SAR image despeckling and denoising method based on DnCNN pre-training model

Abstract

The application provides a SAR image despeckling and denoising method based on a DnCNN pre-training model, and relates to the technical field of image processing. The method comprises the following steps: obtaining a noise SAR image to be processed, inputting the SAR image to be processed into a constructed image preprocessing model, and obtaining a preprocessed image dataset; inputting the preprocessed image dataset into a DnCNN pre-training model, and obtaining a clean SAR image; the construction method of the image preprocessing model is as follows: obtaining an initial SAR image set and calculating a corresponding initial intensity map; adding speckle noise to the initial intensity map according to the view number, and obtaining a noise image set; performing logarithmic transformation on the noise image set, and obtaining a logarithmic intensity map; determining a normalization factor according to the view number and adding the normalization factor to the logarithmic intensity map, and obtaining a converted image set; constructing the image preprocessing model based on a CNN network architecture, taking the initial SAR image set as input, and taking the converted image set as input. The application simplifies the process of using the DnCNN pre-training model to denoise the SAR image, and solves the problem of excessively high complexity.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a SAR image despeckling and denoising method based on a DnCNN pre-training model. BACKGROUND

[0002] SAR (Synthetic Aperture Radar) images have very important applications in environmental monitoring, geological exploration, military reconnaissance and other fields. However, the speckle noise in SAR images greatly affects the clarity of the images and the accuracy of information processing, bringing many inconveniences to the subsequent links. Therefore, for a long time, how to eliminate speckle noise and improve the quality of SAR images has been an important research topic in the relevant field.

[0003] In recent years, methods based on deep learning have been widely used in the field of natural image denoising. In particular, Convolutional Neural Networks (CNN) have a very deep architecture, a large capacity and a high flexibility to represent image features, which are very suitable for image restoration. By designing a special network structure and training a large amount of image data, the CNN method can adaptively learn complex noise patterns, thus achieving a high-efficiency denoising effect that other methods cannot achieve. Despeckling is a long-standing problem in synthetic aperture radar (SAR) imaging. Many different SAR image restoration schemes have been proposed. Among various possible strategies, the method based on Convolutional Neural Networks (CNN) has been proven to produce the best SAR image restoration results.

[0004] However, in order to learn an effective model, a large number of speckle-free SAR images need to be collected in advance for training, which undoubtedly consumes a lot of time and effort.

[0005] The DnCNN model and the MuLoG method are combined together, so that the DnCNN pre-training model is used to eliminate the multiplicative speckle noise in the SAR image, which well achieves the expected goal and obtains excellent results.

[0006] However, learning to use the MuLoG-DnCNN model, especially how to implement the pre-processing of the SAR image, still requires a certain amount of time and effort, even for those skilled in the art. In addition, when using remote sensing data for analysis and processing, a large number of SAR images are often required, and it is not realistic to preprocess them one by one, and a method is needed to batch process the SAR images. SUMMARY

[0007] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a SAR image despeckling and denoising method based on a DnCNN pre-training model, which solves the problem of high cost of obtaining a SAR image training set in the prior art.

[0008] In order to achieve the above-mentioned purpose, the present application provides the following solutions.

[0009] A SAR image despeckling and denoising method based on a DnCNN pre-training model comprises:

[0010] Obtain a noise SAR image to be processed, and input the SAR image to be processed into a constructed image preprocessing model to obtain a preprocessed image data set;

[0011] Input the preprocessed image data set into a DnCNN pre-training model to obtain a clean SAR image;

[0012] The construction method of the image preprocessing model is:

[0013] Obtain an initial SAR image set and calculate a corresponding initial intensity map;

[0014] Add speckle noise to the initial intensity map according to the view number to obtain a noise map set;

[0015] Logarithmically transform the noise map set to obtain a log intensity map;

[0016] Determine a normalization factor according to the view number and add it to the log intensity map to obtain a converted image set;

[0017] Based on a CNN network architecture, an image preprocessing model is constructed by taking the initial SAR image set as input and the converted image set as input.

[0018] Preferably, the relationship expression of the image and the noise of the initial SAR image set is:

[0019] Y=N*X,

[0020] Wherein Y is an image in the initial SAR image set, N is speckle noise, and X is a clean image.

[0021] Preferably, the expression of the clean image after logarithmic transformation in the log intensity map is:

[0022]

[0023] Wherein f is a DnCNN pre-training model, L is a view number, and ψ is a digamma function.

[0024] Preferably, the step of determining a normalization factor according to the view number and adding it to the log intensity map comprises:

[0025] The input variance is calculated;

[0026] The input method is approximated to obtain an approximate variance;

[0027] A normalization factor is obtained according to the approximate variance.

[0028] Preferably, the calculation expression of the input variance is:

[0029]

[0030] wherein σ * is the input variance, Y min is the 0.3% quantile of Y, and Y max is the 99.7% quantile of Y.

[0031] Preferably, the expression of the approximate variance is:

[0032] σ = floor(σ * / 5) * 5;

[0033] wherein σ is the approximate variance.

[0034] Preferably, the expression of the normalization factor is:

[0035]

[0036] wherein NF is the normalization factor.

[0037] The present application discloses the following technical effects:

[0038] The present application provides a SAR image despeckling and denoising method based on a DnCNN pre-training model, comprising: obtaining a noise SAR image to be processed, and inputting the SAR image to be processed into a constructed image preprocessing model to obtain a preprocessed image dataset; inputting the preprocessed image dataset into a DnCNN pre-training model to obtain a clean SAR image; the construction method of the image preprocessing model is: obtaining an initial SAR image set and calculating a corresponding initial intensity map; adding speckle noise to the initial intensity map according to the view number to obtain a noise image set; performing logarithmic transformation on the noise image set to obtain a logarithmic intensity map; determining a normalization factor according to the view number and adding it to the logarithmic intensity map to obtain a converted image set; constructing an image preprocessing model based on a CNN network architecture, taking the initial SAR image set as input and taking the converted image set as input. The present application greatly reduces the learning cost required by using the MuLoG-DnCNN method by constructing an image preprocessing model, quickly obtains a preprocessed SAR image data, and conveniently and quickly uses the DnCNN model to denoise to obtain a relatively clean SAR image. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0040] Figure 1 A flow chart of a SAR image despeckling and denoising method based on a DnCNN pre-training model is provided for the embodiments of the present application.

[0041] Figure 2 A detailed schematic diagram of a SAR image despeckling and denoising method based on a DnCNN pre-training model is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0043] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0044] As shown in Figure 1 The present application provides a SAR image despeckling and denoising method based on a DnCNN pre-training model, which comprises:

[0045] Step 100: obtaining a noise SAR image to be processed, and inputting the SAR image to be processed into a constructed image pre-processing model to obtain a pre-processed image data set;

[0046] Step 200: inputting the pre-processed image data set into a DnCNN pre-training model to obtain a clean SAR image;

[0047] The construction method of the image pre-processing model is:

[0048] Step 101: obtaining an initial SAR image set and calculating a corresponding initial intensity map;

[0049] Step 102: adding speckle noise to the initial intensity map according to the view number to obtain a noise image set;

[0050] Step 103: Logarithmic transformation is performed on the noise set to obtain a log intensity map;

[0051] Step 104: A normalization factor is determined according to the visual number and added to the log intensity map to obtain a converted image set;

[0052] Step 105: Based on the CNN network architecture, an image preprocessing model is constructed by taking the initial SAR image set as input and the converted image set as input.

[0053] As Figure 2 shown, the embodiment of the present application also provides a detailed process of a SAR image despeckling and denoising method based on a DnCNN pre-training model:

[0054] Step 1: Prepare noise images in the training set data

[0055] 400 clean SAR image files or ordinary gray-scale images are obtained; the intensity map of the image is calculated, and speckle noise is added to the image in turn with a visual number L of 1-5; the noise images are saved, and a total of 400*5=2000 images are obtained;

[0056] Step 2: Convert the data distribution and obtain the converted image

[0057] The saved 2000 noise images are read; logarithmic transformation is performed on the image to convert the multiplicative speckle noise into additive noise. This step is because the DnCNN pre-training model is trained for additive noise, and the multiplicative noise must be converted into additive noise to apply the DnCNN pre-training model; the original relationship between the image and the noise (the relationship between the images in the initial SAR image set and the noise) is as follows: Y=N×X, where Y is the noisy image, N is the speckle noise, and X is the clean image. After logarithmic transformation, the composition of the clean image X (the clean image after logarithmic transformation in the log intensity map) becomes: where f is the DnCNN pre-training model, L is the visual number, and ψ is the digamma function. The log-transformed image is normalized. The purpose of this step is to reduce the heavy-tailed effect. The normalization formula is as follows: Input=(Y-Y min ) / (Y max -Y min ), where Y min takes the 0.3% quantile of Y, and Y max takes the 99.7% quantile of Y. Finally, a normalization factor NF is added to obtain the converted image, and the steps for calculating NF are as follows: first, assuming that the input variance of the homogeneous region is similar to the variance of the training data set, the input variance σ * is calculated: Since the pre-trained DnCNN model has a sigma range of 10-75 with a step of 5, we need to use the floor function to approximate the calculated variance to meet the pre-trained model range: sigma = floor(sigma / 5) * 5, finally, the normalized factor NF is calculated and added to the image: * / 5)×5, finally, the normalized factor NF is calculated and added to the image: Save the converted image as the test set, a total of 400*5 = 2000.

[0058] Step 3: Effect verification

[0059] The converted image is directly input into the DnCNN pre-trained model to obtain the denoised image; the denoised image is inversely normalized to obtain the original size; the image is debiased; the final denoised image is obtained by inverse logarithmic transformation and rounding; PSNR, SSIM and ENL are calculated to verify the denoising effect.

[0060] Step 4: Train CNN model

[0061] The noise image in step 1 is input into the model, and the converted image in step 2 is output from the model; a CNN network architecture with convolutional layers and pooling layers is constructed; the training set images are processed (a total of 2000+2000 = 4000), and the model file is obtained; save the model file trainedModel.mat.

[0062] Step 5: Effect verification

[0063] Another part of the noise image is input into the CNN model to obtain the converted image, and the effect is verified by repeating step 3.

[0064] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between each embodiment can be referred to each other.

[0065] The principles and implementation modes of the present application are described by applying specific examples in this paper. The above embodiment description is only used to help understand the method and core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A SAR image speckle removal and noise reduction method based on a DnCNN pre-trained model, characterized in that, include: Acquire a noisy SAR image to be processed, and input the SAR image to be processed into a pre-constructed image preprocessing model to obtain a pre-processed image dataset; The preprocessed image dataset is input into the DnCNN pre-trained model to obtain a clean SAR image; The method for constructing the image preprocessing model is as follows: Acquire the initial SAR image set and calculate the corresponding initial intensity map; Based on the number of views, speckle noise is added to the initial intensity map to obtain a noise map set; The noise dataset is logarithmically transformed to obtain a logarithmic intensity map; The normalization factor is determined based on the number of views and added to the logarithmic intensity map to obtain the converted image set; Based on the CNN network architecture, an image preprocessing model is constructed with the initial SAR image set as input and the transformed image set as output. The step of determining the normalization factor based on the number of views and adding it to the logarithmic intensity map includes: The input variance was calculated. The input variance is approximated to obtain the approximate variance; The normalization factor is obtained based on the approximate variance; The expression for calculating the input variance is: ； in, To input the variance, The 0.3% quantile of Y It is the 99.7th percentile of Y; The expression for the approximate variance is: ； in, This is an approximate variance; The expression for the normalization factor is: ； Where NF is the normalization factor.

2. The SAR image speckle removal and noise reduction method based on a DnCNN pre-trained model according to claim 1, characterized in that, The relationship between images and noise in the initial SAR image set is expressed as follows: ， Where Y is the initial SAR image set, N is speckle noise, and X is the clean image.

3. The SAR image speckle removal and noise reduction method based on a DnCNN pre-trained model according to claim 2, characterized in that, The expression for the clean image in the logarithmic intensity map after logarithmic transformation is: ； Where f is the DnCNN pre-trained model, and L is the number of views. It is the digamma function.

Images