Image denoising method and system, storage medium and electronic equipment

By performing block processing and feature analysis on the image, dynamically selecting the denoising method, combined with traditional and neural network technology, the balance between cost and performance is achieved, and the high cost and insufficient performance of image denoising in the existing technology is solved. Problem, providing flexible denoising solutions.

CN118967497APending Publication Date: 2024-11-15SENSLAB INC
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411007214.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing image denoising technology is costly and consumes high computing resources when improving performance. The traditional method and AI neural network method have their own limitations, making it difficult to find a balance between cost and performance.

Method used

By blocking the input image, the characteristics of each image are analyzed, such as noise standard deviation, frequency components and local contrast, and dynamically select traditional methods or neural network denoising methods with different parameter quantities to realize denoising processing, and through the image Transitional zone fusion processing merges the noise denoising results.

Benefits of technology

While reducing costs, it provides better image denoising performance, flexibly adapts to different image features and noise levels, and improves the denoising effect and efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118967497A_ABST
    Figure CN118967497A_ABST
Patent Text Reader

Abstract

The invention relates to the field of image processing, and provides an image denoising method and system, a storage medium and electronic equipment, and the method comprises the steps: carrying out the blocking processing of an input image, and obtaining N first block images with different sizes; performing feature analysis on the first block image to obtain a target image feature; determining a denoising algorithm of the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm; denoising the first block-shaped image by adopting a denoising algorithm of the target image features to obtain a second block-shaped image; and combining the N second block images to obtain a target image. According to the invention, on the basis of the scene of image segmentation, a novel denoising method combining the traditional denoising technology and the denoising advantages of the neural network or a denoising method based on different neural networks is used, so that the cost is reduced, and effective image denoising performance is provided at the same time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to an image denoising method, system, storage medium and electronic device. Background Art

[0002] In the field of image denoising technology, although traditional frequency domain and spatial domain processing methods can reduce image noise and retain detail texture to a certain extent, their performance is poor in generalization, poor adaptability, insufficient learning ability, poor real-time processing ability, limited detail retention ability, poor adaptability to noise types, and poor adaptive noise removal ability compared to artificial intelligence (AI)-based neural network denoising methods. AI neural network denoising technology can remove noise more effectively while retaining more image details and texture information, but this performance improvement is usually accompanied by higher costs, including greater storage space requirements and greater computing resource consumption.

[0003] The performance of a neural network is closely related to its parameter count. The more parameters a network has, the stronger its potential denoising ability is. After careful training and optimization, these networks show better denoising effects. However, neural networks with high parameter counts may face deployment and computational challenges in practical applications.

[0004] Most existing image denoising solutions use traditional denoising technology or AI neural network denoising technology to directly perform denoising on the input image, which results in unclear denoising effect or high denoising cost. Therefore, how to achieve better denoising effect while keeping costs down? Especially when the image is processed in blocks, the diversity of image content and noise level requires denoising methods to be able to process them in a targeted manner in order to find a balance between cost and performance. Summary of the invention

[0005] The object of the present invention is to provide an image denoising method, system, storage medium and electronic device that can achieve cost reduction while providing effective image denoising performance.

[0006] Technical solution: The present invention provides an image denoising method, comprising:

[0007] The input image is processed into blocks to obtain N first block images with different sizes;

[0008] Performing feature analysis on the first block image to obtain target image features;

[0009] Determining a denoising algorithm for the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm;

[0010] De-noising the first block image using a denoising algorithm based on target image features to obtain a second block image;

[0011] The N second block images are merged to obtain a target image.

[0012] Furthermore, the target image feature is a noise standard deviation of the first block image, and the feature analysis is performed on the first block image to obtain the target image feature, including the steps of:

[0013] A noise level analysis is performed on the first block image to obtain a noise standard deviation of the first block image.

[0014] Furthermore, the target image feature is a frequency component of the first block image, and feature analysis is performed on the first block image to obtain the target image feature, including the steps of:

[0015] A noise distribution analysis is performed on the first block image to obtain a frequency component of the first block image.

[0016] Furthermore, the target image feature is a local contrast of the first block image, and feature analysis is performed on the first block image to obtain the target image feature, including the steps of:

[0017] An image saliency distribution analysis is performed on the first block image to obtain a local contrast of the first block image.

[0018] Further, determining a denoising algorithm for the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm, further comprises the steps of:

[0019] calculating a standard deviation of pixel values ​​in the first block image;

[0020] Setting a first threshold, comparing the calculated standard deviation with the first threshold to determine a corresponding denoising method, wherein the first threshold is a standard deviation calculated based on pixel values ​​of the input image;

[0021] If the standard deviation is smaller than the first threshold, a traditional image denoising method or a neural network denoising method based on a small parameter amount is selected; if the standard deviation is larger than the first threshold, a neural network denoising method based on a large parameter amount is selected.

[0022] Further, determining a denoising algorithm for the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm, further comprises the steps of:

[0023] Performing frequency domain analysis on the first block image to calculate the amplitude of the frequency component;

[0024] Setting a second threshold, comparing the calculated amplitude of the frequency component with the first threshold to determine a corresponding denoising method, wherein the second threshold is an amplitude calculated based on the frequency component of the input image;

[0025] If the amplitude is less than the second threshold, a neural network denoising method based on a large parameter amount is selected; if the amplitude is greater than the first threshold, a traditional image denoising method or a neural network denoising method based on a small parameter amount is selected.

[0026] Further, determining a denoising algorithm for the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm, further comprises the steps of:

[0027] Compare the maximum pixel value and the minimum pixel value in the first block image to obtain the local contrast,

[0028] Setting a third threshold, comparing the local contrast with the third threshold to determine a corresponding denoising method, wherein the third threshold is a local contrast obtained according to a maximum pixel value and a minimum pixel value of the input image;

[0029] If the local contrast is greater than the third threshold, a denoising method based on a neural network with a small parameter amount is selected; if the local contrast is greater than the third threshold, a denoising method based on a neural network with a large parameter amount or a traditional image denoising method is selected.

[0030] Furthermore, the neural network denoising method includes a convolutional neural network, an autoencoder or a generative adversarial network.

[0031] Furthermore, the traditional image denoising method includes Gaussian filtering, Fourier transform method or discrete cosine transform method.

[0032] Furthermore, the method of performing block processing on the input image includes a frequency domain analysis method, a noise evaluation method, a target detection method or a saliency detection method.

[0033] Furthermore, merging the N second block images also includes fusion processing of image transition areas, and the fusion processing method of the image transition areas is based on bilinear interpolation between blocks or multi-band fusion based on the frequency domain.

[0034] Furthermore, the bilinear interpolation between blocks includes the steps of: performing bilinear interpolation by calculating the distance weights between the current pixel and four adjacent pixel points.

[0035] Furthermore, the multi-band fusion based on the frequency domain includes the steps of first decomposing the image into multiple frequency bands, then processing the frequency bands of two adjacent blocks respectively, performing multi-band fusion according to the distance weight, and generating the frequency bands of the transition zone.

[0036] An image denoising system comprises an image segmentation module, an image feature analysis module, an image denoising module and an image merging module, wherein the image denoising module is connected to the image feature analysis module and the image merging module respectively, and the image feature analysis module is connected to the image segmentation module.

[0037] The image block module is used to perform block processing on the input image to obtain N first block images with different sizes;

[0038] The image feature analysis module is used to perform feature analysis on the first block image to obtain target image features;

[0039] The image denoising module is used to determine a denoising algorithm for the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm; and denoise the first block image using the denoising algorithm for the target image feature to obtain a second block image;

[0040] The image merging module is used to merge the N second block images to obtain a target image.

[0041] A computer-readable storage medium stores a program, and when the program is executed, an image denoising method is implemented.

[0042] An electronic device comprises a display screen and an image processing system, wherein the display screen is used for displaying a denoised image.

[0043] Technical effect: Based on image segmentation, image feature analysis is performed according to different image noise distributions, noise levels or image significance. According to different feature analysis results, neural network denoising methods or traditional image denoising methods with different parameter amounts are selected for denoising, thereby providing better image denoising performance while reducing costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of a denoising method according to an embodiment;

[0045] Figure 2 It is a flow chart of the method based on image segmentation combined with traditional image denoising and neural network denoising;

[0046] Figure 3 It is a flow chart of denoising methods based on image segmentation using different neural networks;

[0047] Figure 4 This is a schematic diagram showing the obvious boundaries between sub-blocks after the image is divided;

[0048] Figure 5 is a schematic diagram of an image denoising system in an embodiment;

[0049] Figure 6 is the original noisy image;

[0050] Figure 7 It is the image denoised based on the traditional image denoising method;

[0051] Figure 8 The image shown is denoised based on the neural network denoising method;

[0052] Fig. 9 It is the image denoised by combining the traditional image denoising method and the neural network-based denoising method after the image is divided into blocks;

[0053] Fig.10 It is the denoised image after the image is divided into blocks and then denoised based on different neural network denoising methods. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein should be understood by people with general skills in the field to which the present invention belongs. "Including" and similar words used in this article mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0055] As an embodiment of the present invention, Figure 1 As shown in the figure, the present invention provides a method of combining the traditional denoising based on discrete cosine transform and denoising based on neural network with the idea of ​​image block. In this example, the computational load of the neural network is 60G, and the computational load of the discrete cosine transform denoising is 1.5G. It can be seen that the computational load of the discrete cosine transform is much smaller than that of the denoising based on the neural network. The number of parameters of the neural network is 500k, and the number of parameters of the discrete cosine transform is 16.

[0056] Step 100: block processing is performed on the input image to obtain N first block images with different sizes. The method for block processing on the input image is a frequency domain analysis method, a noise evaluation method, a target detection method and a saliency detection method.

[0057] Frequency domain analysis method: Use the characteristics of high-frequency and low-frequency components after frequency domain transformation (such as discrete cosine transform) to divide image blocks. For example, set a threshold according to the noise level, calculate the size of the high-frequency component after discrete cosine transform for each image block, and determine which high-frequency components are greater than the threshold and which are less than the threshold. For high-frequency components greater than the threshold, they are considered to contain important image information and should be retained or slightly denoised. For high-frequency components less than the threshold, they are considered to contain mainly noise and can be subjected to stronger denoising. This method focuses on distinguishing textures and details (high frequencies) from smooth areas (low frequencies) in the image.

[0058] Noise evaluation method: Image blocks are divided by evaluating the noise level (such as standard deviation) within the image blocks. The standard deviation is calculated for each image block. The standard deviation reflects the degree of change in the pixel value within the image block. The larger the standard deviation, the stronger the noise. For image blocks with a standard deviation greater than a preset threshold, it is considered that the noise in these blocks is large and a stronger denoising process can be performed. For image blocks with a standard deviation less than or equal to the preset threshold, it is considered that the noise in these blocks is small or no noise, and a lighter denoising process or no denoising process can be performed. This method judges the importance of a block based on the degree of fluctuation of the pixel value within the block.

[0059] Saliency detection method: evaluates the saliency of an image block by calculating the local contrast within the image block. Calculate the local contrast for each image block. The local contrast can be calculated by comparing the mean and maximum values ​​of all pixels in the window, or using other similar indicators. For image blocks with large contrast, it is believed that these blocks contain important image content and should be retained or slightly denoised. For image blocks with small contrast, it is believed that the information in these blocks may not be very important and stronger denoising can be performed. This method focuses on the visual importance and saliency of image blocks and is often used to highlight key areas in an image.

[0060] Object detection: Use neural networks or other machine learning methods to detect specific targets (such as people, cars, text, etc.) in an image. Important objects obtained using detection neural networks: In the field of computer vision, detection neural networks (such as YOLOv5, Faster R-CNN, etc.) are often used to identify and locate specific objects in an image. These networks extract features through a convolutional neural network (CNN) structure, and then use a classifier or regressor to predict the category and location of the object. When an important object is detected, the appropriate denoising method can be selected based on the characteristics of the object and the context.

[0061] For detected people, the following methods can be used to remove noise: Spatial filtering: Add some blurring effect around the face area to reduce the impact of background noise on the facial details. This can be achieved using Gaussian filtering or other types of smoothing filters. Color filtering: Depending on the skin color characteristics of the person, the intensity of certain colors can be selectively enhanced or weakened to improve the clarity of the person's outline. This can be achieved by adjusting the parameters of the hue / saturation / brightness (HSV) space. Edge detection: Use edge detection algorithms to highlight the outline of the person, thereby removing background noise.

[0062] Vehicle: For detected vehicles, the following methods can be used to denoise the image: Motion blur: Simulate the dynamic blur effect of a moving vehicle to make the vehicle outline more obvious. This can be achieved by adding a slight motion blur around the vehicle. Color fill: Fill a solid color inside the vehicle outline to highlight its shape and boundaries. This can be achieved by changing the color channel of the image. Morphological operations: Use morphological operations such as dilation and erosion to enhance the outline of the vehicle and make it clearer.

[0063] Text: For detected text, the following methods can be used for denoising: Character segmentation: First separate the text from the image, and then perform denoising on each character individually. This can be achieved by using OCR technology. Frequency domain denoising: Filter the text signal in the frequency domain to remove high-frequency noise components. This can be achieved by using techniques such as wavelet transform. Spatial domain denoising: Filter the text area in the spatial domain to remove background noise. This can be achieved by using techniques such as median filtering. This method focuses on identifying and dividing the areas in the image that contain important targets.

[0064] As a preferred implementation of this embodiment, Figure 2 As shown in the figure, for some blocks, a denoising method based on discrete cosine transform can be used, such as the 8×8 block in the figure, which has low noise, simple image content and is easy to process. This method performs discrete cosine transform on the image block, then removes high-frequency noise components according to a certain threshold, and finally reconstructs the denoised image block through inverse discrete cosine transform. The computational complexity of discrete cosine transform denoising is relatively small, and the number of parameters is also small. It is suitable for situations where the denoising effect is not particularly high or the computing resources are limited.

[0065] For other blocks, a neural network-based denoising method can be used, such as the 16×16 and 32×32 blocks in the figure, the areas with moderate and complex image noise, and the areas with general and complex image content. This method uses a trained neural network model to denoise the image blocks. Neural networks have a large number of parameters and a large amount of computation, but they usually provide better denoising results.

[0066] In this way, a trade-off between denoising effect and cost can be achieved. If a better denoising effect is expected, more blocks can use neural network-based denoising; if a lower cost is expected, more blocks can use discrete cosine transform-based denoising.

[0067] The advantages of the present invention are that a denoising method can be flexibly selected according to specific requirements and resource constraints, and at the same time, a balance between different denoising effects and costs can be achieved by adjusting the ratio of the block strategy and the denoising method.

[0068] As a preferred implementation of this embodiment, Figure 3 As shown, different blocks use different neural networks to perform image denoising. This embodiment adopts a typical convolutional neural network (CNN) workflow in the image denoising task.

[0069] The input image is divided into N image blocks, and each image block is independently fed into the neural network for processing. This parallel processing utilizes the parallel computing power of the graphics processing unit, significantly improving processing speed and efficiency.

[0070] A convolution (conv) operation is performed on each image block. The convolution layer uses a convolution kernel (or filter) to slide over the image to extract local features such as edges, textures, etc. After the convolution (conv) operation, batch normalization (BN) is applied. Batch Normalization (BN) helps stabilize the training process, improve convergence speed, and reduce overfitting.

[0071] After batch normalization, the ReLU activation function is applied. The ReLU function (Rectified Linear Unit) introduces nonlinear factors, allowing the network to learn more complex mapping relationships. The combination of convolution, batch normalization, and ReLU activation may be repeated many times to form a deep network structure. Each repetition may extract higher-level features, thereby capturing more complex image structures.

[0072] At some stage, the network may include a pooling layer (pool). Pooling operations (such as max pooling or average pooling) reduce the dimensionality of the data while retaining key information, helping to reduce computation and prevent overfitting.

[0073] At the final stage of the network, the outputs are merged together using an UpSample operation. Upsampling increases the dimensionality of the data by inserting additional pixels (e.g. using nearest neighbor upsampling, bilinear interpolation, etc.). The upsampled data is then subjected to one more convolution operation to generate the final output. This is often used to generate a denoised image or for further image processing. The final output of the network is the denoised image patches. These patches can be recombined to form the complete denoised image.

[0074] This structure shows the typical process of CNN when processing image denoising tasks, especially when performing feature extraction and image reconstruction tasks. By adjusting the structure and parameters of the network, it can adapt to different denoising needs and performance requirements.

[0075] Step 200: Perform feature analysis on the first block image to obtain target image features.

[0076] As an implementation mode of this embodiment, the target image feature is the standard deviation of the pixel values ​​in the first block image. The first block image is subjected to feature analysis to obtain the target image feature, including the steps of: performing a noise level analysis on the first block image and calculating the standard deviation of the pixel values ​​in the first block image.

[0077] The calculation formula involved in the above implementation is:

[0078] Image pixel mean

[0079] Pixel standard deviation Where ω and h are the width and height of the image, p ij is the pixel value corresponding to the ij position.

[0080] As a preferred implementation manner of this embodiment, the target image feature is the amplitude of the frequency component of the first block image in the frequency domain. The first block image is subjected to feature analysis to obtain the target image feature, including the steps of: performing noise distribution analysis on the first block image, and calculating the amplitude of the frequency component of the first block image in the frequency domain.

[0081] The calculation formula of the above implementation is:

[0082]

[0083] Where X(k) is the transformed frequency component amplitude, x(n) is the original signal, N is the signal length, and k is the frequency component index, ranging from 0 to N-1.

[0084] Discrete cosine transform is particularly important in image and video compression because it can effectively convert image data into the frequency domain and retain important visual information while removing redundant information.

[0085] As another implementation of this embodiment, the target image feature is the local contrast of the first block image, and feature analysis is performed on the first block image to obtain the target image feature, including the steps of: performing image significance distribution analysis on the first block image, comparing the maximum pixel value and the minimum pixel value in the first block image, and obtaining the local contrast.

[0086] The calculation formula of the above implementation method is:

[0087] Local contrast = maximum pixel value - minimum pixel value.

[0088] It should be noted that the use of any of the above image feature analysis methods has little effect on the result of the denoised image obtained by the final selection of the corresponding denoising method.

[0089] Step 300: determining a denoising algorithm for the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm;

[0090] As an implementation of this embodiment, a first threshold is set, and the standard deviation is compared with the first threshold to determine a corresponding denoising method, wherein the first threshold is a standard deviation calculated according to pixel values ​​of the input image;

[0091] If the standard deviation is smaller than the first threshold, a traditional image denoising method or a neural network denoising method based on a small parameter amount is selected; if the standard deviation is larger than the first threshold, a neural network denoising method based on a large parameter amount is selected.

[0092] As an implementation of this embodiment, a second threshold is set, and different values ​​of the second threshold are set according to different user requirements and noise distributions, and the amplitude is compared with the second threshold to determine a corresponding denoising method, and the second threshold is an amplitude calculated according to the frequency components of the input image;

[0093] If the amplitude is less than the second threshold, a neural network denoising method based on a large parameter amount is selected; if the amplitude is greater than the first threshold, a traditional image denoising method or a neural network denoising method based on a small parameter amount is selected.

[0094] As an implementation manner of this embodiment, a third threshold is set, and different values ​​of the third threshold are set according to user requirements and different contrasts, and the local contrast is compared with the third threshold to determine a corresponding denoising method, and the third threshold is a local contrast obtained according to the maximum pixel value and the minimum pixel value of the input image;

[0095] If the local contrast is greater than the third threshold, a denoising method based on a neural network with a small number of parameters is selected; if the local contrast is less than the third threshold, a denoising method based on a neural network with a large number of parameters or a traditional image denoising method is selected.

[0096] It should be noted that no matter which of the above methods is used to determine the corresponding denoising method, the final denoising effect has no obvious difference.

[0097] Step 400: De-noise the first block image using a denoising algorithm based on target image features to obtain a second block image.

[0098] The traditional image denoising methods include Gaussian filtering, Fourier transform or discrete cosine transform. Gaussian filtering uses Gaussian functions as weights to perform weighted averaging on images, which can effectively remove Gaussian noise. The Fourier transform method converts an image from the spatial domain to the frequency domain, removes high-frequency noise through a filter, and then performs an inverse transform to restore the image. The Fourier transform method is very effective in removing periodic noise or noise with specific frequency characteristics. The discrete cosine transform method is similar to the Fourier transform. The discrete cosine transform method converts the image to the transform domain and removes noise through threshold processing.

[0099] The neural network-based denoising method includes a convolutional neural network, an autoencoder or a generative adversarial network. A convolutional neural network (CNN) is a deep learning model that uses a convolutional layer to extract local features of an image and performs classification or regression tasks through a fully connected layer, which can capture the complex structure and contextual information of the image. An autoencoder is an unsupervised learning model consisting of an encoder and a decoder. The encoder compresses the input data into a lower-dimensional representation, while the decoder attempts to reconstruct the original data. It can be used for data dimensionality reduction, feature extraction, and unsupervised learning. A generative adversarial network is a deep learning model consisting of a generator and a discriminator. The generator attempts to generate realistic data, while the discriminator attempts to distinguish between generated data and real data, and can create data with a high sense of reality.

[0100] Step 500: merge N of the second block images to obtain a target image.

[0101] As a preferred implementation, if the denoising effects of various sub-blocks are similar, the denoised blocks are directly merged without any transition processing.

[0102] As a preferred embodiment, Figure 4As shown, if the boundaries between the sub-blocks are obvious, such as the gray area in the figure, in order to make the transition between the sub-blocks more natural, the fusion processing of the image transition area can be used. First, the positions of the important blocks of the previously divided blocks are set to 255, and the unimportant positions are set to 0 to generate a marked image. Then, the transition area is generated by the mean filtering method, so that the boundary area of ​​the block gradually transitions from one denoising effect to another denoising effect.

[0103] There are two ways to fuse the image transition area:

[0104] (1) Bilinear interpolation based on the distance between blocks: This method performs bilinear interpolation by calculating the distance weights between the current pixel and the four adjacent pixels. The formula can be expressed as:

[0105] f(i,j)=w1×p1+w2×p2+w3×p3+w4×p4

[0106] Among them, w1, w2, w3, w4 are the distance weights of four adjacent pixels, and p1, p2, p3, p4 are the pixel values ​​of these four pixels. This method can produce a smooth transition effect between blocks.

[0107] (2) Multi-band fusion based on frequency domain: This method first divides the image into multiple bands, which can be simply obtained by using a 5x5 mean filter or a low-pass filter. Then, the two bands are bilinearly interpolated based on distance in the frequency domain to achieve block fusion. This method can produce a natural transition effect between different frequency bands.

[0108] In general, merging the denoising results after image segmentation is a key step, which directly affects the final denoising effect and the naturalness of the image. Choosing a suitable merging method can make the denoised image look more natural and reduce block effects and boundary effects.

[0109] As a second embodiment of the present invention, Figure 5 As shown, an image denoising system comprises an image segmentation module 1, an image feature analysis module 2, an image denoising module 3 and an image merging module 4, wherein the image denoising module 3 is connected to the image feature analysis module 2 and the image merging module 4 respectively, and the image feature analysis module 2 is connected to the image segmentation module 1.

[0110] The image segmentation module 1 is used to segment the input image into blocks to obtain N first block images with different sizes;

[0111] The image feature analysis module 2 is used to perform feature analysis on first block images of different sizes to obtain target image features;

[0112] The image denoising module 3 is used to determine a denoising algorithm for the target image feature according to the target image feature and the first corresponding relationship, where the first corresponding relationship is the corresponding relationship between the image feature and the image denoising algorithm; and to denoise the first block image using the denoising algorithm for the target image feature to obtain a second block image;

[0113] The image merging module 4 is used to merge the N second block images to obtain a target image.

[0114] As a third embodiment of the present invention, a readable storage medium is provided, wherein a program is stored in the readable storage medium, and when the program is executed, the image denoising method described in any one of the above embodiments is implemented.

[0115] It is worth noting that if the method is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling an electronic device to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk.

[0116] As a fourth embodiment of the present invention, an electronic device includes a display screen and an image processing system, wherein the display screen is used to display a denoised image. The display screen displays: Figure 6 The original noise image is shown. For the image in the red circle area, the signal-to-noise ratio is 27.18dB. Figure 7 The image shown is the image after using traditional image denoising. For the image in the red circle area in the figure, the signal-to-noise ratio is 27.6dB; Figure 8 The image shown is denoised based on the neural network denoising method. For the image in the red circle area in the figure, the signal-to-noise ratio is 27.8dB; Fig. 9 The image shown is the image denoised after image segmentation combined with the traditional neural network denoising method and the neural network-based denoising method. For the image in the red circle area in the figure, the signal-to-noise ratio is 29.2dB; Fig.10 The image is shown as denoised based on different neural network denoising methods after image segmentation. For the image in the red circle area in the figure, the signal-to-noise ratio is 32.42dB. Therefore, it can be seen that the new denoising method that combines the advantages of traditional image denoising technology and neural network denoising or based on different neural network denoising methods provides better image denoising performance while reducing costs.

[0117] Although the embodiments of the present invention are described in detail above, it is obvious to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations are within the scope and spirit of the present invention as described in the claims. Moreover, the present invention described herein may have other embodiments and may be implemented or realized in a variety of ways.

Claims

1. An image denoising method, characterized in that: include: The input image is processed into blocks to obtain N first block images with different sizes; Performing feature analysis on the first block image to obtain target image features; Determining a denoising algorithm for the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm; De-noising the first block image using a denoising algorithm based on target image features to obtain a second block image; The N second block images are merged to obtain a target image.

2. The image denoising method according to claim 1, characterized in that: The target image feature is the standard deviation of pixel values ​​in the first block image. The first block image is subjected to feature analysis to obtain the target image feature, including the steps of: A noise level analysis is performed on the first block image to calculate a standard deviation of pixel values ​​in the first block image.

3. The image denoising method according to claim 1, characterized in that: The target image feature is the amplitude of the frequency component of the first block image in the frequency domain. The first block image is subjected to feature analysis to obtain the target image feature, including the steps of: A noise distribution analysis is performed on the first block image to calculate the amplitude of the frequency component of the first block image in the frequency domain.

4. The image denoising method according to claim 1, characterized in that: The target image feature is the local contrast of the first block image. The first block image is subjected to feature analysis to obtain the target image feature, including the steps of: An image saliency distribution analysis is performed on the first block image, and a maximum pixel value and a minimum pixel value in the first block image are compared to obtain a local contrast.

5. The image denoising method according to claim 2, characterized in that: Determining a denoising algorithm for the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm, further comprising the steps of: Setting a first threshold, comparing the standard deviation with the first threshold to determine a corresponding denoising method, wherein the first threshold is a standard deviation calculated based on pixel values ​​of the input image; If the standard deviation is smaller than the first threshold, a traditional image denoising method or a neural network denoising method based on a small parameter amount is selected; if the standard deviation is larger than the first threshold, a neural network denoising method based on a large parameter amount is selected.

6. The image denoising method according to claim 3, characterized in that: Determining a denoising algorithm for the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm, further comprising the steps of: Setting a second threshold, comparing the amplitude with the second threshold to determine a corresponding denoising method, wherein the second threshold is an amplitude calculated based on a frequency component of the input image; If the amplitude is less than a second threshold, selecting a neural network denoising method based on a large number of parameters; If the amplitude is greater than the first threshold, a traditional image denoising method or a neural network denoising method based on a small number of parameters is selected.

7. The image denoising method according to claim 4, characterized in that: Determining a denoising algorithm for the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm, further comprising the steps of: Setting a third threshold, comparing the local contrast with the third threshold to determine a corresponding denoising method, wherein the third threshold is a local contrast obtained according to a maximum pixel value and a minimum pixel value of the input image; If the local contrast is greater than the third threshold, a denoising method based on a neural network with a small number of parameters is selected; if the local contrast is less than the third threshold, a denoising method based on a neural network with a large number of parameters or a traditional image denoising method is selected.

8. The image denoising method according to any one of claims 5 to 7, characterized in that: The neural network denoising method includes a convolutional neural network, an autoencoder or a generative adversarial network.

9. The image denoising method according to any one of claims 5 to 7, characterized in that: The conventional image denoising method includes Gaussian filtering, Fourier transform method or discrete cosine transform method.

10. The image denoising method according to claim 1, characterized in that: Methods for processing the input image in blocks include frequency domain analysis, noise assessment, target detection or saliency detection.

11. The image denoising method according to claim 1, characterized in that: The N second block images are merged, and the fusion processing of the image transition area is also included. The fusion processing method of the image transition area is based on bilinear interpolation between blocks or multi-band fusion based on frequency domain.

12. The image denoising method according to claim 11, characterized in that: The bilinear interpolation between blocks includes the following steps: performing bilinear interpolation by calculating the distance weights between a current pixel and four adjacent pixels.

13. The image denoising method according to claim 12, characterized in that: The multi-band fusion based on the frequency domain includes the following steps: firstly, the image is decomposed into multiple frequency bands, then the frequency bands of two adjacent blocks are processed respectively, and the multi-band fusion is performed according to the distance weight to generate the frequency band of the transition zone.

14. An image denoising system, characterized in that: It includes an image segmentation module, an image feature analysis module, an image denoising module and an image merging module, wherein the image denoising module is connected to the image feature analysis module and the image merging module respectively, and the image feature analysis module is connected to the image segmentation module. The image block module is used to perform block processing on the input image to obtain N first block images with different sizes; The image feature analysis module is used to perform feature analysis on the first block image to obtain target image features; The image denoising module is used to determine a denoising algorithm for the target image feature according to the target image feature and a first corresponding relationship, wherein the first corresponding relationship is a corresponding relationship between the image feature and the image denoising algorithm; and denoising the first block image using a denoising algorithm based on target image features to obtain a second block image; The image merging module is used to merge the N second block images to obtain a target image.

15. A computer-readable storage medium, wherein a program is stored in the computer-readable storage medium, characterized in that: When the program is executed, the image denoising method according to any one of claims 1 to 13 is implemented.

16. An electronic device, characterized in that: It comprises a display screen and the image processing system according to claim 14, wherein the display screen is used to display the denoised image.

Citation Information

Cited By

  • Image filter optimization method and system

    CN120182101A