Remote sensing image composite noise removal method based on deep learning

Abstract

The present invention relates to a remote sensing image composite noise removal method based on deep learning, which belongs to the field of image recognition. The original remote sensing image is input into a noise recognition and classification network for noise type recognition and classification, and a single-class noise image or a mixed noise image and its corresponding noise type are obtained; the single-class noise image is input into the corresponding trained single noise denoising model for denoising to obtain the final denoised image; the mixed noise image is copied according to the number of its noise types, and the copied mixed noise images are respectively input into the corresponding single noise denoising model to obtain multiple partial noise denoised images; the denoised image background and the remaining noise information are adjusted according to the partial noise denoised image to obtain a preliminary denoised image; the preliminary denoised image is reorganized by discrete wavelet transform and inverse discrete wavelet transform to obtain the final denoised image. The present invention can classify and denoise the mixed noise, and retain the image detail information as much as possible.

Description

Technical Field

[0001] The present invention belongs to the field of image recognition and relates to a method for removing composite noise from remote sensing images based on deep learning. Background Art

[0002] Remote sensing images are often polluted by various noises during the imaging process, including the noise of the sensor itself, atmospheric disturbances, the influence of surface features on light, errors in digital signal processing, and error factors in data transmission and storage. Due to the existence of these noises, the performance of related image vision tasks will be seriously affected, so denoising is a very important part of image processing. Common noise models include Gaussian noise, impulse noise, Poisson noise, stripe noise, etc.

[0003] According to the methods and techniques of remote sensing image denoising in the past decades, remote sensing image denoising algorithms can be divided into traditional algorithms and algorithms based on deep learning. Traditional algorithms are usually based on mathematical models and statistical methods, including filtering, wavelet transform, edge preservation and other technologies. Among them, common filtering methods include mean filtering, median filtering, Gaussian filtering, etc. These methods try to remove noise and retain image details by processing the image at the pixel level. Algorithms based on deep learning use deep learning neural network models to learn feature representations for image denoising. In 2017, the proposed DnCNN network was the first deep learning algorithm used in image denoising. Compared with traditional algorithms, the denoising speed, effect and generalization ability have been greatly improved; in 2018, the proposed FFDNet network effectively removed noise from the image and maintained the image details by introducing a step-by-step feedback mechanism and a variable weight strategy; in 2019, the RIDNet network based on residuals and dense connections introduced the structure of residuals and dense blocks to learn richer image features and achieve better denoising effects. Although the convolutional neural network method has achieved good results in the field of remote sensing image denoising, the remote sensing image denoising network based on deep learning is dependent on specific noise models or has some noise model assumption restrictions, which is unreasonable in some scenarios and will reduce the denoising effect of the algorithm to a certain extent.

[0004] Considering the denoising characteristics of remote sensing images, the type of noise will determine the way and degree of image degradation. In order to adapt to diverse noise interference, it is necessary to design a remote sensing image composite noise removal method that can simultaneously handle mixed noise interference. Summary of the invention

[0005] In view of this, the object of the present invention is to provide a method for removing composite noise from remote sensing images based on deep learning.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A remote sensing image composite noise removal method based on deep learning, the method comprising the following steps:

[0008] S1. Inputting the original remote sensing image into the noise recognition and classification network to perform noise type recognition and classification, and obtaining a single-class noise image or a mixed noise image and its corresponding noise type;

[0009] S2, using a noise data set with a single noise to train a denoising network model, and obtaining a corresponding single noise denoising model;

[0010] S3, inputting the single-type noise image into the corresponding single noise denoising model for denoising to obtain the final denoised image; duplicating the mixed noise image according to the number of its noise types, and respectively inputting the duplicated mixed noise images into the corresponding single noise denoising model to obtain multiple partial noise denoised images;

[0011] S4, adjusting the denoised image background and the remaining noise information according to the partial noise denoised image to obtain a preliminary denoised image;

[0012] S5. Input multiple preliminary denoised images into a multi-level wavelet convolutional neural network, and reorganize them through discrete wavelet transform and inverse discrete wavelet transform to obtain the final denoised image.

[0013] Further, in step S1, the noise recognition classification network extracts features and reduces the dimension of the image by a combination of convolution layer + ReLU and maximum pooling, obtains the category tensor through the fully connected layer, and finally normalizes the values ​​through the Sigmoid function to represent the probability of each type of noise, and determines whether the noise exists through the threshold;

[0014] The noise recognition classification network consists of several convolutional layers, several downsampling layers, a fully connected layer, and a classification layer. The softmax function is used to generate a vector representing the probability of the classification label.

[0015] The category labels of the noise recognition classification network use the commonly used encoding for multi-classification tasks;

[0016] The cross entropy loss function shown in the following formula is used as the loss function of the noise recognition classification network:

[0017]

[0018] in, is the predicted value; is the true value; n is the number of samples; c is the number of noise categories.

[0019] Further, in step S1, the image noise types identified by the noise recognition and classification network include Gaussian noise, salt and pepper noise, and stripe noise.

[0020] Further, in step S2, the denoising network model is divided into an upper network and a lower network, wherein the first layer of the 1-12 layers of the upper network layer is a deformable convolution layer Deformable Conv+LReLU, the 2-12 layers are Conv+BN+LReLU convolution layers, and the 1-12 layers are used to extract noise features; the 13th layer of the 13-16 layers is a dilated convolution layer Dilated Conv+BN+LReLU, the 14-16 layers are Conv+BN+LReLU convolution layers, and the 13-16 layers are used to extract robust noise information; the 17th layer is a Conv convolution layer;

[0021] The 1st layer and the 7th to 16th layers of the lower network layer are Conv+BN+LReLU, the 2nd, 3rd, 5th and 6th layers are dilated convolutions Dilated Conv BN+LReLU, the 4th layer is the MSFE module, and the MSFE module includes multiple convolution operations of different scales; the 17th layer is a Conv convolution layer.

[0022] Further, in step S2, when training the denoising network model, the following steps are included:

[0023] Select a large-scale remote sensing image dataset and process it to obtain the corresponding grayscale image;

[0024] A single type of Gaussian noise, salt and pepper noise or stripe noise is added to obtain three color and gray noise data sets, which are respectively input into the denoising network for training, and six denoising network models are obtained, including three grayscale image denoising models and three color image denoising models, corresponding to Gaussian denoising, salt and pepper denoising and stripe denoising respectively;

[0025] The loss function of the single-class noise denoising network adopts L 1 Norm:

[0026]

[0027] Among them, x i is the predicted value; i is the true value; f(·) is the network operation; n is the number of samples.

[0028] Further, in step S3, for a single-class noise image containing only a single noise type, the single-class noise image is input into a corresponding denoising model to obtain a final denoised image;

[0029] For mixed noise images including multiple types of noise, the image is copied to the same number of sheets as the number of noise types inside it, and each sheet is input into a different single denoising model. When denoising a color image, for a color image with non-additive Gaussian noise, the color image is separated into three RGB channels, and then each channel image is denoised using the grayscale denoising model corresponding to the noise type, and the denoised channel results are integrated into the color image.

[0030] Further, in step S4, the image degradation formula under the composite noise interference is as follows:

[0031]

[0032] Where y is the noise image; x is the noise-free image; n j is the noise; j represents the corresponding serial number of the noise type;

[0033] When the composite noise passes through the single-class denoising network, the image background and the remaining noise information change, and the initial denoising result after denoising is expressed as:

[0034] y j =D j (x+Σ k∈K n k )=x+Σ k∈K,k≠j n k +Δ

[0035] Among them, y j represents the image after removing the noise number j; D j represents the single-class noise removal network corresponding to j noise; Δ represents D j The slight changes caused to the background information and other noise; K is the range of noise numbers. j =D j (x+∑ k∈K n k )=x+∑ k∈K,k≠j n k +Δ, we get t single-class noise removal result images, add them and subtract (t-1)y to get the preliminary denoised image y', and continue the subsequent operations on y'.

[0036] Further, in step S5, the noise of several preliminary denoised images obtained is removed by using the MWCNN network structure, wherein wavelet transform is introduced to reduce the size of feature map and increase the receptive field, feature extraction and reorganization are performed through the convolution layer, and then the inverse wavelet transform is deployed to reorganize the image, wherein,

[0037] The specific operation of discrete wavelet transform (DWT) is as follows:

[0038] x LL (m,n)=x(2m-1,2n-1)+x(2m-1,2n)+x(2m,2n-1)+x(2m,2n)

[0039] x LH (m,n)=-x(2m-1,2n-1)-x(2m-1,2n)+x(2m,2n-1)+x(2m,2n)

[0040] x HL (m,n)=-x(2m-1,2n-1)+x(2m-1,2n)-x(2m,2n-1)+x(2m,2n)

[0041] x HH (m,n)=x(2m-1,2n-1)-x(2m-1,2n)-x(2m,2n-1)+x(2m,2n)

[0042] The specific operation of inverse discrete wavelet transform (IWT) is as follows:

[0043]

[0044]

[0045]

[0046]

[0047] Among them, x LL (m,n), x LH (m,n), x HL (m,n), x HH (m,n) represents the value of the wavelet decomposition image in different directions at point (m,n); x(m,n) represents the value of the image at point (m,n). The final denoised image is obtained by recombining through discrete wavelet transform and inverse discrete wavelet transform.

[0048] The beneficial effects of the present invention are:

[0049] The present invention can not only classify and denoise a single type of noise, but also classify and denoise mixed types of noise according to actual needs, thereby effectively removing the composite noise while retaining the image detail information as much as possible.

[0050] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0052] Figure 1 It is the overall flow chart of the present invention;

[0053] Figure 2 It is the overall architecture diagram of the present invention;

[0054] Figure 3 It is a structural diagram of the noise identification and classification network of the present invention;

[0055] Figure 4 It is a structural diagram of the denoising network model of the present invention;

[0056] Figure 5 It is a structural diagram of a multi-scale feature extraction module in the denoising network model of the present invention;

[0057] Figure 6 It is a structural diagram of the multi-level wavelet convolutional neural network of the present invention. DETAILED DESCRIPTION

[0058] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0059] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0060] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0061] See also Figure 1 to Figure 6 , which is a remote sensing image composite noise removal method based on deep learning. It can not only classify and denoise a single type of noise, but also classify and denoise mixed types of noise according to actual needs, while effectively removing composite noise and retaining image detail information as much as possible.

[0062] like Figure 1 The overall flow chart of the method for removing complex noise from remote sensing images based on deep learning of the present invention is shown as well as Figure 2 The overall architecture diagram of the remote sensing image complex noise removal method based on deep learning of the present invention is shown, which includes the following steps:

[0063] S1. Inputting the original remote sensing image into the noise recognition and classification network to perform noise type recognition and classification, and obtaining a single-class noise image or a mixed noise image and its corresponding noise type;

[0064] S2, using a noise data set with a single noise to train a denoising network model, and obtaining a corresponding single noise denoising model;

[0065] S3, inputting the single-type noise image into the corresponding single noise denoising model for denoising to obtain the final denoised image; duplicating the mixed noise image according to the number of its noise types, and respectively inputting the duplicated mixed noise images into the corresponding single noise denoising model to obtain multiple partial noise denoised images;

[0066] S4, adjusting the denoised image background and the remaining noise information according to the partial noise denoised image to obtain a preliminary denoised image;

[0067] S5. Input multiple preliminary denoised images into a multi-level wavelet convolutional neural network, and reorganize them through discrete wavelet transform and inverse discrete wavelet transform to obtain the final denoised image.

[0068] Example

[0069] This embodiment provides a specific implementation method of the remote sensing image complex noise removal method based on deep learning of the present invention. First, the original remote sensing image is input, and the noise category of the remote sensing image is identified and classified through the remote sensing image noise recognition and classification network to obtain a single-class noise image or a mixed noise image of Gaussian noise, salt and pepper noise, and stripe noise. Then, the single-class noise image and the mixed noise image are processed separately, and the single-class noise image passes through the corresponding single-class denoising network; and the mixed noise image is copied according to the number of noise types contained, and passes through the corresponding single-class noise removal network to obtain multiple partial noise denoising images, and then these images are used by mathematical methods to obtain a preliminary denoising image, and finally the preliminary denoising image is passed through a multi-level wavelet convolutional neural network to obtain the final denoising image.

[0070] In step S1 of this embodiment, the recognition and classification of noise can effectively provide information support for the subsequent denoising method. Therefore, training a noise recognition network that can accurately identify noise categories is the first step in conducting composite noise research. Compared with common visual classification tasks, image noise recognition has fewer categories and noise is a low-level feature. Therefore, a simple classification network with a shallow number of layers can accurately identify noise categories. The specific structure of the noise recognition classification network is as follows: Figure 3 As shown, the image is feature extracted and dimension reduced by a combination of convolution layer + ReLU and maximum pooling, and a category tensor is obtained through a fully connected layer. Finally, the value is normalized by the Sigmoid function to represent the probability of various types of noise, and the presence of noise is determined by a threshold. Ideally, if such noise exists, the output value of the corresponding bit of the tensor is 1, and if it does not exist, it is 0. In this embodiment, the framework is based on VGG16, including 4 convolutional layers, 4 downsampling layers, and 1 fully connected layer. The last layer is a classification layer, and a softmax function is used to generate a vector representing the probability of a classification label.

[0071] The category labels of the noise recognition classification network use the commonly used encoding for multi-classification tasks, and the loss function uses the cross entropy loss function corresponding to formula (1).

[0072]

[0073] in, is the predicted value; is the true value; n is the number of samples; c is the number of noise categories.

[0074] In step S2 of this embodiment, the single-category noise removal network can select a denoising network suitable for the noise type for removal. Different network architectures can extract different features, and these extracted features are complementary in image denoising. The overall double-branch structure is divided into an upper network and a lower network. By increasing the width of the network, the denoising performance of the model is improved. The network model is as follows: Figure 4As shown, it includes an upper network and a lower network.

[0075] In the upper network, more representative noise features are extracted through 1-12 layers of deformable and stacked convolutional architecture (1 deformable convolutional layer + 11 convolutional layers). The deformable convolutional layer is Deformable Conv+LReLU, and the remaining 11 layers are Conv+BN+LReLU. At the same time, in order to mine more structural information at a lower cost, the 13-16 layers promote contextual interaction through dilated convolution. The dilated convolution process includes Dilated Conv+BN+LReLU in the 13th layer and Conv+BN+LReLU in the 14-16 layers to extract more robust noise information. The 17th layer is Conv.

[0076] The lower network considers the balance between performance and efficiency, uses dilated convolution to capture more contextual information, and adopts a multi-scale feature extraction module and residual learning strategy. The multi-scale feature extraction MSFE module is used to extract features from multi-scale noisy images. The first layer, 7-16 layers of the lower network are Conv+BN+LReLU, the second, third, fifth, and sixth layers are dilated convolution, which is Dilated Conv BN+LReLU; the fourth layer is the MSFE module, and the last layer, the 17th layer, is Conv.

[0077] like Figure 5 As shown in Figure 1, the MSFE block consists of five convolution operations at different scales. The features of the noise image at five different scales are analyzed and finally concatenated to achieve the final multi-scale feature extraction graph.

[0078] The output of the network is a residual noise image, which needs to be subtracted from the noisy input image to get the final denoised image.

[0079] When training the denoising network model, first, a large-scale remote sensing image dataset is selected and processed to obtain the corresponding grayscale image. Then a single type of noise (Gaussian noise, salt and pepper noise, stripe noise) is added to obtain three color and gray noise datasets, which are input into the denoising network for training, and 6 denoising models are obtained, including 3 grayscale image denoising models and 3 color image denoising models, corresponding to Gaussian denoising, salt and pepper denoising, and stripe denoising.

[0080] The loss function of the single-class noise denoising network adopts L 1 Norm.

[0081]

[0082] Among them, x i is the predicted value; iis the true value; f(·) is the network operation; n is the number of samples.

[0083] In step S3 of this embodiment, for a single-class noise image containing only a single noise type, the single-class noise image is input into the corresponding denoising model to obtain the final denoised image, and the denoising process of the single-class noise image is relatively simple. For a mixed noise image including multiple types of noise, the image is copied to a number of sheets that is consistent with the number of noise types inside it, and each sheet is input into a different single denoising model. When denoising a color image, for a color image with non-additive Gaussian noise, the color image is separated into three RGB channels, and then each channel image is denoised using a grayscale denoising model corresponding to the noise type, and the denoised channel results are integrated into the color image.

[0084] In step S4 of this embodiment, the image degradation formula under the interference of composite noise is shown in formula (3):

[0085]

[0086] Where y is the noise image; x is the noise-free image; n j is the noise; j represents the corresponding serial number of the noise type.

[0087] When the composite noise passes through a single-class denoising network, the removal of the corresponding class of noise is complete, but the image background and the remaining noise information will also change, thus we can get formula (4):

[0088] y j =D j (x+∑ k∈K n k )=x+∑ k∈K,k≠j n k +Δ (4)

[0089] Among them, y j represents the image after removing the noise number j; D j represents the single-class noise removal network corresponding to j noise; Δ represents D j The slight changes caused to the background information and other noise; K is the range of noise numbers, K refers to the number i of the noise category, where 1, 2, and 3 refer to Gaussian, salt and pepper, and stripe noise respectively. j =D j (x+∑ k∈K n k )=x+∑ k∈K,k≠j n k +Δ, we get t (t is the number of noise categories in the image) single-class noise removal result images, add them together and subtract (t-1)y to get the preliminary denoising result y', and continue the subsequent operations with y'

[0090] In step S5, based on the preliminary denoising result, this embodiment continues to perform the following steps: Figure 6 The MWCNN structure shown in the figure is removed. The network introduces wavelet transform to reduce the size of feature maps and increase the receptive field, extracts and reorganizes features through convolutional layers, and then deploys inverse wavelet transform to reconstruct the image, which has excellent performance in texture information recovery. The loss function of this network is L 1 Norm.

[0091] The specific operation of discrete wavelet transform (DWT) is as follows:

[0092]

[0093] The specific operation of inverse discrete wavelet transform (IWT) is as follows:

[0094]

[0095] Among them, x LL (m,n), x LH (m,n), x HL (m,n), x HH (m,n) represents the value of the wavelet decomposition image in different directions at point (m,n); x(m,n) represents the value of the image at point (m,n). The final denoised image is obtained by recombining through discrete wavelet transform and inverse discrete wavelet transform.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A remote sensing image composite noise removal method based on deep learning, characterized by: The method comprises the following steps: S1. Inputting the original remote sensing image into the noise recognition and classification network to perform noise type recognition and classification, and obtaining a single-class noise image or a mixed noise image and its corresponding noise type; S2, using a noise data set with a single noise to train a denoising network model, and obtaining a corresponding single noise denoising model; S3, inputting the single-type noise image into the corresponding single noise denoising model for denoising to obtain the final denoised image; duplicating the mixed noise image according to the number of its noise types, and inputting the duplicated mixed noise images into the corresponding single noise denoising model respectively to obtain multiple partial noise denoised images; S4, adjusting the denoised image background and the remaining noise information according to the partial noise denoised image to obtain a preliminary denoised image; S5. Input multiple preliminary denoised images into a multi-level wavelet convolutional neural network, and reorganize them through discrete wavelet transform and inverse discrete wavelet transform to obtain the final denoised image.

2. The method for removing composite noise from remote sensing images based on deep learning according to claim 1, characterized in that: In step S1, the noise recognition classification network extracts features and reduces the dimension of the image by a combination of convolution layer + ReLU and maximum pooling, obtains the category tensor through the fully connected layer, and finally normalizes the value through the Sigmoid function to represent the probability of each type of noise, and determines whether the noise exists through the threshold; The noise recognition classification network consists of several convolutional layers, several downsampling layers, a fully connected layer, and a classification layer. The softmax function is used to generate a vector representing the probability of the classification label. The category labels of the noise recognition classification network use the commonly used encoding for multi-classification tasks; The cross entropy loss function shown in the following formula is used as the loss function of the noise recognition classification network: in, is the predicted value; is the true value; n is the number of samples; c is the number of noise categories.

3. The method for removing composite noise from remote sensing images based on deep learning according to claim 2, characterized in that: In step S1, the image noise types identified by the noise recognition and classification network include Gaussian noise, salt and pepper noise, and stripe noise.

4. The method for removing composite noise from remote sensing images based on deep learning according to claim 3, characterized in that: In step S2, the denoising network model is divided into an upper network and a lower network, wherein the first layer of the 1-12 layers of the upper network layer is a deformable convolution layer Deformable Conv+LReLU, the 2-12 layers are Conv+BN+LReLU convolution layers, and the 1-12 layers are used to extract noise features; the 13th layer of the 13-16 layers is a dilated convolution layer Dilated Conv+BN+LReLU, the 14-16 layers are Conv+BN+LReLU convolution layers, and the 13-16 layers are used to extract robust noise information; the 17th layer is a Conv convolution layer; The 1st layer and the 7th to 16th layers of the lower network layer are Conv+BN+LReLU, the 2nd, 3rd, 5th and 6th layers are dilated convolution DilatedConv BN+LReLU, the 4th layer is the MSFE module, and the MSFE module includes multiple convolution operations of different scales; the 17th layer is the Conv convolution layer.

5. The method for removing composite noise from remote sensing images based on deep learning according to claim 4, characterized in that: In step S2, when training the denoising network model, the following steps are included: Select a large-scale remote sensing image dataset and process it to obtain the corresponding grayscale image; A single type of Gaussian noise, salt and pepper noise or stripe noise is added to obtain three color and gray noise data sets, which are respectively input into the denoising network for training, and six denoising network models are obtained, including three grayscale image denoising models and three color image denoising models, corresponding to Gaussian denoising, salt and pepper denoising and stripe denoising respectively; The loss function of the single-class noise denoising network uses the L1 norm: Among them, x i is the predicted value; i is the true value; f(·) is the network operation; n is the number of samples.

6. The method for removing composite noise from remote sensing images based on deep learning according to claim 5, characterized in that: In step S3, for a single-class noise image containing only a single noise type, the single-class noise image is input into a corresponding denoising model to obtain a final denoised image; For mixed noise images including multiple types of noise, the image is copied to the same number of sheets as the number of noise types inside it, and each sheet is input into a different single denoising model. When denoising a color image, for a color image with non-additive Gaussian noise, the color image is separated into three RGB channels, and then each channel image is denoised using the grayscale denoising model corresponding to the noise type, and the denoised channel results are integrated into the color image.

7. The method for removing composite noise from remote sensing images based on deep learning according to claim 6, characterized in that: In step S4, the image degradation formula under the composite noise interference is as follows: Where y is the noise image; x is the noise-free image; n j is the noise; j represents the corresponding serial number of the noise type; When the composite noise passes through the single-class denoising network, the image background and the remaining noise information change, and the initial denoising result after denoising is expressed as: y j =D j (x+∑ k∈K n k )=x+∑ k∈K,k≠j n i +Δ Among them, y j represents the image after removing the noise number j; D j represents the single-class noise removal network corresponding to j noise; Δ represents D j The slight changes caused to the background information and other noise; K is the range of noise number; through the formula y j =D j (x+∑ k∈K n k )=x+∑ k∈K,k≠j n i +Δ, we get t single-class noise removal result images, add them and subtract (t-1)y to get the preliminary denoised image y', and continue the subsequent operations on y'.

8. The method for removing composite noise from remote sensing images based on deep learning according to claim 7, characterized in that: In step S5, the noise of several preliminary denoised images obtained is removed by using the MWCNN network structure, wherein wavelet transform is introduced to reduce the size of feature maps and increase the receptive field, feature extraction and reorganization are performed through the convolution layer, and then the inverse wavelet transform is deployed to reorganize the image, wherein, The specific operation of discrete wavelet transform (DWT) is as follows: x LL (m,n)=x(2m-1,2n-1)+x(2m-1,2n)+x(2m,2n-1)+x(2m,2n) x LH (m,n)=-x(2m-1,2n-1)-x(2m-1,2n)+x(2m,2n-1)+x(2m,2n) x HL (m,n)=-x(2m-1,2n-1)+x(2m-1,2n)-x(2m,2n-1)+x(2m,2n) x HH (m,n)=x(2m-1,2n-1)-x(2m-1,2n)-x(2m,2n-1)+x(2m,2n) The specific operation of inverse discrete wavelet transform (IWT) is as follows: Among them, x LL (m,n), x LH (m,n), x HL (m,n), x HH (m,n) represents the value of the wavelet decomposition image in different directions at point (m,n); x(m,n) represents the value of the image at point (m,n). The final denoised image is obtained by recombining through discrete wavelet transform and inverse discrete wavelet transform.

Images