A remote sensing image denoising method based on a new optimizer

Abstract

This invention discloses a remote sensing image denoising method based on a novel optimizer, which relates to the field of remote sensing image denoising. The method comprises: acquiring a noisy remote sensing image and dividing the image into a training set and a test set; performing gradient optimization on a DudeNet denoising model using the GAF algorithm, and training the gradient-optimized DudeNet denoising model using the training set; testing the trained DudeNet denoising model using the test set and evaluating it using image quality metrics, thereby completing remote sensing image denoising based on the novel optimizer. This invention solves the problems of poor denoising performance and high computational overhead in existing denoising methods.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image denoising, and in particular to a remote sensing image denoising method based on a novel optimizer. Background Art

[0002] Remote sensing images are widely used in fields such as geographic information monitoring, environmental monitoring, agricultural surveys, and urban planning. However, remote sensing images are often affected by noise during the acquisition process. In particular, in complex environments such as low light, fog, sensor failure, or atmospheric interference, the image quality is poor, which seriously affects subsequent image analysis and interpretation. Therefore, the problem of remote sensing image denoising has become an important research direction in remote sensing image processing. Traditional remote sensing image denoising methods mainly include filtering algorithms and statistical models, but these methods often fail to effectively remove noise while preserving image details. In recent years, denoising methods based on deep learning have become mainstream, but due to problems such as gradient vanishing and local optimality in the training process of deep neural networks, they often result in poor denoising effects or high computational overhead. Therefore, how to design an efficient and accurate remote sensing image denoising model remains a current technical challenge. Summary of the Invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a remote sensing image denoising method based on a novel optimizer, which solves the problems of poor denoising effect and high computational overhead of the existing denoising methods.

[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a remote sensing image denoising method based on a new optimizer, comprising the following steps:

[0005] S1: Collecting remote sensing noise images and dividing the remote sensing noise images into a training set and a test set;

[0006] S2: Use the GAF algorithm to perform gradient optimization on the DudeNet denoising model, and use the training set to train the gradient-optimized DudeNet denoising model;

[0007] S3: Test the trained DudeNet denoising model using the test set and evaluate it using image quality metrics to complete remote sensing image denoising based on the new optimizer.

[0008] Furthermore, the step S2 includes the following sub-steps:

[0009] S21: using the scaling factor α and sensitivity factor β in the GAF algorithm to adjust the gradient range of the GAF algorithm;

[0010] S22: Use the loss function to update the gradient of the GAF algorithm, and replace the gradient of the DudeNet denoising model with the updated gradient of the GAF algorithm to obtain the gradient-optimized DudeNet denoising model;

[0011] S23: Use the training set to train the gradient-optimized DudeNet denoising model.

[0012] Furthermore, the scaling factor α and the sensitivity factor β in S21 are set to {α=0.1, β=20} and {α=0.2, β=10}.

[0013] Furthermore, in S22, the loss function is used to update the gradient of the GAF algorithm, and the formula is:

[0014]

[0015] in, is the gradient of the random training process, ▽ is the gradient symbol, ω is the weight vector, i is the number of iterations, k is the number of data sets, L(·) is the loss function, M(·) is the output of the model for each input sample in each iteration, X (·) is the input sample in the training batch, ω (·) is the weight vector, y (·) is the corresponding label of the input sample, g k is the gradient of the kth iteration, μ m is the momentum coefficient, g k-1 is the gradient of the k-1th iteration, η is the learning rate, g' k The gradient after momentum update is applied to the GAF function to obtain the adjusted gradient, g'(·) is a different type of GAF function, ω k is the weight vector for the kth iteration.

[0016] Furthermore, the DudeNet denoising model in S22 includes a feature extraction module FEB, an image enhancement module EB, a compression module CB and a reconstruction module RB.

[0017] Furthermore, the feature extraction module FEB includes two parallel first feature extraction modules FEBnet1 and second feature extraction modules FEBnet2, wherein the first feature extraction module FEBnet1 extracts global features and local features from the input noisy image through a sparse mechanism, and the second feature extraction module FEBnet2 is used to extract supplementary features from the input noisy image;

[0018] The first feature extraction module FEBnet1 adopts a sparse mechanism, including a convolutional layer, a batch normalization layer and a ReLU activation layer;

[0019] The output FEB1 of the first feature extraction module FEBnet1 is expressed as:

[0020] FEB1=C(CBR3(S(CBR1(Y))))

[0021] where C(·) is a 3×3 convolution function, CBR3(·) is a module consisting of three convolutional layers, a batch normalization layer, and a ReLU activation layer, S(·) is a sparse mechanism, CBR1(·) is a module consisting of one convolutional layer, a batch normalization layer, and a ReLU activation layer, and Y is the input noisy image.

[0022] The second feature extraction module FEBnet2 includes a convolutional layer, a ReLU activation layer and a convolutional layer connected in sequence;

[0023] The output FEB2 of the second feature extraction module FEBnet2 is expressed as:

[0024] FEB2=C1(CR 15 (Y))

[0025] Among them, C1(·) is a 3×3 convolution function, CR 15 (·) is a functional module that executes 15 convolutional layers and ReLU activation layers in sequence.

[0026] Furthermore, the image enhancement module EB includes a first image enhancement module EB1 and a second image enhancement module EB2, wherein the first image enhancement module EB1 is used to fuse the output features of the first feature extraction module FEBnet1 and the second feature extraction module FEBnet2, and process them through a batch normalization layer and a ReLU activation layer; the second image enhancement module EB2 is used to fuse the output features of the first image enhancement module EB1 and the input image, and compress the output features using 1×1 convolution;

[0027] The output OE1 of the first image enhancement module EB1 is expressed as:

[0028] OE1=E(FEB1,FEB2)

[0029] OE1=R(B(CON(C(CBR3(S(CBR1(Y))))),CB(CR 15 (Y)))))

[0030] Among them, E(·) is the fusion operation, R(·) is the activation layer, B(·) is the batch normalization layer, CON(·) is the convolution function, and CB(·) is the compression module;

[0031] The output OE2 of the second image enhancement module EB2 is expressed as:

[0032] OE2=E(OCB2,Y)

[0033] OCB2=C1(OE1)

[0034] Here, OCB2 represents convolution of the output OE1 of the first image enhancement module EB1 to compress data.

[0035] Furthermore, the compression module CB includes a first convolutional layer CB1, a second convolutional layer CB2, and a third convolutional layer CB3, wherein the first convolutional layer CB1 is located at the end of the second feature extraction module FEBnet2, the second convolutional layer CB2 is located between the first image enhancement module EB1 and the second image enhancement module EB2, and the third convolutional layer CB3 is located between the second image enhancement module EB2 and the reconstruction module RB;

[0036] The output of the compression module CB is represented as:

[0037] OCB3=C1(OE2)

[0038] Among them, OCB3 is the residual feature output by the third convolutional layer CB3.

[0039] Furthermore, the reconstruction module RB uses a residual connection to map the compressed features to the denoised image, and the output is expressed as:

[0040] X=Y-OCB3

[0041] Where X is the reconstructed clear image.

[0042] Furthermore, the image quality measurement index in S3 is:

[0043]

[0044] Where PSNR is the image quality measure, MAX is the maximum image intensity, MSE is the mean square error, p is the number of rows of image pixels, q is the number of columns of image pixels, r is the value of p, s is the value of q, I(·) is the pixel value of the original input image, and K(·) is the pixel value of the processed image.

[0045] The beneficial effects of the present invention are:

[0046] (1) The GAF of the present invention modifies gradients by applying a specific activation function. This function reduces excessive gradients and amplifies small gradients, thereby preventing oscillations in sharp directions while accelerating convergence in flat directions. In terms of theoretical support: The optimizer used in this invention provides theoretical proof that under certain conditions, GAF reduces the condition number of the optimization problem, which means greater stability and faster convergence.

[0047] (2) The design of the GAF of the present invention ensures that it avoids gradient values ​​that are too large or too small, preventing the problem of gradient vanishing or exploding, especially in deep networks. The arctan-type GAF is emphasized as being particularly effective. In terms of theoretical support: the present invention provides a theorem showing that under certain conditions, the GAF output will have a larger amplitude than the small gradient input, thus preventing the gradient from vanishing. It also inherently limits the output of the gradient, helping to prevent gradient explosion.

[0048] (3) By modifying the gradient landscape, GAF allows the optimizer to more efficiently move away from saddle points. This is achieved by reshaping the loss function, effectively making saddle points less pronounced. In terms of theoretical support, the present technique shows that under certain conditions, GAF-modified gradient descent converges to a minimum faster near saddle points than standard methods.

[0049] (4) The present invention solves the ill-posed problem and prevents vanishing / exploding gradients. GAF promotes faster convergence of the optimization algorithm. Theoretical Support: The present invention theoretically proves that under certain assumptions, including strong convexity of the loss function and Lipschitz continuity conditions, SGD with GAF converges faster than SGD without GAF.

[0050] (5) The GAF of the present invention is easy to implement and integrate into existing deep learning frameworks (PyTorch and TensorFlow), and the computational cost of adding GAF is very small, requiring only element activation on the gradient.

[0051] (6) The properties of GAF (amplifying small gradients and limiting large gradients) enable DudeNet to better adapt to various noise intensities and types. The present invention uses multiple GAFs, such as arctan, tanh, and log types, to explore which GAF is more suitable for image denoising. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of a remote sensing image denoising method based on a new optimizer.

[0053] Figure 2 A comparison chart of the PSNR values ​​of different optimizers in the image denoising model. DETAILED DESCRIPTION

[0054] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0055] like Figure 1 As shown, a remote sensing image denoising method based on a new optimizer is characterized by comprising the following steps:

[0056] S1: Collecting remote sensing noise images and dividing the remote sensing noise images into a training set and a test set;

[0057] S2: Use the GAF algorithm to perform gradient optimization on the DudeNet denoising model, and use the training set to train the gradient-optimized DudeNet denoising model;

[0058] S3: Test the trained DudeNet denoising model using the test set and evaluate it using image quality metrics to complete remote sensing image denoising based on the new optimizer.

[0059] In this example, the training set uses 100 JPEG-compressed images, each sized 512x512. The test set uses the online dataset CC, which contains 15 512x512 real-world noise images captured using three digital cameras: a Canon 5D Mark III, a Nikon D600, and a Nikon D800.

[0060] The S2 includes the following steps:

[0061] S21: using the scaling factor α and sensitivity factor β in the GAF algorithm to adjust the gradient range of the GAF algorithm;

[0062] S22: Use the loss function to update the gradient of the GAF algorithm, and replace the gradient of the DudeNet denoising model with the updated gradient of the GAF algorithm to obtain the gradient-optimized DudeNet denoising model;

[0063] S23: Use the training set to train the gradient-optimized DudeNet denoising model.

[0064] The scaling factor α and the sensitivity factor β in S21 are set to {α=0.1, β=20} and {α=0.2, β=10}.

[0065] In GAF, α and β are two important parameters used to control the shape and strength of the gradient activation function. Together, they determine how GAF amplifies small gradients and limits large gradients.

[0066] α primarily controls the range of GAF output values. It determines the maximum extent to which the gradient value is scaled after GAF processing. When α is large, the range of GAF output values ​​is also large, which means that the gradient may be amplified or reduced more after GAF processing. This is usually used to deal with situations where the gradient value is relatively small. When α is small, the range of GAF output values ​​is also small, which means that the gradient is amplified or reduced less after GAF processing. This is usually used to deal with situations where the gradient value is relatively large.

[0067] β primarily controls the sensitivity of the GAF when the gradient approaches zero. It determines the steepness of the GAF curve in regions with small gradients. When β is large, the slope of the GAF is steeper in regions with gradients close to zero. This means that in regions with small gradients, the GAF significantly amplifies the gradient, allowing these small gradients to be updated more quickly, thus preventing gradient vanishing. When β is small, the slope of the GAF is gentler in regions with gradients close to zero, meaning that the amplification effect of the GAF is weaker in regions with small gradients.

[0068] Train a neural network model without GAF and record the maximum gradient. Use visualization to plot the loss curve and compare it with the quadratic equation curve to confirm. Increase the gradient in areas of the loss curve that are flatter than the quadratic curve, and decrease the gradient in areas that are steeper than the quadratic curve.

[0069] After many experiments, it was found that {α=0.1, β=20} and {α=0.2, β=10} are relatively stable, and initial attempts and fine-tuning can be carried out on this basis.

[0070] In S22, the gradient of the GAF algorithm is updated using the loss function, and the formula is:

[0071]

[0072] in, is the gradient of the random training process, ▽ is the gradient symbol, ω is the weight vector, i is the number of iterations, k is the number of data sets, L(·) is the loss function, M(·) is the output of the model for each input sample in each iteration, X (·) is the input sample in the training batch, ω (·) is the weight vector, y (·) is the corresponding label of the input sample, g k is the gradient of the kth iteration, μ m is the momentum coefficient, g k-1 is the gradient of the k-1th iteration, η is the learning rate, g' k The gradient after momentum update is applied to the GAF function to obtain the adjusted gradient, g'(·) is a different type of GAF function, ω k is the weight vector for the kth iteration.

[0073] For different types of GAF, g'(g k ) The corresponding formulas are different and are as follows:

[0074] Arctan-type GAF:g'(g k )=αarctan(βg k )

[0075] Tanh-type GAF:g'(g k )=αtanh(βg k )

[0076] Log-type GAF:g'(g k )=α(ln(ReLU(βg k )+1)-ln(ReLU(-βg k )+1))

[0077] After the above cycle is completed, the original gradient is replaced by the gradient activated by the GAF algorithm.

[0078] The DudeNet denoising model in S22 includes a feature extraction module FEB, an image enhancement module EB, a compression module CB and a reconstruction module RB.

[0079] The feature extraction module FEB includes two parallel first feature extraction modules FEBnet1 and second feature extraction modules FEBnet2, wherein the first feature extraction module FEBnet1 extracts global features and local features from the input noisy image through a sparse mechanism, and the second feature extraction module FEBnet2 is used to extract supplementary features from the input noisy image;

[0080] The first feature extraction module FEBnet1 adopts a sparse mechanism, including a convolutional layer, a batch normalization layer and a ReLU activation layer;

[0081] The output FEB1 of the first feature extraction module FEBnet1 is expressed as:

[0082] FEB1=C(CBR3(S(CBR1(Y))))

[0083] where C(·) is a 3×3 convolution function, CBR3(·) is a module consisting of three convolutional layers, a batch normalization layer, and a ReLU activation layer. CBR3 can include dilated convolutions, S(·) is a sparsifying mechanism, CBR1(·) is a module consisting of one convolutional layer, a batch normalization layer, and a ReLU activation layer, and Y is the input noisy image.

[0084] The second feature extraction module FEBnet2 includes a convolutional layer, a ReLU activation layer and a convolutional layer connected in sequence;

[0085] The output FEB2 of the second feature extraction module FEBnet2 is expressed as:

[0086] FEB2=C1(CR 15 (Y))

[0087] Among them, C1(·) is a 3×3 convolution function, CR 15 (·) is a functional module that executes 15 convolutional layers and ReLU activation layers in sequence.

[0088] The image enhancement module EB includes a first image enhancement module EB1 and a second image enhancement module EB2. The first image enhancement module EB1 is used to fuse the output features of the first feature extraction module FEBnet1 and the second feature extraction module FEBnet2, and process them through a batch normalization layer and a ReLU activation layer; the second image enhancement module EB2 is used to fuse the output features of the first image enhancement module EB1 and the input image, and compress the output features using 1×1 convolution;

[0089] The output OE1 of the first image enhancement module EB1 is expressed as:

[0090] OE1=E(FEB1,FEB2)

[0091] OE1=R(B(CON(C(CBR3(S(CBR1(Y))))),CB(CR 15 (Y)))))

[0092] Among them, E(·) is the fusion operation, R(·) is the activation layer, B(·) is the batch normalization layer, CON(·) is the convolution function, and CB(·) is the compression module;

[0093] The output OE2 of the second image enhancement module EB2 is expressed as:

[0094] OE2=E(OCB2,Y)

[0095] OCB2=C1(OE1)

[0096] Here, OCB2 represents convolution of the output OE1 of the first image enhancement module EB1 to compress data.

[0097] The compression module CB includes a first convolutional layer CB1, a second convolutional layer CB2, and a third convolutional layer CB3. The first convolutional layer CB1 is located at the end of the second feature extraction module FEBnet2, the second convolutional layer CB2 is located between the first image enhancement module EB1 and the second image enhancement module EB2, and the third convolutional layer CB3 is located between the second image enhancement module EB2 and the reconstruction module RB.

[0098] The output of the compression module CB is represented as:

[0099] OCB3=C1(OE2)

[0100] Among them, OCB3 is the residual feature output by the third convolutional layer CB3.

[0101] The reconstruction module RB uses residual connections to map the compressed features to the denoised image, and the output is expressed as:

[0102] X=Y-OCB3

[0103] Where X is the reconstructed clear image.

[0104] The feature extraction module uses a sparsity mechanism to extract diverse features and reduce network depth. The image augmentation module enhances the extracted features by fusing the features of the two sub-networks, which is particularly useful for images contaminated by unknown noise. The compression module compresses the network to reduce computational costs. The reconstruction module is responsible for reconstructing the clean image.

[0105] The optimized model is trained and the dataset images are denoised to obtain images with better denoising effects and the PSNR value is calculated for evaluation.

[0106] PSNR is a measure of image quality calculated by comparing the difference between the original image and the distorted image. Specifically, PSNR is calculated by comparing every pixel value in the two images. If the two images are identical, the noise is zero and the PSNR is infinite. If the two images are significantly different, the noise is high and the PSNR decreases accordingly. Therefore, a higher PSNR indicates better image quality.

[0107] The image quality measurement indicators in S3 are:

[0108]

[0109] Where PSNR is the image quality measure, MAX is the maximum image intensity, MSE is the mean square error, p is the number of rows of image pixels, q is the number of columns of image pixels, r is the value of p, s is the value of q, I(·) is the pixel value of the original input image, and K(·) is the pixel value of the processed image.

[0110] In one embodiment of the present invention, Figure 2 As shown in Table 1, the PSNR differences of different optimizers in the image denoising model are compared.

[0111] Table 1 PSNR differences of different optimizers in image denoising models

[0112]

[0113] Compared to the above experimental cases, the optimizer sgd-alt, optimized using the GAF algorithm, introduces a momentum term. This factor not only considers the current gradient but also the cumulative effect of past gradients during each parameter update. This allows it to escape saddle points during training and find a better optimal solution, effectively alleviating gradient explosion, gradient vanishing, saddle point problems, and other ill-conditioned problems. Compared to the original adam optimizer in denoising tasks, sgd-alt performs more stably and efficiently, achieving superior PSNR (peak signal-to-noise ratio) values ​​for the same image denoising, demonstrating superior denoising results.

[0114] The proposed GAF ​​optimizer provides a unified solution to several common deep learning optimization challenges. The technique supports its claims with theoretical justification and empirical evidence from experiments with various datasets and network architectures. The simplicity of its implementation increases its appeal, suggesting the potential for widespread adoption.

[0115] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the invention.

Claims

1. A remote sensing image denoising method based on a new optimizer, characterized in that: The following steps are involved: S1: Collecting remote sensing noise images and dividing the remote sensing noise images into a training set and a test set; S2: Use the GAF algorithm to perform gradient optimization on the DudeNet denoising model, and use the training set to train the gradient-optimized DudeNet denoising model; The S2 includes the following steps: S21: Using the scaling factor in the GAF algorithm and sensitivity factor Adjust the gradient range of the GAF algorithm; S22: Use the loss function to update the gradient of the GAF algorithm, and replace the gradient of the DudeNet denoising model with the updated gradient of the GAF algorithm to obtain the gradient-optimized DudeNet denoising model; In S22, the gradient of the GAF algorithm is updated using the loss function, and the formula is: ； in, is the gradient of the stochastic training process, is the gradient symbol, is the weight vector, is the number of iterations, is the number of data sets, is the loss function, is the output of the model for each input sample in each iteration, is the input sample in the training batch, is the weight vector, is the corresponding label of the input sample, For the The gradient of the iteration, is the momentum coefficient, For the The gradient of the iteration, is the learning rate, The updated gradient for momentum is applied to the GAF function to obtain the adjusted gradient, For different types of GAF functions, For the The weight vector of the iteration; S23: Use the training set to train the gradient-optimized DudeNet denoising model; S3: Test the trained DudeNet denoising model using the test set and evaluate it using image quality metrics to complete remote sensing image denoising based on the new optimizer.

2. The remote sensing image denoising method based on the novel optimizer according to claim 1, characterized in that: The S21 scaling factor and sensitivity factor Set to and .

3. The remote sensing image denoising method based on a novel optimizer according to claim 1, characterized in that: The DudeNet denoising model in S22 includes a feature extraction module FEB, an image enhancement module EB, a compression module CB and a reconstruction module RB.

4. The remote sensing image denoising method based on the novel optimizer according to claim 3 is characterized in that: The feature extraction module FEB includes two parallel first feature extraction modules FEBnet1 and second feature extraction modules FEBnet2, wherein the first feature extraction module FEBnet1 extracts global features and local features from the input noisy image through a sparse mechanism, and the second feature extraction module FEBnet2 is used to extract supplementary features from the input noisy image; The first feature extraction module FEBnet1 adopts a sparse mechanism, including a convolutional layer, a batch normalization layer and a ReLU activation layer; The output of the first feature extraction module FEBnet1 Expressed as: ； in, for Convolution function, is a module consisting of three convolutional layers, a batch normalization layer, and a ReLU activation layer. is a sparse mechanism, is a module consisting of a convolutional layer, a batch normalization layer, and a ReLU activation layer. is the input noise image; The second feature extraction module FEBnet2 includes a convolutional layer, a ReLU activation layer and a convolutional layer connected in sequence; The output of the second feature extraction module FEBnet2 Expressed as: ； in, for Convolution function, It is a functional module that executes 15 convolutional layers and ReLU activation layers in sequence.

5. The remote sensing image denoising method based on the novel optimizer according to claim 4 is characterized in that: The image enhancement module EB includes a first image enhancement module EB1 and a second image enhancement module EB2. The first image enhancement module EB1 is used to fuse the output features of the first feature extraction module FEBnet1 and the second feature extraction module FEBnet2, and process them through a batch normalization layer and a ReLU activation layer; the second image enhancement module EB2 is used to fuse the output features of the first image enhancement module EB1 and the input image, and compress the output features using 1×1 convolution; The output of the first image enhancement module EB1 Expressed as: ； ； in, represents the fusion operation, is the activation layer, is the batch normalization layer, is the convolution function, For compression module; The output of the second image enhancement module EB2 Expressed as: ； ； in, Represents the output of the first image enhancement module EB1 Convolution is performed to compress the data.

6. The remote sensing image denoising method based on a novel optimizer according to claim 5, characterized in that: The compression module CB includes a first convolutional layer CB1, a second convolutional layer CB2, and a third convolutional layer CB3. The first convolutional layer CB1 is located at the end of the second feature extraction module FEBnet2, the second convolutional layer CB2 is located between the first image enhancement module EB1 and the second image enhancement module EB2, and the third convolutional layer CB3 is located between the second image enhancement module EB2 and the reconstruction module RB. The output of the compression module CB is represented as: ； in, It is the residual feature output by the third convolutional layer CB3.

7. The remote sensing image denoising method based on a novel optimizer according to claim 6, characterized in that: The reconstruction module RB uses residual connections to map the compressed features to the denoised image, and the output is expressed as: ； in, Reconstructed clear image.

8. The remote sensing image denoising method based on a novel optimizer according to claim 1, characterized in that: The image quality measurement indicators in S3 are: ； ； in, is an indicator of image quality. is the maximum value of the image intensity, is the mean square error, is the number of rows of image pixels, is the number of columns of image pixels, for The value of for The value of is the pixel value of the original input image, is the pixel value of the processed image.

Images