Image processing method and system based on big data

Abstract

The invention provides an image processing method and system based on big data. The method comprises the following steps: performing deep denoising on a sample image to obtain a deep denoised image; performing pixel point denoising on the sample image to obtain an image after pixel point denoising; fusing the images after the deep denoising and the pixel point denoising; taking the fused image and the sample image as training data; inputting the training data into a convolutional neural network for training to obtain a fusion denoising model; and de-noising an image to be processed by using the fusion de-noising model to obtain a de-noised image. According to the method, the sample image subjected to deep denoising and the sample image subjected to pixel point denoising are fused to obtain the training sample, the training sample is trained to obtain the fused denoising model, and finally the sample image is denoised by using the fused denoising model, so that noise in the sample image can be effectively eliminated, and the accuracy of denoising is improved. Moreover, the geometric detail information representation capability of the image can be enhanced, and the image authenticity is improved.

Description

Technical Field

[0001] This invention relates to the field of image denoising technology, and more specifically, to an image processing method and system based on big data. Background Technology

[0002] In modern society, digital image processing has become an important task. However, traditional image processing methods often have limitations in terms of processing speed, accuracy, and scalability. Therefore, there is a need for image processing methods based on large datasets to meet the ever-increasing demands for image processing. Existing methods for denoising sample images generally rely on grayscale values. However, since the grayscale thresholds in sample images are usually set manually, methods that directly rely on the grayscale features of the image cannot effectively remove noise. Summary of the Invention

[0003] The purpose of this invention is to provide an image processing method and system based on big data, which aims to solve the problem that existing image denoising methods cannot effectively remove noise.

[0004] A big data-based image processing method includes:

[0005] Step 1: Obtain sample images;

[0006] Step 2: Perform depth denoising on the sample image to obtain a depth-denoised image;

[0007] Step 3: Perform pixel-by-pixel denoising on the sample image to obtain the denoised image;

[0008] Step 4: Fuse the depth-denoised image and the pixel-denoised image to obtain the fused image;

[0009] Step 5: Use the fused image and the sample image as training data; the training data includes a target set and an input set, with the fused image as the target set and the sample image as the input set;

[0010] Step 6: Input the training data into a convolutional neural network for training to obtain a fusion and denoising model;

[0011] Step 7: Use the fusion denoising model to denoise the image to be processed to obtain the denoised image.

[0012] Preferably, step 2: performing depth denoising on the sample image to obtain a depth-denoised image includes:

[0013] Step 2.1: Construct a depth denoising threshold based on the size of the sample image; wherein, the depth denoising threshold is:

[0014]

[0015] Among them, w j Denotes the j-th wavelet coefficient. The mean of the wavelet coefficients, median|w j | represents the median of the wavelet coefficients, λ represents the depth denoising threshold, M represents the length of the sample image, N represents the width of the sample image, and L represents the decomposition scale of the sample image.

[0016] Step 2.2: Construct a depth denoising model based on the depth denoising threshold;

[0017] Step 2.3: Use the depth denoising model to perform depth denoising on the sample image to obtain the depth-denoised image.

[0018] Preferably, the deep denoising model includes:

[0019]

[0020] Where sign is the sign function and a is a preset coefficient. This represents the j-th wavelet coefficient at the i-th decomposition scale. This represents the wavelet coefficients after deep denoising.

[0021] Preferably, step 3: performing pixel-by-pixel denoising on the sample image to obtain a pixel-by-pixel denoised image includes:

[0022] Formula used:

[0023]

[0024] The sample image is denoised pixel by pixel to obtain the denoised image; where f(a,b) represents the pixel value of pixel (a,b) after pixel denoising, D is a user-defined coefficient, x(a,b) represents the pixel value of pixel (a,b) in the sample image, mean(a,b) represents the pixel mean of the sample image, and σ x (a,b) represents the variance of pixel values ​​within the sample image.

[0025] Preferably, step 4: fusing the depth-denoised image and the pixel-denoised image to obtain a fused image includes:

[0026] Step 4.1: Divide the depth-denoised image and the pixel-denoised image into image blocks of the same size;

[0027] Step 4.2: Calculate the saliency value for each image patch;

[0028] Step 4.3: Remove the corresponding image blocks whose significance values ​​are less than a preset threshold to obtain the image blocks to be fused;

[0029] Step 4.4: Calculate the weights of the image patches to be fused;

[0030] Step 4.5: Use the weights to linearly weight the image blocks to be fused to obtain the fused image.

[0031] Preferably, the formula for calculating the significance value is:

[0032] λ=exp((1-m d )×CV×H)-1

[0033] Where λ represents the significance value, m d represents the average depth value of the corresponding image block, CV represents the preset weight, and H represents the depth frequency entropy.

[0034] Preferably, the formula for calculating the weights of the image blocks to be fused is:

[0035]

[0036] Among them, SQ i Let N(p) represent the weight of the i-th image patch, and let W(p) represent the set of eight neighboring pixels of pixel p. pq c represents the correlation between adjacent pixels p and q. p This represents the L*a*b value of pixel p, c q Let d represent the L*a*b value of pixel q. p d represents the depth value of pixel p. q σ represents the depth value of pixel q. 2 S is the nearest neighbor parameter. t B represents the target region of an image patch. t This represents the background area of ​​an image block.

[0037] Preferably, step 4.5: using the weights to linearly weight the image blocks to be fused to obtain the fused image, includes:

[0038] Formula used:

[0039]

[0040] The image blocks to be fused are linearly weighted to obtain the fused image; where M i Let represent the i-th image patch, and Norm represent the normalization function.

[0041] Preferably, step 6: inputting the training data into a convolutional neural network for training to obtain a fusion denoising model includes:

[0042] Step 6.1: Construct the loss function of the convolutional neural network using the target set and the input set; wherein the loss function is:

[0043]

[0044] Where n represents the number of training samples, θ represents the set of convolution kernel weights and neural network biases, and F(Y) i ;θ) represents the image block Y processed by the convolutional neural network. i The actual output, X i Indicates Y i The corresponding target output block;

[0045] Step 6.2: Optimize and train the loss function using the network parameter optimization model to obtain the fusion denoising model; wherein, the network parameter optimization model is:

[0046]

[0047] Where, Δ i W represents the convolution kernel weights after the i-th iteration. i l Let represent the bias value of the l-th layer of the neural network, and η represent the learning rate. This represents the partial derivative of the loss function with respect to the convolution kernel weights.

[0048] The present invention also provides an image processing system based on big data, comprising:

[0049] The sample acquisition module is used to acquire sample images;

[0050] A depth denoising module is used to perform depth denoising on the sample image to obtain a depth-denoised image;

[0051] A pixel-level denoising module is used to perform pixel-level denoising on the sample image to obtain a pixel-level denoised image;

[0052] The fusion module is used to fuse the depth-denoised image and the pixel-denoised image to obtain a fused image;

[0053] A training data construction module is used to use the fused image and the sample image as training data; the training data includes a target set and an input set, with the fused image as the target set and the sample image as the input set.

[0054] The training module is used to input the training data into a convolutional neural network for training to obtain a fusion and denoising model;

[0055] The denoising module is used to denoise the image to be processed using the fusion denoising model to obtain a denoised image.

[0056] The beneficial effects of the image processing method and system based on big data provided by this invention are as follows: Compared with the prior art, this invention obtains training samples by fusing deep denoised sample images and pixel-denoised sample images, and inputs the training samples into a convolutional neural network for training to obtain a fusion denoising model. Finally, the fusion denoising model is used to denoise the sample images, which can not only effectively eliminate noise in the sample images, but also enhance the geometric detail information representation ability of the images and improve the image realism.

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 A flowchart of an image processing method based on big data provided by an embodiment of the present invention is shown;

[0060] Figure 2 The diagram illustrates the principle of an image processing system based on big data, as provided in an embodiment of the present invention. Detailed Implementation

[0061] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0062] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0063] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0064] The purpose of this invention is to provide an image processing method and system based on big data, which aims to solve the problem that existing image denoising methods cannot effectively remove noise.

[0065] Please see Figure 1 A big data-based image processing method includes:

[0066] Step 1: Obtain sample images;

[0067] Step 2: Perform depth denoising on the sample image to obtain a depth-denoised image;

[0068] Wavelet transform is a common data processing technique. After decomposing an image, wavelet coefficients of different sizes are generated. Some of these coefficients contain noise, while others do not. Choosing an appropriate denoising threshold allows for image denoising. The specific denoising process is as follows:

[0069] Furthermore, step 2 includes:

[0070] Step 2.1: Construct a depth denoising threshold based on the size of the sample image; wherein, the depth denoising threshold is:

[0071]

[0072] Among them, w j Denotes the j-th wavelet coefficient. The mean of the wavelet coefficients, median|w j | represents the median of the wavelet coefficients, λ represents the depth denoising threshold, M represents the length of the sample image, N represents the width of the sample image, and L represents the decomposition scale of the sample image.

[0073] Step 2.2: Construct a depth denoising model based on the depth denoising threshold; the depth denoising model includes:

[0074]

[0075] Where sign is the sign function and a is a preset coefficient. This represents the j-th wavelet coefficient at the i-th decomposition scale. This represents the wavelet coefficients after deep denoising.

[0076] Step 2.3: Use the depth denoising model to perform depth denoising on the sample image to obtain the depth-denoised image.

[0077] Step 3: Perform pixel-by-pixel denoising on the sample image to obtain the denoised image;

[0078] In this invention, step 3 includes:

[0079] Formula used:

[0080]

[0081] The sample image is denoised pixel by pixel to obtain the denoised image; where f(a,b) represents the pixel value of pixel (a,b) after pixel denoising, D is a user-defined coefficient, x(a,b) represents the pixel value of pixel (a,b) in the sample image, mean(a,b) represents the pixel mean of the sample image, and σ x (a,b) represents the variance of pixel values ​​within the sample image.

[0082] Step 4: Fuse the depth-denoised image and the pixel-denoised image to obtain the fused image;

[0083] Furthermore, step 4 includes:

[0084] Step 4.1: Divide the depth-denoised image and the pixel-denoised image into image blocks of the same size;

[0085] Step 4.2: Calculate the saliency value for each image patch; Before fusing the image patches, this invention requires evaluating the quality of each image patch (calculating the saliency value) to enrich the information contained in each image patch. The saliency value calculation formula in this invention is:

[0086] λ=exp((1-m d )×CV×H)-1

[0087] Where λ represents the significance value, m drepresents the average depth value of the corresponding image block, CV represents the preset weight, and H represents the depth frequency entropy.

[0088] Step 4.3: Remove the corresponding image blocks whose significance values ​​are less than a preset threshold to obtain the image blocks to be fused;

[0089] Step 4.4: Calculate the weights of the image blocks to be fused; wherein the formula for calculating the weights of the image blocks to be fused is:

[0090]

[0091] Among them, SQ i Let N(p) represent the weight of the i-th image patch, and let W(p) represent the set of eight neighboring pixels of pixel p. pq c represents the correlation between adjacent pixels p and q. p This represents the L*a*b value of pixel p, c q Let d represent the L*a*b value of pixel q. p d represents the depth value of pixel p. q σ represents the depth value of pixel q. 2 S is the nearest neighbor parameter. t B represents the target region of an image patch. t This represents the background area of ​​an image block.

[0092] Step 4.5: Apply the weights to the image blocks to be fused using linear weighting to obtain the fused image. The linear weighting formula is as follows:

[0093]

[0094] The image blocks to be fused are linearly weighted to obtain the fused image; where M i Let represent the i-th image patch, and Norm represent the normalization function.

[0095] This invention uses weighted fusion of image blocks according to their weights to alleviate image smoothing problems, encourage feature reuse, compensate for the lack of detail in image-level fusion, and promote complementary advantages in the entire fusion process, making the denoised image more realistic and free from distortion.

[0096] Step 5: Use the fused image and the sample image as training data; the training data includes a target set and an input set, with the fused image as the target set and the sample image as the input set;

[0097] Step 6: Input the training data into a convolutional neural network for training to obtain a fusion and denoising model;

[0098] Furthermore, step 6 includes:

[0099] Step 6.1: Construct the loss function of the convolutional neural network using the target set and the input set; wherein the loss function is:

[0100]

[0101] Where n represents the number of training samples, θ represents the set of convolution kernel weights and neural network biases, and F(Y) i ;θ) represents the image block Y processed by the convolutional neural network. i The actual output, X i Indicates Y i The corresponding target output block;

[0102] Step 6.2: Optimize and train the loss function using the network parameter optimization model to obtain the fusion denoising model; wherein, the network parameter optimization model is:

[0103]

[0104] Where, Δ i W represents the convolution kernel weights after the i-th iteration. i l Let represent the bias value of the l-th layer of the neural network, and η represent the learning rate. This represents the partial derivative of the loss function with respect to the convolution kernel weights.

[0105] Furthermore, the convolutional neural network of the present invention includes a first convolutional layer and a second reconstruction convolutional layer; wherein, the first convolutional layer has 64 convolutional kernels and the kernel size is 9*9; the second reconstruction convolutional layer has 32 convolutional kernels and the kernel size is 5*5.

[0106] Step 7: Use the fusion denoising model to denoise the image to be processed to obtain the denoised image.

[0107] This invention utilizes a convolutional neural network to fit the mapping relationship between noisy and denoised images, while optimizing the error between the pixel values ​​of the actual output image and the pixel values ​​of the corresponding target output block as the objective function, thereby maximizing the removal of noise from the image.

[0108] Please see Figure 2 The present invention also provides an image processing system based on big data, comprising:

[0109] The sample acquisition module is used to acquire sample images;

[0110] A depth denoising module is used to perform depth denoising on the sample image to obtain a depth-denoised image;

[0111] A pixel-level denoising module is used to perform pixel-level denoising on the sample image to obtain a pixel-level denoised image;

[0112] The fusion module is used to fuse the depth-denoised image and the pixel-denoised image to obtain a fused image;

[0113] A training data construction module is used to use the fused image and the sample image as training data; the training data includes a target set and an input set, with the fused image as the target set and the sample image as the input set.

[0114] The training module is used to input the training data into a convolutional neural network for training to obtain a fusion and denoising model;

[0115] The denoising module is used to denoise the image to be processed using the fusion denoising model to obtain a denoised image.

[0116] This invention obtains training samples by fusing depth-denoised sample images and pixel-denoised sample images, and then inputs the training samples into a convolutional neural network for training to obtain a fusion denoising model. Finally, the fusion denoising model is used to denoise the sample images, which can not only effectively eliminate noise in the sample images, but also enhance the geometric detail information representation ability of the images and improve the image realism.

[0117] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An image processing method based on big data, characterized in that, include: Step 1: Obtain sample images; Step 2: Perform depth denoising on the sample image to obtain a depth-denoised image; Step 3: Perform pixel-by-pixel denoising on the sample image to obtain the denoised image; Step 4: Fuse the depth-denoised image and the pixel-denoised image to obtain the fused image; Step 5: Use the fused image and the sample image as training data; the training data includes a target set and an input set, with the fused image as the target set and the sample image as the input set; Step 6: Input the training data into a convolutional neural network for training to obtain a fusion and denoising model; Step 7: Use the fusion denoising model to denoise the image to be processed to obtain the denoised image.

2. The image processing method based on big data according to claim 1, characterized in that, Step 2: Perform depth denoising on the sample image to obtain a depth-denoised image, including: Step 2.1: Construct a depth denoising threshold based on the size of the sample image; wherein, the depth denoising threshold is: Among them, w j Denotes the j-th wavelet coefficient. The mean of the wavelet coefficients, median|w j | represents the median of the wavelet coefficients, λ represents the depth denoising threshold, M represents the length of the sample image, N represents the width of the sample image, and L represents the decomposition scale of the sample image. Step 2.2: Construct a depth denoising model based on the depth denoising threshold; Step 2.3: Use the depth denoising model to perform depth denoising on the sample image to obtain the depth-denoised image.

3. The image processing method based on big data according to claim 2, characterized in that, The deep denoising model includes: Where sign is the sign function and a is a preset coefficient. This represents the j-th wavelet coefficient at the i-th decomposition scale. This represents the wavelet coefficients after deep denoising.

4. The image processing method based on big data according to claim 3, characterized in that, Step 3: Perform pixel-by-pixel denoising on the sample image to obtain a denoised image, including: Formula used: The sample image is pixel point denoised to obtain an image after pixel point denoising; wherein f(a,b) represents a pixel value of a pixel point (a,b) after pixel point denoising, D is a self-defined coefficient, x(a,b) represents a pixel value of a pixel point (a,b) of the sample image, mean(a,b) represents a pixel mean value of the sample image, and σ x (a,b) represents a variance of a pixel value in the sample image.

5. The image processing method based on big data according to claim 4, characterized in that, Step 4: Fusing the depth-denoised image and the pixel-denoised image to obtain a fused image, including: Step 4.1: Divide the depth-denoised image and the pixel-denoised image into image blocks of the same size; Step 4.2: Calculate the saliency value for each image patch; Step 4.3: Remove the corresponding image blocks whose significance values ​​are less than a preset threshold to obtain the image blocks to be fused; Step 4.4: Calculate the weights of the image patches to be fused; Step 4.5: Use the weights to linearly weight the image blocks to be fused to obtain the fused image.

6. The image processing method based on big data according to claim 5, characterized in that, The formula for calculating the significance value is: λ=exp((1-m d )×CV×H)-1 Where λ represents the significance value, m d represents the average depth value of the corresponding image patch, CV represents the preset weight, and H represents the depth frequency entropy.

7. The image processing method based on big data according to claim 6, characterized in that, The formula for calculating the weight of the image blocks to be fused is: Among them, SQ i Let N(p) represent the weight of the i-th image patch, and let W(p) represent the set of eight neighboring pixels of pixel p. pq c represents the correlation between adjacent pixels p and q. p This represents the L*a*b value of pixel p, c q Let d represent the L*a*b value of pixel q. p d represents the depth value of pixel p. q σ represents the depth value of pixel q. 2 S is the nearest neighbor parameter. t B represents the target region of an image patch. t This represents the background area of ​​an image block.

8. The image processing method based on big data according to claim 7, characterized in that, Step 4.5: Using the weights, linearly weight the image blocks to be fused to obtain the fused image, including: Formula used: The image patches to be fused are linearly weighted to obtain the fused image; where M i Let represent the i-th image patch, and Norm represent the normalization function.

9. The image processing method based on big data according to claim 8, characterized in that, Step 6: Inputting the training data into a convolutional neural network for training to obtain a fusion denoising model, including: Step 6.1: Construct the loss function of the convolutional neural network using the target set and the input set; wherein the loss function is: Where n represents the number of training samples, θ represents the set of convolution kernel weights and neural network biases, and F(Y) i ;θ) represents the image patch Y processed by the convolutional neural network. i The actual output, X i Indicates Y i The corresponding target output block; Step 6.2: Optimize and train the loss function using the network parameter optimization model to obtain the fusion denoising model; wherein, the network parameter optimization model is: Where, Δ i W represents the convolution kernel weights after the i-th iteration. i l Let represent the bias value of the l-th layer of the neural network, and η represent the learning rate. This represents the partial derivative of the loss function with respect to the convolution kernel weights.

10. An image processing system based on big data, characterized in that, include: The sample acquisition module is used to acquire sample images. A depth denoising module is used to perform depth denoising on the sample image to obtain a depth-denoised image; A pixel-level denoising module is used to perform pixel-level denoising on the sample image to obtain a pixel-level denoised image; The fusion module is used to fuse the depth-denoised image and the pixel-denoised image to obtain a fused image; A training data construction module is used to use the fused image and the sample image as training data; the training data includes a target set and an input set, with the fused image as the target set and the sample image as the input set. The training module is used to input the training data into a convolutional neural network for training to obtain a fusion and denoising model; The denoising module is used to denoise the image to be processed using the fusion denoising model to obtain a denoised image.

Images