Image noise reduction method and device, equipment and storage medium

Abstract

The invention discloses an image noise reduction method, device and equipment and a storage medium, and is applied to the field of image processing, and the method comprises the steps: obtaining an original image, and carrying out the wavelet transformation of the original image, and obtaining a multi-scale feature map; constructing a multi-scale structure tensor based on the multi-scale feature map, and smoothing the multi-scale structure tensor to obtain a dynamic smoothing kernel; filtering the original image based on a dynamic smoothing kernel control filter to obtain a filtered image; determining a difference graph between the filtered image and the original image, and generating a structural weight graph based on the difference graph; and performing weighted fusion on the original image and the filtered image based on the structure weight map to obtain a noise reduction image. According to the method, multi-scale wavelet analysis, structure tensor feature extraction and a dynamic smoothing strategy are combined, a weight map is generated based on the structure difference between a filtered image and an original image, and the optimal balance of image detail reservation and noise suppression is achieved.

Description

Technical Field

[0001] This invention relates to the field of image processing, and particularly to an image denoising method, an image denoising apparatus, an electronic device, and a computer-readable storage medium. Background Technology

[0002] High-precision industrial inspection relies on image clarity and detail representation. However, in actual acquisition processes, industrial images are often affected by noise interference, uneven lighting, and blurred edges, which can impact subsequent recognition and analysis during information extraction and processing. Traditional image filtering methods for noise removal, such as Gaussian filtering, median filtering, and bilateral filtering, while effective to some extent, often sacrifice edge and detail, making it difficult to balance image smoothness and structural information preservation. Therefore, ensuring effective image denoising while maximizing the preservation of important structural features remains a key challenge in current industrial image processing. Summary of the Invention

[0003] The purpose of this invention is to provide an image denoising method, apparatus, device, and storage medium for use in the field of image processing. This method combines multi-scale wavelet analysis, structural tensor feature extraction, and dynamic smoothing strategies, and generates a weight map based on the structural differences between the filtered image and the original image, thereby achieving an optimal balance between image detail preservation and noise suppression.

[0004] To address the aforementioned technical problems, this invention provides an image denoising method, comprising:

[0005] Obtain the original image, and perform wavelet transform on the original image to obtain a multi-scale feature map;

[0006] A multi-scale structure tensor is constructed based on the multi-scale feature map, and a dynamic smoothing kernel is obtained by smoothing the multi-scale structure tensor.

[0007] The original image is filtered using the dynamic smoothing kernel control filter to obtain a filtered image.

[0008] Determine the difference map between the filtered image and the original image, and generate a structure weight map based on the difference map;

[0009] The original image and the filtered image are weighted and fused based on the structure weight map to obtain a denoised image.

[0010] Optionally, generating a structure weight map based on the difference map includes:

[0011] The difference map is normalized to obtain a normalized difference map;

[0012] The normalized difference map is input into the weight map function to obtain the output structure weight map.

[0013] Optionally, wavelet transform is performed on the original image to obtain a multi-scale feature map, including:

[0014] Gaussian filtering at different scales is applied to the original image to simulate the wavelet approximation coefficients of each decomposition layer during the wavelet transform process;

[0015] The wavelet approximation coefficients of each of the decomposition layers obtained from the simulation are determined as the multi-scale feature map.

[0016] Optionally, constructing a multi-scale structure tensor based on the multi-scale feature map includes:

[0017] Determine the gradient of the feature map at each scale, and determine the magnitude tensor of each feature map based on the gradient;

[0018] The amplitude tensors of each feature map are fused to obtain the multi-scale structure tensor.

[0019] Optionally, filtering the original image based on the dynamic smoothing kernel controlled filter to obtain a filtered image includes:

[0020] The filter window of the filter is constructed based on the dynamic smoothing kernel;

[0021] The original image is filtered using the filter to obtain the filtered image.

[0022] Optionally, the multi-scale structure tensor is smoothed to obtain a dynamic smoothing kernel, including:

[0023] The dynamic smoothing kernel is obtained by smoothing the multi-scale structure tensor based on the Gaussian kernel.

[0024] Optionally, the original image is a metal image, a ceramic image, or a plastic image;

[0025] It also includes: inputting the denoised image into the contour recognition model to obtain the output contour recognition result.

[0026] To solve the above-mentioned technical problems, the present invention provides an image noise reduction device, comprising:

[0027] The first module is used to acquire the original image and perform wavelet transform on the original image to obtain a multi-scale feature map;

[0028] The second module is used to construct a multi-scale structure tensor based on the multi-scale feature map, and to smooth the multi-scale structure tensor to obtain a dynamic smoothing kernel.

[0029] The third module is used to filter the original image based on the dynamic smoothing kernel control filter to obtain a filtered image;

[0030] The fourth module is used to determine the difference map between the filtered image and the original image, and generate a structure weight map based on the difference map;

[0031] The fifth module is used to perform weighted fusion of the original image and the filtered image based on the structure weight map to obtain a denoised image.

[0032] To solve the above-mentioned technical problems, the present invention provides an electronic device comprising:

[0033] Memory, used to store computer programs;

[0034] A processor is used to implement the image noise reduction method described above when executing the computer program.

[0035] To solve the above-mentioned technical problems, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the image noise reduction method described above.

[0036] As can be seen, this invention acquires the original image, performs wavelet transform on the original image to obtain a multi-scale feature map; constructs a multi-scale structure tensor based on the multi-scale feature map, and smooths the multi-scale structure tensor to obtain a dynamic smoothing kernel; uses the dynamic smoothing kernel to control the filter to filter the original image to obtain a filtered image; determines the difference map between the filtered image and the original image, and generates a structure weight map based on the difference map; and performs weighted fusion of the original image and the filtered image based on the structure weight map to obtain a denoised image. This invention combines multi-scale wavelet analysis, structure tensor feature extraction, and dynamic smoothing strategies, and generates a weight map based on the structural differences between the filtered image and the original image, achieving an optimal balance between image detail preservation and noise suppression. Attached Figure Description

[0037] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0038] Figure 1 A flowchart of an image noise reduction method provided in an embodiment of the present invention;

[0039] Figure 2 A flowchart of a MOD-STF algorithm provided in an embodiment of the present invention;

[0040] Figure 3 This is a structural block diagram of an image noise reduction device provided in an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] With the development of industrial automation and intelligent manufacturing, image processing technology has been widely applied in fields such as surface defect detection, structural recognition, and precision measurement. High-precision industrial inspection relies on the clarity and detail representation of images. However, in actual acquisition processes, industrial images are often affected by noise interference, uneven lighting, and blurred edges, which can impact subsequent recognition and analysis results during information extraction and processing. Traditional image filtering methods for noise removal, such as Gaussian filtering, median filtering, and bilateral filtering, can remove noise to some extent, but often at the cost of sacrificing edges and details, making it difficult to balance image smoothness and the preservation of structural information. Therefore, how to preserve important structural features to the greatest extent while ensuring image denoising effectiveness is a key challenge in current industrial image processing. In recent years, multi-scale analysis and structural tensors have shown strong capabilities in representing image structure, and deep learning has also provided inspiration for adaptive adjustment of image features. However, introducing these ideas into industrial filtering algorithms still suffers from weak interpretability and uncontrollable structure.

[0043] To address the aforementioned issues, this invention provides an adaptive filtering method based on multi-scale structural tensors. By combining multi-scale wavelet analysis, structural tensor feature extraction, and dynamic smoothing strategies, and introducing a bidirectional LSTM (Long Short-term Memory Networks) mechanism to simulate image structural differences, this method can achieve an optimal balance between detail preservation and noise suppression in industrial images, making it particularly suitable for edge-sensitive tasks such as defect detection.

[0044] The following combination Figure 1 , Figure 1 A flowchart of an image noise reduction method provided in an embodiment of the present invention, the method may include:

[0045] S101: Obtain the original image and perform wavelet transform on the original image to obtain a multi-scale feature map.

[0046] This embodiment first obtains the original image that needs to be denoised. This embodiment does not limit the specific format and type of the original image. Generally, it can be an image acquired in industrial applications such as surface defect detection, structure recognition and precision measurement.

[0047] This embodiment uses a multi-scale oriented denoising with structure tensor fusion (MOD-STF) adaptive filtering algorithm for industrial image detection to denoise the original image. The algorithm flow is as follows: Figure 2 As shown, it mainly includes four core steps: multi-scale image feature extraction, construction of multi-scale structural tensors, dynamic smoothing kernel generation, and weighted graph-guided structural fusion.

[0048] This embodiment first extracts features from the original image to obtain a multi-scale feature map. This embodiment does not limit the specific method of feature extraction; generally, wavelet transform can be performed on the original image to obtain the multi-scale feature map.

[0049] Specifically, in this embodiment, Gaussian filtering at different scales can be applied to the original image to simulate the wavelet approximation coefficients of each decomposition layer during the wavelet transform process; the simulated wavelet approximation coefficients of each decomposition layer are then used as multi-scale feature maps.

[0050] First, let the grayscale image of the original image be I(x,y), where (x,y) are the coordinates of the pixels in the image. In this embodiment, wavelet approximation coefficients can be simulated using Gaussian filtering at different scales:

[0051] ;

[0052] In the formula, W i (x,y) represents the feature map at the i-th scale layer, and L represents the total number of scale layers. The scale is σ i The two-dimensional Gaussian kernel function, σ i =a×i, where a is the preset smoothing scale, typically 1.2. As the number of scale layers i increases, the blur radius of the kernel gradually increases. * indicates a two-dimensional convolution operation. Multiple different σ values ​​can be set... i Values, such as [1.2, 2.4, 3.6], can yield a series of images from the "detail layer" to the "coarse layer," also known as the "Gaussian pyramid" or "Gaussian scale space."

[0053] S102: Construct a multi-scale structural tensor based on multi-scale feature maps, and perform smoothing processing on the multi-scale structural tensor to obtain a dynamic smoothing kernel.

[0054] This embodiment can construct a multi-scale structure tensor based on multi-scale feature maps. This embodiment does not limit the specific method of constructing a multi-scale structure tensor based on multi-scale feature maps. Generally, the gradient of the feature map at each scale can be determined, and the magnitude tensor of each feature map can be determined based on the gradient. The magnitude tensors of each feature map are then fused to obtain the multi-scale structure tensor.

[0055] Specifically, the gradient is calculated for the feature map at each scale:

[0056] ;

[0057] In the formula, W i (x,y) represents the gradient of the feature map at the i-th scale layer.

[0058] Furthermore, the magnitude tensor of each feature map is calculated:

[0059] ;

[0060] In the formula, T i (x,y) is the magnitude tensor of the feature map at the i-th scale layer.

[0061] Finally, the multi-scale structure tensor is obtained by fusion:

[0062] ;

[0063] In the formula, T MOST (x,y) is a multi-scale structure tensor.

[0064] MOD-STF utilizes multi-scale wavelet coefficients to construct a multi-scale oriented structure tensor (MOST), which in this embodiment is a multi-scale Gaussian smoothing map. This captures texture, orientation, and variation information at different scales and generates a dynamic smoothing kernel. This allows the filtering to adaptively enhance or suppress complex structural regions, avoiding a one-size-fits-all approach to the entire image and enhancing the algorithm's generalization ability on various complex surfaces (such as metals, plastics, and ceramics).

[0065] This embodiment can obtain a dynamic smoothing kernel (DSEC Kernel) by smoothing a multi-scale structure tensor. This embodiment does not limit the specific method of obtaining the dynamic smoothing kernel. Generally, a dynamic smoothing kernel can be obtained by smoothing the multi-scale structure tensor based on a Gaussian kernel, as shown in the following equation:

[0066] ;

[0067] In the formula, KDSEC (x,y) is the dynamic smoothing kernel. The basic smoothing scale is The Gaussian kernel, in this embodiment, can be set to 1.2, and can be adjusted based on the actual application.

[0068] S103: The original image is filtered using a dynamic smoothing kernel control filter to obtain a filtered image.

[0069] This embodiment can filter the original image based on a dynamic smoothing kernel-controlled filter to obtain a filtered image. This embodiment does not limit the specific filtering method. Generally, the filtering window of the filter can be constructed based on the dynamic smoothing kernel. The original image is then filtered based on the filter to obtain a filtered image.

[0070] Specifically, the filter window is constructed based on the dynamic smoothing kernel:

[0071] ;

[0072] In the formula, b(x,y) is a parameter used to adjust the size (or intensity) of the filter window. base α is the minimum filter strength (can be 3 or 5), α is a parameter to control the flyback degree (can be 1~2), β is to control the shape of the response curve (can be 3~10), and tanh is the activation function to make the structural response converge to outliers.

[0073] Based on the dynamic smoothing kernel size control of the filter window size (or intensity) b(x,y), perform filtering (simplified to mean or Gaussian):

[0074] ;

[0075] In the formula, I local (x,y) is the filtered image. The GaussianBlur 2D Gaussian kernel convolution operation performs a weighted average in the local region of the image.

[0076] S104: Determine the difference map between the filtered image and the original image, and generate a structure weight map based on the difference map.

[0077] This embodiment can determine the difference map between the filtered image and the original image, as shown in the following formula:

[0078] ;

[0079] In the formula, D(x,y) is the difference graph.

[0080] Furthermore, the difference map can be normalized to obtain a normalized difference map, as shown in the following formula:

[0081] ;

[0082] In the formula, D norm (x,y) is the normalized difference plot, max is the maximum value function, and z is the error-proofing parameter.

[0083] Finally, in this embodiment, the normalized difference map can be input into the weight map function to obtain the output structure weight map. The expression of the weight map function can be:

[0084] ;

[0085] In the formula, W(x,y) is the structure weight map, (x,y) are the pixel coordinates, and D... norm (x,y) is the normalized weighted graph, K DSEC (x,y) is the dynamic smoothing kernel, γ is the nonlinear enhancement exponent, typically taken as 1.2~2.0, δ is a preset constant, and + indicates that the result is positive to ensure numerical stability and prevent negative weights from being generated in abnormal difference regions. By introducing a structure-aware adjustment term 1 / (K+δ), the more obvious the structure, the smaller the penalty, i.e., the higher the weight. Furthermore, a full multiplication mechanism is introduced to strengthen the two-factor adjustment of difference + structure.

[0086] In this embodiment, when the accuracy requirement for image recognition is low in industrial applications, a simpler method can be used to obtain the structure weight map, thereby saving computational load while reducing the accuracy requirement, as shown in the following formula:

[0087] .

[0088] By constructing a difference map before and after image filtering, an "attention mechanism" type weight map is introduced for adaptive fusion processing. This simulates bidirectional LSTM extraction of local similarity, resulting in a weight map used for the fusion operation. This further enhances effective features and suppresses redundant information. Compared to the uniform smoothness of Gaussian filtering, MOD-STF's "region-selective enhancement" better meets the needs of industrial defect detection, especially suitable for scenarios with high noise and small targets.

[0089] S105: The original image and the filtered image are weighted and fused based on the structure weight map to obtain the denoised image.

[0090] This embodiment can obtain a denoised image by weighted fusion of the original image and the filtered image based on the structure weight map. This embodiment does not limit the specific method of image fusion, but it can generally be shown in the following formula:

[0091] ;

[0092] In the formula, I final(x,y) represents the final denoised image. This algorithm simulates the attention mechanism in deep learning and uses structure-driven region-weighted fusion to effectively balance denoising and structure preservation.

[0093] Based on the above embodiments, the present invention combines multi-scale wavelet analysis, structural tensor feature extraction and dynamic smoothing strategy, and generates a weight map based on the structural differences between the filtered image and the original image, thereby achieving the optimal balance between image detail preservation and noise suppression.

[0094] The following is an application example of the MOD-STF algorithm provided by this invention. In industrial high-exposure camera image fusion scenarios, MOD-STF provides a cleaner and less noisy input image. While preserving the contour information of the target captured by the industrial vision camera, which is inevitably affected by fine-grained defects such as contamination, it provides a more suitable and accurate localization method for traditional contour localization algorithms. The core principle is that after the original image is input into this filtering algorithm, an LSTM-extracted weight map is obtained, and σ is adjusted according to specific needs. i Obtain the weight map, and then use the weight map and the original image to obtain the final MOD-STF image.

[0095] In practical applications, the original image can be an image of a material with a complex surface, such as a metal image, a ceramic image, or a plastic image. This embodiment can take a metal image as an example to extract the structure weight map of the metal image; after processing by the MOD-STF algorithm, a denoised image is obtained; the anti-light noise points at the edges of the metal surface in the denoised image are filtered out, making the difference between the black blocks and the metallic color more obvious, which helps to extract different color difference blocks in specific tasks.

[0096] The denoised image is input into the contour recognition model to obtain the output contour recognition result, thereby obtaining the contour of the metal region. A comparison of the difference between the denoised image and the original image (diff algorithm) reveals that the noise on the outer contour at the bottom of the contour region is relatively scattered, resulting in lower accuracy in overall contour identification. After denoising, the edge features of the metal edges are more obvious after MOD-STF, significantly increasing the accuracy of contour extraction and leading to more precise contour recognition of the original image.

[0097] The following combination Figure 3 , Figure 3 This is a structural block diagram of an image noise reduction device provided in an embodiment of the present invention. The device may include:

[0098] The first module 100 is used to acquire the original image and perform wavelet transform on the original image to obtain a multi-scale feature map;

[0099] The second module 200 is used to construct a multi-scale structure tensor based on the multi-scale feature map and to perform smoothing processing on the multi-scale structure tensor to obtain a dynamic smoothing kernel.

[0100] The third module 300 is used to filter the original image based on the dynamic smoothing kernel control filter to obtain a filtered image;

[0101] The fourth module 400 is used to determine the difference map between the filtered image and the original image, and generate a structure weight map based on the difference map;

[0102] The fifth module 500 is used to perform weighted fusion of the original image and the filtered image based on the structure weight map to obtain a denoised image.

[0103] Based on the above embodiments, the present invention combines multi-scale wavelet analysis, structural tensor feature extraction and dynamic smoothing strategy, and generates a weight map based on the structural differences between the filtered image and the original image, thereby achieving the optimal balance between image detail preservation and noise suppression.

[0104] Based on the above embodiments, the fourth module 400 may include:

[0105] The first unit is used to normalize the difference map to obtain a normalized difference map.

[0106] The second unit is used to input the normalized difference map into the weight map function to obtain the output structure weight map.

[0107] Based on the above embodiments, the first module 100 may include:

[0108] The third unit is used to perform Gaussian filtering on the original image at different scales to simulate the wavelet approximation coefficients of each decomposition layer in the wavelet transform process.

[0109] The fourth unit is used to determine the wavelet approximation coefficients of each of the decomposition layers obtained from the simulation as the multi-scale feature map.

[0110] Based on the above embodiments, the second module 200 may include:

[0111] The fifth unit is used to determine the gradient of the feature map at each scale, and to determine the magnitude tensor of each feature map based on the gradient;

[0112] The sixth unit is used to fuse the magnitude tensors of each feature map to obtain the multi-scale structure tensor.

[0113] Based on the above embodiments, the third module 300 may include:

[0114] The seventh unit is used to construct the filter window of the filter based on the dynamic smoothing kernel;

[0115] The eighth unit is used to filter the original image based on the filter to obtain the filtered image.

[0116] Based on the above embodiments, the second module 200 may include:

[0117] The ninth unit is used to smooth the multi-scale structure tensor based on a Gaussian kernel to obtain the dynamic smoothing kernel.

[0118] Based on the above embodiments, the original image is a metal image, a ceramic image, or a plastic image;

[0119] It also includes a sixth module, used to input the denoised image into the contour recognition model to obtain the output contour recognition result.

[0120] Based on the above embodiments, the present invention also provides an electronic device, which may include a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it can implement the steps provided in the above embodiments. Of course, the device may also include various necessary network interfaces, a power supply, and other components.

[0121] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by an execution terminal or processor, can implement the method provided in the embodiments of the present invention; the storage medium may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0122] In this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, without necessarily requiring or implying any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. An image denoising method, characterized in that, include: Obtain the original image, and perform wavelet transform on the original image to obtain a multi-scale feature map; A multi-scale structure tensor is constructed based on the multi-scale feature map, and a dynamic smoothing kernel is obtained by smoothing the multi-scale structure tensor. The original image is filtered using the dynamic smoothing kernel control filter to obtain a filtered image. Determine the difference map between the filtered image and the original image, and generate a structure weight map based on the difference map; The original image and the filtered image are weighted and fused based on the structure weight map to obtain a denoised image.

2. The image denoising method according to claim 1, characterized in that, Generate a structure weight map based on the difference map, including: The difference map is normalized to obtain a normalized difference map; The normalized difference map is input into the weight map function to obtain the output structure weight map.

3. The image denoising method according to claim 1, characterized in that, The original image is subjected to wavelet transform to obtain a multi-scale feature map, including: The original image is subjected to Gaussian filtering at different scales to simulate the wavelet approximation coefficients of each decomposition layer in the wavelet transform process; The wavelet approximation coefficients of each of the decomposition layers obtained from the simulation are determined as the multi-scale feature map.

4. The image denoising method according to claim 1, characterized in that, Constructing a multi-scale structure tensor based on the multi-scale feature map includes: Determine the gradient of the feature map at each scale, and determine the magnitude tensor of each feature map based on the gradient; The amplitude tensors of each feature map are fused to obtain the multi-scale structure tensor.

5. The image denoising method according to claim 1, characterized in that, The original image is filtered using the dynamic smoothing kernel control filter to obtain a filtered image, including: The filter window of the filter is constructed based on the dynamic smoothing kernel; The original image is filtered using the filter to obtain the filtered image.

6. The image denoising method according to claim 1, characterized in that, The dynamic smoothing kernel is obtained by smoothing the multi-scale structure tensor, including: The dynamic smoothing kernel is obtained by smoothing the multi-scale structure tensor based on the Gaussian kernel.

7. The image denoising method according to claim 1, characterized in that, The original image is a metal image, a ceramic image, or a plastic image; It also includes: inputting the denoised image into the contour recognition model to obtain the output contour recognition result.

8. An image noise reduction device, characterized in that, include: The first module is used to acquire the original image and perform wavelet transform on the original image to obtain a multi-scale feature map; The second module is used to construct a multi-scale structure tensor based on the multi-scale feature map, and to smooth the multi-scale structure tensor to obtain a dynamic smoothing kernel. The third module is used to filter the original image based on the dynamic smoothing kernel control filter to obtain a filtered image; The fourth module is used to determine the difference map between the filtered image and the original image, and generate a structure weight map based on the difference map; The fifth module is used to perform weighted fusion of the original image and the filtered image based on the structure weight map to obtain a denoised image.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the image denoising method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the image denoising method as described in any one of claims 1 to 7.

Images