, S103: the steps for obtaining the trained denoising model include: S103-1: Obtain the remote sensing image dataset, add noise to all remote sensing images in the dataset, and obtain a noisy remote sensing image dataset. S103-2: Perform non-local similarity sampling on each noisy remote sensing image in the noisy remote sensing image dataset to obtain a second noise spatially independent similar image; S103-3: Perform neighborhood Bernoulli sampling on the second noise-spatial-independent similar image to obtain the set of second image pairs; S103-4: Input the second set of image pairs into the denoising model, train the model, and obtain the trained denoising model.

[0027] Further, S103-1: Obtain a remote sensing image dataset, add noise to all remote sensing images in the dataset to obtain a noisy remote sensing image dataset, including: Using a known remote sensing image dataset, Gaussian-Poisson noise is added to the known remote sensing image dataset, and the image size is normalized to obtain a noisy remote sensing image dataset.

[0028] It should be understood that, due to the lack of large-scale, dedicated noisy remote sensing image datasets for denoising tasks, this invention uses the UC-Merced and Optical-31 remote sensing image classification datasets and adds Gaussian-Poisson joint distribution noise to construct a noisy image dataset for training. The constructed noisy dataset is then randomly divided into training and test sets at a ratio of 10:1. Because of the self-supervised learning nature of this invention, the ground truth images from the original UC-Merced and Optical-31 sets are not used in the training process but are only used as reference images for evaluation metrics.

[0029] Since some high-resolution remote sensing images are often very large, but computational resources are limited, the original images need to be adjusted and cropped before being fed into the model. In this invention, the model supports direct input of images with a size of 512×512 pixels. For high-resolution images much larger than this size, they are cropped to obtain multiple sub-images that meet this size. For input images smaller than this size or with an aspect ratio not of 1:1, they are upsampled and resized to adjust the input image to meet the 512×512 pixel input size.

[0030] Furthermore, the network structure of the trained denoising model includes: The encoders connected in sequence, the first 3 3 convolutional layers, the second 3 3 convolutional layers and decoder; The encoder includes: a first encoding module, a second encoding module, a third encoding module, a fourth encoding module, a fifth encoding module, and a sixth encoding module connected in sequence; The decoder includes: a first decoding module, a second decoding module, a third decoding module, a fourth decoding module, a fifth decoding module, a sixth decoding module, a first upsampling layer, and 1, connected in sequence. 1 convolutional layer; The first encoding module includes: a first convolutional layer, a second convolutional layer, a first activation function layer, a first attention mechanism layer, a first element-wise multiplication unit, and a first max pooling layer connected in sequence; The second encoding module includes: a third convolutional layer, a fourth convolutional layer, a second activation function layer, a second attention mechanism layer, a second element-wise multiplication unit, and a second max pooling layer connected in sequence; The third encoding module includes: a fifth convolutional layer, a sixth convolutional layer, a third activation function layer, a third attention mechanism layer, a third element-wise multiplication unit, and a third max pooling layer connected in sequence; The fourth encoding module includes: a seventh convolutional layer, an eighth convolutional layer, a fourth activation function layer, a fourth attention mechanism layer, a fourth element-wise multiplication unit, and a fourth max pooling layer connected in sequence; The fifth encoding module includes: a ninth convolutional layer, a tenth convolutional layer, a fifth activation function layer, a fifth attention mechanism layer, a fifth element-wise multiplication unit, and a fifth max pooling layer connected in sequence; The sixth encoding module includes: an eleventh convolutional layer, a twelfth convolutional layer, a sixth activation function layer, a sixth attention mechanism layer, and a sixth element-wise multiplication unit connected in sequence; The first decoding module includes: a seventh element-wise multiplication unit, a thirteenth convolutional layer, a fourteenth convolutional layer, and a first adder connected in sequence; The second decoding module includes: a second upsampling layer, an eighth element-wise multiplication unit, a fifteenth convolutional layer, a sixteenth convolutional layer, and a second adder connected in sequence; The third decoding module includes: a third upsampling layer, a ninth element-wise multiplication unit, a seventeenth convolutional layer, an eighteenth convolutional layer, and a third adder connected in sequence; The fourth decoding module includes: a fourth upsampling layer, a tenth element-wise multiplication unit, a nineteenth convolutional layer, a twentieth convolutional layer, and a fourth adder connected in sequence; The fifth decoding module includes: a fifth upsampling layer, an eleventh element-wise multiplication unit, a twenty-first convolutional layer, a twenty-second convolutional layer, and a fifth adder connected in sequence; The sixth decoding module includes: a sixth upsampling layer, a twelfth element-wise multiplication unit, a twenty-third convolutional layer, a twenty-fourth convolutional layer, and a sixth adder connected in sequence.

[0031] The output of the first attention mechanism layer is connected to the input of the twelfth element-wise multiplication unit; The output of the second attention mechanism layer is connected to the input of the eleventh element-wise multiplication unit; The output of the third attention mechanism layer is connected to the input of the tenth element-wise multiplication unit; The output of the fourth attention mechanism layer is connected to the input of the ninth element-wise multiplication unit; The output of the fifth attention mechanism layer is connected to the input of the eighth element-wise multiplication unit; The output of the sixth attention mechanism layer is connected to the input of the seventh element-wise multiplication unit; The output of the first convolutional layer is connected to the input of the sixth adder; the output of the second convolutional layer is connected to the input of the fifth adder; the output of the third convolutional layer is connected to the input of the fourth adder; the output of the fourth convolutional layer is connected to the input of the third adder; the output of the fifth convolutional layer is connected to the input of the second adder; and the output of the sixth convolutional layer is connected to the input of the first adder.

[0032] For example, the model in this invention is based on the U-Net architecture. In the encoding phase of this invention, the model consists of six encoder blocks (EBs). The first five EBs consist of two layers. The first EB consists of partial convolutional layers plus LeakyReLu activation layers, a parameterless attention mechanism layer Sim-AM, and a max-pooling layer, while the sixth EB does not include a max-pooling layer. In each EB, the attention map output by Sim-AM guides the model to pay more attention to the detailed features of meaningful parts in high-level semantics during the encoding process.

[0033] Furthermore, in the denoising model of this invention, the feature map input to the convolutional layer undergoes random Bernoulli sampling, causing some pixels to be masked and take a value of 0. Conventional convolutional operations, when processing such feature maps, do not distinguish between valid pixels and masked pixels with a value of 0, which easily leads to undesirable results such as color differences and artifacts. To overcome this problem, this method uses partial convolution to replace the original convolutional operation in the encoder.

[0034] Partial Convolution performs convolution operations only on the valid pixel portion and also has the function of automatically updating the mask. This ensures that the convolution process is not affected by pixels that are masked to 0, thus effectively solving the problems caused by improper convolution operations in image completion tasks.

[0035] Compared to typical convolutional neural networks, partial convolution requires an input mask matrix to avoid convolution on invalid pixels. The formula is as follows: (4) in, For convolution kernel, The input image is within the receptive field of the current convolution. This is the mask matrix within the receptive field of the current convolution. Represents the size of the convolution kernel. The receptive field representing convolution The region in the matrix where the value is 1.

[0036] In partial convolution, the masking matrix in each convolutional layer is updated as the layer number increases. The update rule is as follows: if the convolution can compute the output based on at least one valid input value, then in the next step, we update the masking matrix of this pixel. Marked as 1. That is, the partial convolution of the first... This point in the layer The value is determined by the first Within the receptive field of the layer The value determines this. For example: if the first... Within the receptive field of the layer If the value is 1, then the point is at the th position. Partial convolution of layers The value is 1 if it is positive and 0 otherwise. This can be expressed as the following formula: For the updated The value of the element in the middle: (5) Equation (5) can be used to explain the process of updating the masking matrix mask in each partial convolution layer as described above: This is an element in the masking matrix mask in the next layer of partial convolution. If at least one element in the mask matrix within the receptive field of the currently computed convolution kernel in this layer is 1, then the mask element value at the corresponding position of the convolution kernel in the next layer is still 1; otherwise, it is 0.

[0037] Furthermore, to reduce the number of parameters required for attention map computation and thus save memory, this invention uses the parameterless attention mechanism Sim-AM to compute the attention map. According to relevant research in neuroscience, neurons with unique firing patterns typically carry richer information and can efficiently encode and transmit complex information, thus playing a crucial role in visual processing. During neuronal activity, a phenomenon called neuronal spatial inhibition exists, where active neurons inhibit the activity of surrounding neurons. Specifically, an active neuron can reduce the activity level of nearby neurons. This characteristic provides important clues for studying the activity patterns of neurons. Estimating and calculating the neuron's energy function can be a key means of identifying and locating activated neurons. Given the existence of neuronal spatial inhibition, there is a significant difference between activated neurons and surrounding neurons. Therefore, the activation level of the target neuron can be reflected by calculating the linear separability between the target neuron and its neighboring neurons. Thus, the neuron energy estimation function is: (6) in, for linear transformation , for linear transformation .in For the current target neuron, These are neurons other than the target neuron. The number of all neurons . and These are two different constants. For simplicity, let's... and Defined as binary tags respectively and The purpose of this formula is to find the linear separability between the current target neuron and other neurons. For this, we have: and When, in formula (6) When the minimum value is reached, the target neuron is present. The neuron that differs most from other neurons should carry richer and more effective feature information.

[0038] Furthermore, to simplify the computational workload, a fast solution for formula (6) is provided: (7-1) (7-2) Among them, there are and .

[0039] Based on this, the minimum energy of the target neuron can be derived. as follows: (8) In formula (8), The smaller the value, the greater the difference between the target neuron and other neurons, and the more attention should be given to the target neuron in the denoising task.

[0040] Therefore, the reciprocal of the result of formula (8) is used. This serves as a metric for determining the importance of neurons. After traversing all neurons, all... Arrange the neurons according to their positions in the original input feature map to obtain the output value of the attention mechanism layer: the attention map weight matrix. .

[0041] Overall, the encoding stage of the model in this invention is a feature extraction network with six encoder blocks. Each encoder block contains partial convolutions with 3×3 kernels and utilizes a Sim-AM attention module to guide the model to focus on detailed features. Simultaneously, leveraging the structural characteristics of the U-Net network, the feature maps and attention maps from the encoder block at this level are passed across to the same level of the decoder. Specifically, regarding the differences in encoder structure, the first five encoders each consist of two convolutional layers and one max-pooling layer, while the sixth encoder has only two convolutional layers and no pooling layer. All six encoder blocks include an attention module. When the input image size is... At that time, the final output feature map size of the model encoding stage is .

[0042] In this invention, the decoding stage of the model consists of six decoder blocks and one output layer. The first decoder includes two convolutional kernels. The remaining decoder adds an upsampling layer with a scaling factor of 2 at input to gradually recover the size of the feature maps. The final output stage consists of a 1×1 convolutional layer that compresses the 96 feature map channels to the number of channels in the input image (3 in this method) to ensure the output image size matches the input image size. Furthermore, the U-Net cascade operation combines the feature maps passed from the encoder at the same level with the encoded attention... Figure 1 When fed into the decoder, this design allows the attention map weights learned from the encoder to be effectively passed to the decoder, realizing the efficient transfer of attention features between the encoder and decoder, and ensuring the consistency and coherence of the feature extraction and reconstruction stages throughout the denoising task.

[0043] Furthermore, the loss function used during the training of the trained denoising model is: Assume random Bernoulli sampling is Then the first noise space-independent similar image The two sub-images obtained after random sampling are: , .structure As input to the denoising model, the optimization objective of the denoising model is: (9) in, and It is a complementary image, that is, it satisfies , The masking matrix in Bernoulli sampling. It is a sampling operation. It is the image output after y is sampled. It is a denoising model.

[0044] Meanwhile, to avoid color deviation and distortion in the denoised image, a perceptual loss term is added to the target optimization problem to ensure the consistency of high-level features in the denoised image. The feature extraction model in the perceptual loss uses a pre-trained ResNet18 network, and its calculation formula is as follows: (10) in, For the pre-trained network ResNet18 Feature maps output by each residual block This represents the number of residual blocks in the currently used pre-trained ResNet18 network. This applies when the feature extraction model is ResNet18. .

[0045] The complete objective function of the denoising model is: (11) in, is a hyperparameter representing the weighting coefficient of the perceptual loss.

[0046] It should be understood that during the training phase, multiple random Bernoulli sampling is required on the first noise-space-independent similar images after similarity replacement to obtain a set of multiple image pairs. As input to the model, the overall denoising model needs to be optimized, and its objective function is: (12) in, Let m be the masking matrix for the m-th sampling. For the denoising model, For model parameters, The number of times to perform repeated sampling. It is the square of the 2-norm.

[0047] As can be seen from the above formula, only when... The masked pixels are used to measure each pair of images. The loss. Due to Generated completely randomly according to the Bernoulli distribution, when the number of repetitions... When large enough, the difference between all image pixels can be measured by the total loss of all pairs.

[0048] During training, this invention uses the Adam optimizer. The network is then optimized using a dynamic learning rate. The detailed learning rate settings are as follows: the initial learning rate is... The decay rate is progressively reduced at 20%, 40%, 60%, and 80% of the total number of iterations, with a decay coefficient of 1. .

[0049] Hyperparameters in the loss function The calculation formula is as follows: (15) in, For the current iteration round, This represents the total number of iteration rounds. The decay coefficients are 2 and 0.5 respectively in this invention. The random Bernoulli iteration of the model is... That is, a total of 100 pairs of sampled subgraphs are randomly obtained as training data, and the total number of iterations is [number missing]. When the model converges to its optimum, i.e., when the loss function reaches its minimum, the model from the current training epoch is used as the optimal model for the inference process.

[0050] During the inference phase, the input image Multiple random Bernoulli sampling is performed to obtain multiple sets of new images, and the newly generated image sets are input into the trained denoising model.

[0051] Finally, for all output denoised images The process of averaging to obtain the final result can be represented as follows: (16) in, For the input image The An example output after the second Bernoulli sampling.

[0052] The two multiple random Bernoulli samplings performed during the training and inference phases are independent of each other. Although this method is computationally expensive, it effectively avoids model overfitting during training. In actual training, this method further suppresses overfitting during model training by performing data augmentation operations such as rotation and mirroring on the input image.

[0053] Example 2 This embodiment provides a self-supervised optical remote sensing image blind denoising system, including: The acquisition module is configured to: acquire the remote sensing image to be denoised, and perform non-local similarity sampling on the remote sensing image to be denoised to obtain a first noise space-independent similar image; The sampling module is configured to perform neighborhood Bernoulli sampling on the first noise spatially independent similar image to obtain a first set of image pairs; The denoising module is configured to input the first set of image pairs into the trained denoising model to obtain the denoised remote sensing image.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A self-supervised blind denoising method for optical remote sensing images, characterized by: include: Acquire the remote sensing image to be denoised, and perform non-local similarity sampling on the remote sensing image to be denoised to obtain a first noise spatially independent similar image, specifically including: Calculate the cosine similarity between each pixel in the remote sensing image to be denoised and all other non-neighboring pixels in the image; Sort all non-neighborhood pixels in descending order of similarity value to obtain the top m non-neighborhood pixels with the highest similarity. Each pixel in the remote sensing image to be denoised is regarded as a target pixel. From the first m non-neighbor pixels with the highest similarity, a non-neighbor pixel is randomly selected as a similar pixel of each target pixel. Then, the target pixel and the similar pixel are used as a matching pair. Finally, the set of matching pairs of all pixels corresponding to the remote sensing image to be denoised is obtained. Then, based on the set of matching pairs, the similar pixels in each matching pair are replaced with the corresponding target pixels to obtain a similar noise image; the size of the similar noise image is the same as the size of the remote sensing image to be denoised; the similar noise image is the first noise spatially independent similar image; The first set of image pairs is obtained by performing neighborhood Bernoulli sampling on the first noise-spatial-independent similar images. The first set of image pairs is input into the trained denoising model to obtain the denoised remote sensing image; The trained denoising model has the following network structure: The encoders connected in sequence, the first 3 3 convolutional layers, the second 3 3 convolutional layers and decoder; The encoder includes: a first encoding module, a second encoding module, a third encoding module, a fourth encoding module, a fifth encoding module, and a sixth encoding module connected in sequence; The first 5 encoder blocks consist of two layers The encoder block consists of a partial convolutional layer, a LeakyReLu activation layer, a parameterless attention mechanism layer (Sim-AM), and a max pooling layer, while the sixth encoder block does not include a max pooling layer. The decoder includes: a first decoding module, a second decoding module, a third decoding module, a fourth decoding module, a fifth decoding module, a sixth decoding module, a first upsampling layer, and 1, connected in sequence. 1 convolutional layer.

2. The self-supervised optical remote sensing image blind denoising method as described in claim 1, characterized in that, The first set of image pairs is obtained by performing neighborhood Bernoulli sampling on the first noise-spatial-independent similar images, specifically including: For the first noise-space-independent similar image Repeated random Bernoulli sampling is performed multiple times to generate multiple pairs of images with different pixels covered. : ； ； in, and This represents the complementary image generated after sampling. The number of times to perform repeated sampling. This represents the element-wise multiplication operation of a matrix. The masking matrix is ​​randomly generated during Bernoulli sampling. Each element The rules for determining the value of are as follows: ； For each element Generate a random number between 0 and 1. If the random number is less than or equal to 1, then... Then element The value is 1; if the random number is greater than 1, the random number is greater than 1. and less than or equal to Then element The value of is 1.

3. The self-supervised optical remote sensing image blind denoising method as described in claim 1, characterized in that, The steps to obtain the trained denoising model include: Obtain a remote sensing image dataset, add noise to all remote sensing images in the dataset, and obtain a noisy remote sensing image dataset; For each noisy remote sensing image in the noisy remote sensing image dataset, non-local similarity sampling is performed to obtain a second noise spatially independent similar image; For the second noise-spatial-independent similar image, perform neighborhood Bernoulli sampling to obtain the set of second image pairs; The second set of image pairs is input into the denoising model, and the model is trained to obtain the trained denoising model.

4. The self-supervised blind denoising method for optical remote sensing images as described in claim 1, characterized in that, The first encoding module includes: a first convolutional layer, a second convolutional layer, a first activation function layer, a first attention mechanism layer, a first element-wise multiplication unit, and a first max pooling layer connected in sequence; The second encoding module includes: a third convolutional layer, a fourth convolutional layer, a second activation function layer, a second attention mechanism layer, a second element-wise multiplication unit, and a second max pooling layer connected in sequence; The third encoding module includes: a fifth convolutional layer, a sixth convolutional layer, a third activation function layer, a third attention mechanism layer, a third element-wise multiplication unit, and a third max pooling layer connected in sequence; The fourth encoding module includes: a seventh convolutional layer, an eighth convolutional layer, a fourth activation function layer, a fourth attention mechanism layer, a fourth element-wise multiplication unit, and a fourth max pooling layer connected in sequence; The fifth encoding module includes: a ninth convolutional layer, a tenth convolutional layer, a fifth activation function layer, a fifth attention mechanism layer, a fifth element-wise multiplication unit, and a fifth max pooling layer connected in sequence; The sixth encoding module includes: an eleventh convolutional layer, a twelfth convolutional layer, a sixth activation function layer, a sixth attention mechanism layer, and a sixth element-wise multiplication unit connected in sequence.

5. The self-supervised blind denoising method for optical remote sensing images as described in claim 4, characterized in that, The first decoding module includes: a seventh element-wise multiplication unit, a thirteenth convolutional layer, a fourteenth convolutional layer, and a first adder connected in sequence; The second decoding module includes: a second upsampling layer, an eighth element-wise multiplication unit, a fifteenth convolutional layer, a sixteenth convolutional layer, and a second adder connected in sequence; The third decoding module includes: a third upsampling layer, a ninth element-wise multiplication unit, a seventeenth convolutional layer, an eighteenth convolutional layer, and a third adder connected in sequence; The fourth decoding module includes: a fourth upsampling layer, a tenth element-wise multiplication unit, a nineteenth convolutional layer, a twentieth convolutional layer, and a fourth adder connected in sequence; The fifth decoding module includes: a fifth upsampling layer, an eleventh element-wise multiplication unit, a twenty-first convolutional layer, a twenty-second convolutional layer, and a fifth adder connected in sequence; The sixth decoding module includes: a sixth upsampling layer, a twelfth element-wise multiplication unit, a twenty-third convolutional layer, a twenty-fourth convolutional layer, and a sixth adder connected in sequence; The output of the first attention mechanism layer is connected to the input of the twelfth element-wise multiplication unit; the output of the second attention mechanism layer is connected to the input of the eleventh element-wise multiplication unit; the output of the third attention mechanism layer is connected to the input of the tenth element-wise multiplication unit; the output of the fourth attention mechanism layer is connected to the input of the ninth element-wise multiplication unit; the output of the fifth attention mechanism layer is connected to the input of the eighth element-wise multiplication unit; and the output of the sixth attention mechanism layer is connected to the input of the seventh element-wise multiplication unit. The output of the first convolutional layer is connected to the input of the sixth adder; the output of the second convolutional layer is connected to the input of the fifth adder; the output of the third convolutional layer is connected to the input of the fourth adder; the output of the fourth convolutional layer is connected to the input of the third adder; the output of the fifth convolutional layer is connected to the input of the second adder; and the output of the sixth convolutional layer is connected to the input of the first adder.

6. The self-supervised blind denoising method for optical remote sensing images as described in claim 1, characterized in that, The loss function used during the training of the trained denoising model is: Assume random Bernoulli sampling is Then the first noise space-independent similar image The two sub-images obtained after random sampling are: , ;structure As input to the denoising model, the optimization objective of the denoising model is: ； in, and It is a complementary image, that is, it satisfies , The masking matrix in Bernoulli sampling. yes The image output after sampling operation It is a denoising model; ； in, For the pre-trained network ResNet18 Feature maps output by each residual block This represents the number of residual blocks in the currently used pre-trained ResNet18 network. The complete objective function of the denoising model is: ； in, is a hyperparameter representing the weighting coefficient of the perceptual loss.

7. The self-supervised optical remote sensing image blind denoising method as described in claim 6, characterized in that, Hyperparameters in the loss function The calculation formula is as follows: ； in, For the current iteration round, This represents the total number of iteration rounds. This is the attenuation coefficient.

8. The self-supervised blind denoising method for optical remote sensing images as described in claim 3, characterized in that, To obtain a noisy remote sensing image dataset, noise is added to all remote sensing images in the dataset. This includes using a known remote sensing image dataset and adding Gaussian-Poisson noise to it, as well as normalizing the image sizes.

9. A self-supervised blind denoising system for optical remote sensing images, characterized in that, include: The acquisition module is configured to: acquire the remote sensing image to be denoised; perform non-local similarity sampling on the remote sensing image to be denoised to obtain a first noise-space-independent similar image, specifically including: Calculate the cosine similarity between each pixel in the remote sensing image to be denoised and all other non-neighboring pixels in the image; Sort all non-neighborhood pixels in descending order of similarity value to obtain the top m non-neighborhood pixels with the highest similarity. Each pixel in the remote sensing image to be denoised is regarded as a target pixel. From the first m non-neighbor pixels with the highest similarity, a non-neighbor pixel is randomly selected as a similar pixel of each target pixel. Then, the target pixel and the similar pixel are used as a matching pair. Finally, the set of matching pairs of all pixels corresponding to the remote sensing image to be denoised is obtained. Then, based on the set of matching pairs, the similar pixels in each matching pair are replaced with the corresponding target pixels to obtain a similar noise image; the size of the similar noise image is the same as the size of the remote sensing image to be denoised; the similar noise image is the first noise spatially independent similar image; The sampling module is configured to perform neighborhood Bernoulli sampling on the first noise spatially independent similar image to obtain a first set of image pairs; The denoising module is configured to input the first set of image pairs into the trained denoising model to obtain the denoised remote sensing image. The trained denoising model has the following network structure: The encoders connected in sequence, the first 3 3 convolutional layers, the second 3 3 convolutional layers and decoder; The encoder includes: a first encoding module, a second encoding module, a third encoding module, a fourth encoding module, a fifth encoding module, and a sixth encoding module connected in sequence; The first 5 encoder blocks consist of two layers The encoder block consists of a partial convolutional layer, a LeakyReLu activation layer, a parameterless attention mechanism layer (Sim-AM), and a max pooling layer, while the sixth encoder block does not include a max pooling layer. The decoder includes: a first decoding module, a second decoding module, a third decoding module, a fourth decoding module, a fifth decoding module, a sixth decoding module, a first upsampling layer, and 1, connected in sequence. 1 convolutional layer.

Images