Fruit detection image denoising method based on image signal processing
By constructing a multi-layer non-Euclidean graph structure model and a compactly supported bioorthogonal graph wavelet filter bank, the problem of insufficient noise suppression in fruit detection is solved, achieving efficient image denoising and accurate foreign object detection.
Patent Information
- Application Number
- CN202511226211.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-29
- Publication Date
- 2026-01-09
AI Technical Summary
In the detection of subtropical fruits, existing technologies based on regular grid modeling methods are insufficient to effectively describe the irregular surface texture of fruits and sensor networks, resulting in inadequate noise suppression, degraded image quality, and reduced accuracy in foreign object detection.
A graph-based signal processing approach is adopted to construct a multi-layer non-Euclidean graph structure model. Multi-scale decomposition and filtering are performed using a compactly supported bioorthogonal graph wavelet filter bank. Noise is removed by combining the hard thresholding method, and useful signals are retained.
It effectively removes noise from fruit images, improves image quality, ensures the accuracy of foreign object detection and the stability of the production line, and avoids misjudgment and misoperation.
Smart Images
Figure CN121305099A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image processing technology, and specifically relates to a method for denoising fruit detection images based on graph signal processing. Background Technology
[0002] In the industrial production line of subtropical fruits (such as passion fruit and mango), foreign object detection is often carried out using images captured by high-speed industrial cameras. Then, image recognition algorithms are used to identify foreign objects (such as mold spots and insect holes) on the surface of the fruit, providing a basis for screening out unqualified fruits in subsequent quality inspection stages.
[0003] Currently, high-speed industrial cameras are easily affected by environmental noise interference, such as uneven lighting, when capturing images of subtropical fruits or processed fruit products. Furthermore, the transmission channel characteristics are susceptible to sudden instability from external and internal factors. In sudden situations, the industrial camera acquisition system on the production line can be affected by noise, causing image data distortion and a sharp drop in the Peak Signal-to-Noise Ratio (PSNR), leading to a decrease in the accuracy of foreign object detection. If the inspection line cannot detect this sudden drop in PSNR in the acquired images in time, it can cause the vision system to misjudge foreign objects or even malfunction, thus exposing the inspection line to significant risks. When the image PSNR drops to 20dB-30dB, the image quality is insufficient for effective judgment by the vision system, requiring the timely introduction of image preprocessing algorithms to ensure the vision system can correctly identify foreign objects in fruits and jams. When the image PSNR drops below 20dB, a production line fault warning should be issued immediately, and the invalid inspection process on the inspection line should be stopped.
[0004] Traditional preprocessing methods for images of fruits or fruit pulp products captured by industrial cameras often employ regular grid (Euclidean space) modeling. Image data exists in a regular, uniform Euclidean space, where each pixel has fixed, well-defined neighbors, resulting in structural limitations. This approach cannot effectively describe complex data with irregular structures (where each node may have a different number of neighbors, and the "distance" between neighbors does not follow regular grid spacing), such as data on fruit surface textures or sensor networks. Mapping irregular data onto a regular grid introduces errors or leads to the loss of topological information. Secondly, it lacks sufficient ability to suppress noise such as speckle and multiplicative noise caused by sensor gain, easily resulting in the loss of detail. Summary of the Invention
[0005] The purpose of this invention is to provide a fruit detection image denoising method based on graph signal processing, in order to solve the technical problems of existing regular grid (Euclidean space) modeling methods, which are prone to introducing errors or losing topological information when describing irregular structures such as fruit surface texture and sensor networks, as well as insufficient ability to suppress multiplicative noise.
[0006] To achieve the above objectives, the present invention provides a method for denoising fruit detection images based on graph signal processing, comprising:
[0007] Obtain the original fruit image and construct a multi-layer non-Euclidean graph structure model of the original fruit image;
[0008] Construct a compactly supported bioorthogonal graph wavelet filter bank, including an analysis filter bank and a synthesis filter bank;
[0009] The multi-layer non-Euclidean graph structure model is decomposed into multiple scales using the aforementioned analysis filter bank to obtain the high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure.
[0010] The high-frequency coefficients are filtered using a hard thresholding method to obtain the denoised high-frequency coefficients.
[0011] The low-frequency coefficients and the denoised high-frequency coefficients are input into the integrated filter bank for inverse graph wavelet transform to obtain the denoised fruit image.
[0012] Furthermore, in the above technical solution, the expression for the analysis filter bank is:
[0013]
[0014] In the formula, h k These are the weighting coefficients. Let Λ represent the asymmetric normalized Laplace matrix raised to the power of k. k Let U be a diagonal matrix raised to the power of k, and U be the eigenvector matrix. T h is the rank matrix of the eigenvector matrix. i (λ) represents the analysis filter bank H i The spectral kernel function, where λ represents the eigenvalue and K is the Chebyshev order;
[0015] The expression for the synthesized filter bank is:
[0016]
[0017] In the formula, g k g is the weighting coefficient. i (λ) represents the synthesis filter bank G. i The spectral kernel function.
[0018] Furthermore, in the above technical solution, the spectral kernel function of the analysis filter bank and the spectral kernel function of the synthesis filter bank satisfy the following biorthogonality:
[0019] h1(λ)=g0(2-λ),g1(λ)=h1(2-λ) (3)
[0020] The spectral kernel functions of the analysis filter bank and the synthesis filter bank satisfy the following tightly supported structure:
[0021] h0(λ)g0(λ)+h0(2-λ)g0(2-λ)=2 (4)
[0022] Furthermore, in the above technical solution, the process of constructing a compactly supported bioorthogonal graph wavelet filter bank includes:
[0023] The coefficients of the spectral kernel function h0(λ) of the analysis filter and the spectral kernel function g0(λ) of the synthesis filter are calculated using the maximum flatness design method. The formulas are as follows:
[0024]
[0025] p0(λ)=h0(λ)g0(λ) (6)
[0026] In the formula, l is the normalized frequency, J is the root of the half-band filter, and r m Let J0 be the order of the half-band filter, J0 be the number of predefined half-band filters, and p0(λ) be the half-band spectral kernel.
[0027] Substituting h0(λ) and g0(λ) into Equation 3, we obtain the coefficients of the spectral kernel function h1(λ) in the analysis filter bank and the spectral kernel function g1(λ) in the synthesis filter bank.
[0028] The frequency responses of the analytical filter bank's spectral kernel function {h0(λ), h1(λ)} and the synthetic filter bank's spectral kernel function {g0(λ), g1(λ)} are approximated using k-th order Chebyshev polynomials, respectively, to obtain the analytical and synthetic filter banks. Their calculation formulas are as follows:
[0029]
[0030] In the formula, T k (λ) is a Chebyshev polynomial, c k is the coefficient, and K is the Chebyshev order.
[0031] Furthermore, in the above technical solution, the tightly supported biorthogonal graph wavelet filter bank includes an upsampling operator and a downsampling operator. The downsampling operator is used to downsample the high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure to obtain the high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure data after dimensionality reduction. Then, the upsampling operator is used to upsample the high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure data after dimensionality reduction to obtain the high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure data after dimensionality upsampling and restoration.
[0032] Furthermore, in the above technical solution, during the process of filtering the high-frequency coefficients using the hard threshold method to obtain the denoised high-frequency coefficients, the high-frequency coefficients include the high-frequency coefficients after dimensionality reduction of each layer of non-Euclidean graph structure data and the high-frequency coefficients after dimensionality restoration of each layer of non-Euclidean graph structure data.
[0033] Furthermore, in the above technical solution, the calculation formula for filtering the high-frequency coefficients using the hard threshold method is as follows:
[0034] T=α·σ (9)
[0035] In the formula, T is the threshold, α is the empirical coefficient, and σ is the noise standard deviation;
[0036] The hard thresholding method mainly uses the following rules to determine the high-frequency coefficients:
[0037]
[0038] The method uses a hard thresholding approach to suppress low-amplitude noise figures while preserving significant signal components.
[0039] Furthermore, in the above technical solution, the process of constructing a multi-layer non-Euclidean graph structure model of the original fruit image includes:
[0040] The original fruit image is converted into a standardized grayscale image;
[0041] The standardized grayscale image is decomposed into a hierarchical structure to obtain a multi-layer non-Euclidean graph structure model;
[0042] Based on the pixel spatial location and gray-level similarity of graph nodes in a multi-layer non-Euclidean graph structure model, an adjacency matrix is constructed. The weight function of the adjacency matrix is:
[0043]
[0044] In the formula, loc i Let loc be the coordinates of pixel i in the image. j Let I be the coordinates of pixel j in the image. i Let I be the grayscale value of pixel i. j Let σ represent the grayscale value of pixel j, σ represent the distance between different pixel nodes, and γ represent the grayscale difference tolerance parameter.
[0045] Based on the adjacency matrix, calculate the asymmetric normalized Laplacian matrix L of each level of the graph structure in the multi-layer non-Euclidean graph structure model. n =ID -1 A, where I is the identity matrix, D -1 Let A be the inverse of the degree matrix of the graph structure, and let A be the adjacency matrix of the graph structure.
[0046] Furthermore, the above technical solution also includes: evaluating the denoising effect of the tightly supported bioorthogonal graph wavelet filter bank on the original fruit image using the peak signal-to-noise ratio (PSNR) index. The PNR characterization formula is as follows:
[0047]
[0048] Where MSE represents mean squared error and Lev is the maximum pixel value.
[0049] Compared with existing technologies, the present invention has the following advantages:
[0050] This invention constructs a multi-layer non-Euclidean graph structure model from the original fruit image, then inputs this model into a compactly supported bioorthogonal graph wavelet filter bank for multi-scale decomposition, filtering, and reconstruction to obtain a denoised fruit image. Utilizing graph signal processing techniques, it maps the complex, irregular data in the fruit image to a graph structure, achieving non-Euclidean signal analysis. The filter bank decomposes the graph signal into coefficients of different frequency bands. Since noise in the fruit image is concentrated in the high-frequency region, a hard thresholding method is used to filter the noise in the high-frequency region, while retaining the useful low-frequency coefficients and the denoised high-frequency signal. The input to the integrated filter bank for wavelet inverse reconstruction and restoration can effectively remove noise from the original fruit image. At the same time, the analysis filter bank and the integrated filter bank have a tightly supported structure and are biorthogonal. The tightly supported structure of the analysis filter bank has the characteristic of acting only on the "local small region where the noise is located" and will not spread to the surrounding effective pixels. It can effectively remove multiplicative noise concentrated in the texture-dense or high-brightness areas of the fruit. The biorthogonality allows the analysis filter bank and the integrated filter bank to be designed independently, so that the integrated filter bank can be designed to be high-fidelity when reconstructing and restoring image data, avoiding data loss of image data. Attached Figure Description
[0051] Figure 1 This is a schematic diagram illustrating the implementation process of the fruit detection image denoising method based on graph signal processing provided by the present invention.
[0052] Figure 2 shows images of noisy passion fruit pulp and noisy mango with PSNR ≥ 30dB, captured using an industrial camera.
[0053] Figure 3 The image shows a quality comparison of the fruit detection image denoising method based on graph signal processing provided in this invention when the PSNR of the noisy passion fruit pulp image is in the range of [20dB, 30dB].
[0054] Figure 4This is a comparison image showing the quality of fruit detection image denoising after preprocessing by the graph signal processing-based method provided in this invention when the PSNR of the noisy mango image is in the range of [20dB, 30dB]. Detailed Implementation
[0055] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the invention is not limited to the specific embodiments.
[0056] refer to Figure 1 A method for denoising fruit detection images based on graph signal processing includes the following steps:
[0057] Step S1: Obtain the original fruit image and construct a multi-layer non-Euclidean graph structure model of the original fruit image;
[0058] In this embodiment, the original fruit image is obtained by an industrial camera. Since the surface texture of the fruit and the data of the sensor network in the camera system are irregular structures (each node may have a different number of neighbors, and the "distance" between neighbors does not follow the regular grid spacing), the non-Euclidean graph structure model uses graph wavelet transform to map these irregular structure data in the fruit image to a graph structure (graph signal). This can effectively preserve the topological information of the irregular structure data in the fruit image, avoiding the problem of errors or loss of topological information that are introduced when mapping irregular data to a regular grid based on the traditional modeling method based on regular grid (Euclidean space).
[0059] Step S2: Construct a compactly supported bioorthogonal graph wavelet filter bank, including an analysis filter bank and a synthesis filter bank;
[0060] Step S3: Use the analysis filter bank to perform multi-scale decomposition on the multi-layer non-Euclidean graph structure model to obtain the high-frequency coefficients and low-frequency coefficients of each layer of the non-Euclidean graph structure model.
[0061] Step S4: Filter the high-frequency coefficients using the hard thresholding method to obtain the denoised high-frequency coefficients;
[0062] Step S5: Input the low-frequency coefficients and the denoised high-frequency coefficients into the integrated filter bank to perform graph wavelet inverse transform to obtain the denoised fruit image.
[0063] In step S2, the expression for the analysis filter bank is:
[0064]
[0065] In the formula, h k These are the weighting coefficients. Let Λ represent the asymmetric normalized Laplace matrix raised to the power of k. kLet U be a diagonal matrix raised to the power of k, and U be the eigenvector matrix. T h is the rank matrix of the eigenvector matrix. i (λ) represents the analysis filter bank H i The spectral kernel function, where λ is the eigenvalue and K is the Chebyshev order;
[0066] Specifically, h i The formula for calculating (λ) is:
[0067]
[0068] In the formula, λ l These are the eigenvalues.
[0069] The expression for the synthesized filter bank is:
[0070]
[0071] In the formula, g i (λ) represents the synthesis filter bank G. i Spectral kernel function;
[0072] Specifically, g i The formula for calculating (λ) is:
[0073]
[0074] in, g k λ is the weighting coefficient. l For eigenvalues;
[0075] The spectral kernel functions of the analysis filter bank and the spectral kernel function of the synthesis filter bank satisfy the following bioorthogonality property:
[0076] h1(λ)=g0(2-λ),g1(λ)=h1(2-λ) (3)
[0077] It should be noted that traditional orthogonal wavelets, in order to achieve orthogonality, must sacrifice either the "support domain size" or "smoothness." This can result in either an excessively large support domain leading to blurred details, or the generation of a "ringing effect" during reconstruction (such as artifacts appearing at the edges of fruit). The "signal dependence" of multiplicative noise exacerbates this distortion (after removing noise from high-brightness areas, brightness discontinuities are easily observed during reconstruction). In contrast, the analysis filter bank and synthesis filter bank, which satisfy the biorthogonality property, do not require "strict orthogonality" and can be optimized independently. In practical implementation, the analysis filter bank can be designed with "high noise separation capability," accurately distinguishing between "wavelet coefficients corresponding to multiplicative noise" (high frequency, low correlation) and "useful signal coefficients" (low frequency, high correlation) in the image domain. The synthesis filter bank can be designed with "high reconstruction fidelity," ensuring that even after threshold suppression of noise coefficients (the core denoising operation), the original brightness distribution and texture details of the fruit image are still restored after reconstruction, without ringing effects.
[0078] The analysis filter bank and the synthesis filter bank satisfy the following tight-support structure:
[0079] h0(λ)g0(λ)+h0(2-λ)g0(2-λ)=2 (4)
[0080] It's important to note that multiplicative noise often concentrates in areas of dense texture on fruit (such as fruit wrinkles) or in high-brightness areas (such as reflective spots on fruit). Traditional filters with excessively large support domains (such as traditional large-kernel Gaussian filters) can cause the noise to smooth along with adjacent effective textures / edges, resulting in blurred textures (such as the disappearance of fruit wrinkles) and blurred edges (such as the fruit stem edge becoming blurred). In contrast, the tightly supported analytical and synthetic filter banks only act on the localized small area where the noise is located and do not spread to surrounding effective pixels. For example, when processing multiplicative noise in the striped area of a fruit, the tightly supported analytical and synthetic filter banks only act on 3-5 pixels near the stripes (the support domain size can be designed), which can suppress noise while fully preserving the alternation of light and dark in the stripes, avoiding "over-smoothing".
[0081] Furthermore, based on the above technical solution, the construction process of the closely supported biorthogonal graph wavelet filter bank includes the following steps:
[0082] Step S2.1: Calculate the coefficients of the spectral kernel function h0(λ) of the analysis filter and the spectral kernel function g0(λ) of the synthesis filter using the maximum flatness design method. The calculation formula is as follows:
[0083]
[0084] p0(λ)=h0(λ)g0(λ) (6)
[0085] In the formula, l is the normalized frequency, J is the root of the half-band filter, and rm Let J0 be the order of the half-band filter, J0 be the number of predefined half-band filters, and p0(λ) be the half-band spectral kernel.
[0086] Step S2.2: Substitute h0(λ) and g0(λ) into Formula 3 to calculate the coefficients of the spectral kernel function h1(λ) in the analysis filter bank and the spectral kernel function g1(λ) in the synthesis filter bank.
[0087] Step S2.3: Using k-th order Chebyshev polynomials, approximate the frequency responses of the spectral kernel functions {h0(λ), h1(λ)} of the analysis filter bank and the spectral kernel functions {g0(λ), g1(λ)} of the synthesis filter bank, respectively, to obtain the analysis filter bank and the synthesis filter bank. Their calculation formulas are as follows:
[0088]
[0089]
[0090] In the formula, T k (λ) is a Chebyshev polynomial, c k is the coefficient, and K is the Chebyshev order.
[0091] It should be noted that, in this embodiment, a k-th order Chebyshev polynomial is introduced to approximate the frequency response of {h0(λ), h1(λ)} and {g0(λ), g1(λ)}, thereby reducing the h-value of the analysis filter bank. i (λ) and g of the integrated filter bank i This reduces the computational complexity of (λ), thereby improving the efficiency of solving analytical and synthetic filter banks.
[0092] Furthermore, based on the above technical solution, the support for the bioorthogonal graph wavelet filter bank includes an upsampling operator and a downsampling operator. The downsampling operator is used to downsample the high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure to obtain the high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure data after dimensionality reduction. Then, the upsampling operator is used to upsample the high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure data after dimensionality reduction to obtain the high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure data after dimensionality upsampling. In this embodiment, the upsampling operator can be Gaussian weighted interpolation, and the downsampling operator can be generated by an average pooling filter. Through downsampling and upsampling operations, high-frequency and low-frequency coefficients of non-Euclidean graph structures at different scales are provided.
[0093] Furthermore, based on the above technical solution, the high-frequency coefficients are filtered using the hard threshold method to obtain the denoised high-frequency coefficients. The high-frequency coefficients include the high-frequency coefficients after dimensionality reduction of each layer of non-Euclidean graph structure data and the high-frequency coefficients after dimensionality restoration of each layer of non-Euclidean graph structure data. This enables signal processing on graph structures at different scales (resolutions), balancing the filtering effect (such as denoising and feature extraction) with computational efficiency, and realizing multi-scale analysis of high-frequency and low-frequency coefficients of multi-layer non-Euclidean graph structures.
[0094] Furthermore, based on the above technical solution, the calculation formula for filtering high-frequency coefficients using the hard threshold method is as follows:
[0095] T=α·σ (9)
[0096] In the formula, T is the threshold, α is the empirical coefficient, and σ is the noise standard deviation;
[0097] The hard thresholding method mainly uses the following rules to determine the high-frequency coefficients:
[0098]
[0099] Low-amplitude noise figures are suppressed by using a hard thresholding method, while significant signal components are preserved.
[0100] Furthermore, based on the above technical solution, the process of constructing a multi-layer non-Euclidean graph structure of the original fruit image in step S1 includes:
[0101] Step S1.1: Convert the original fruit image into a standardized grayscale image;
[0102] Step S1.2: Perform hierarchical decomposition on the standardized grayscale image to obtain a multi-layer non-Euclidean graph structure model;
[0103] Step S1.3: Based on the pixel spatial location and gray-level similarity of graph nodes in the multi-layer non-Euclidean graph structure model, construct an adjacency matrix. The weight function of the adjacency matrix is:
[0104]
[0105] In the formula, loc i Let loc be the coordinates of pixel i in the image. j Let I be the coordinates of pixel j in the image. i Let I be the grayscale value of pixel i. j Let σ represent the grayscale value of pixel j, σ represent the distance between different pixel nodes, and γ represent the grayscale difference tolerance parameter.
[0106] Step S1.4: Based on the adjacency matrix, calculate the asymmetric normalized Laplacian matrix L of each level of the graph structure in the multi-level non-Euclidean graph structure model.n =ID -1 A, where I is the identity matrix, D -1 Let A be the inverse of the degree matrix of the graph structure, and let A be the adjacency matrix of the graph structure.
[0107] It should be noted that, in this embodiment, the degree matrix D and adjacency matrix A of the graph structure are used for image frequency domain analysis, design of tightly supported bioorthogonal graph wavelet filter banks, and graph wavelet stable reconstruction.
[0108] Furthermore, based on the above technical solution, it also includes:
[0109] The peak signal-to-noise ratio (PSNR) is used to evaluate the denoising effect of the tightly supported bioorthogonal graph wavelet filter bank on the original fruit image. The formula for PSNR is as follows:
[0110]
[0111] Where MSE is the mean squared error, Lev is the maximum pixel value, and the higher the PSNR, the better the image quality.
[0112] In practical application, the images of passion fruit, mango, and other fruits captured by industrial cameras on the fruit inspection production line using this invention are processed in two ways, each corresponding to different noise pollution conditions.
[0113] (1) When there is noisy image, as shown in Figure 2(a) and (b), the passion fruit pulp image with PSNR=31.86dB and the mango image with PSNR=30.4dB are respectively. Practice and theory show that such images are less affected by noise and have higher image clarity. After the PSNR is judged to meet the standard by the preprocessing method of the present invention, it can be sent to the visual inspection system for jam foreign object detection.
[0114] (2) When the PSNR of noisy images is between 20dB and 30dB, these images already contain a significant amount of speckle noise and require image preprocessing to improve image quality before being sent to the visual inspection system for foreign object detection. Such images include... Figure 3 and Figure 4 As shown, the images represent before-and-after quality comparisons of noisy passion fruit and mango images after processing using the preprocessing method of this invention. Figure 3 and Figure 4 It can be seen that the PSNR of the processed image can be improved by 4-5dB, reaching more than 30dB, and can be sent to the visual inspection system for inspection, which helps to improve accuracy and efficiency.
[0115] The above experiments verified that the introduced image denoising method for fruit detection based on graph signal processing can effectively handle noise in fruit images, demonstrating its image denoising effect in subtropical fruit detection. This helps ensure the normal judgment of the subsequent visual inspection system, and can determine the PSNR of noisy images, promptly issue production line fault warnings, and stop invalid inspections from continuing on the inspection line, thus meeting the real-time requirements of industrial production lines.
[0116] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.
Claims
1. A method for denoising fruit detection images based on graph signal processing, characterized in that, include: Obtain the original fruit image and construct a multi-layer non-Euclidean graph structure model of the original fruit image; Construct a compactly supported bioorthogonal graph wavelet filter bank, including an analysis filter bank and a synthesis filter bank; The multi-layer non-Euclidean graph structure model is decomposed into multiple scales using the aforementioned analysis filter bank to obtain the high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure. The high-frequency coefficients are filtered using a hard thresholding method to obtain the denoised high-frequency coefficients. The low-frequency coefficients and the denoised high-frequency coefficients are input into the integrated filter bank for inverse graph wavelet transform to obtain the denoised fruit image.
2. The fruit detection image denoising method based on graph signal processing according to claim 1, characterized in that, The expression for the analysis filter bank is: In the formula, h k These are the weighting coefficients. Let Λ represent the asymmetric normalized Laplace matrix raised to the power of k. k Let U be a diagonal matrix raised to the power of k, and U be the eigenvector matrix. T h is the rank matrix of the eigenvector matrix. i (λ) represents the analysis filter bank H i The spectral kernel function, where λ is the eigenvalue and K is the Chebyshev order; The expression for the synthesized filter bank is: In the formula, g k g is the weighting coefficient. i (λ) represents the synthesis filter bank G. i The spectral kernel function.
3. The fruit detection image denoising method based on graph signal processing according to claim 2, wherein the spectral kernel function of the analysis filter bank and the spectral kernel function of the synthesis filter bank satisfy the following biorthogonal property: h1(λ)=g0(2-λ),g1(λ)=h1(2-λ) (3) The spectral kernel functions of the analysis filter bank and the synthesis filter bank satisfy the following tightly supported structure: h0(λ)g0(λ)+h0(2-λ)g0(2-λ)=2 (4).
4. The fruit detection image denoising method based on graph signal processing according to claim 3, characterized in that, The process of constructing the compactly supported bioorthogonal graph wavelet filter bank includes: The coefficients of the spectral kernel function h0(λ) of the analysis filter and the spectral kernel function g0(λ) of the synthesis filter are calculated using the maximum flatness design method. The formulas are as follows: p0(λ)=h0(λ)g0(λ) (6) In the formula, l is the normalized frequency, J is the root of the half-band filter, and r m Let J0 be the order of the half-band filter, J0 be the number of predefined half-band filters, and p0(λ) be the half-band spectral kernel. Substituting h0(λ) and g0(λ) into Equation 3, we obtain the coefficients of the spectral kernel function h1(λ) in the analysis filter bank and the spectral kernel function g1(λ) in the synthesis filter bank. The frequency responses of the analytical filter bank's spectral kernel function {h0(λ), h1(λ)} and the synthetic filter bank's spectral kernel function {g0(λ), g1(λ)} are approximated using k-th order Chebyshev polynomials, respectively, to obtain the analytical and synthetic filter banks. Their calculation formulas are as follows: In the formula, T k (λ) is a Chebyshev polynomial, c k is the coefficient, and K is the Chebyshev order.
5. The fruit detection image denoising method based on graph signal processing according to claim 1, characterized in that, The tightly supported bioorthogonal graph wavelet filter bank includes an upsampling operator and a downsampling operator. The downsampling operator is used to downsample the high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure to obtain the dimensionality-reduced high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure data. Then, the upsampling operator is used to upsample the dimensionality-reduced high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure data to obtain the dimensionality-upgraded high-frequency and low-frequency coefficients of each layer of the non-Euclidean graph structure data.
6. The fruit detection image denoising method based on graph signal processing according to claim 5, characterized in that, In the process of filtering the high-frequency coefficients using the hard threshold method to obtain the denoised high-frequency coefficients, the high-frequency coefficients include the high-frequency coefficients after dimensionality reduction of each layer of non-Euclidean graph structure data and the high-frequency coefficients after dimensionality restoration of each layer of non-Euclidean graph structure data.
7. The fruit detection image denoising method based on graph signal processing according to claim 6, characterized in that, The formula for filtering the high-frequency coefficients using the hard threshold method is as follows: T=α·σ (9) In the formula, T is the threshold, α is the empirical coefficient, and σ is the noise standard deviation; The hard thresholding method mainly uses the following rules to determine the high-frequency coefficients: The method uses a hard thresholding approach to suppress low-amplitude noise figures while preserving significant signal components.
8. The fruit detection image denoising method based on graph signal processing according to claim 1, characterized in that, The process of constructing a multi-layer non-Euclidean graph structure model of the original fruit image includes: The original fruit image is converted into a standardized grayscale image; The standardized grayscale image is decomposed into a hierarchical structure to obtain a multi-layer non-Euclidean graph structure model; Based on the pixel spatial location and gray-level similarity of graph nodes in a multi-layer non-Euclidean graph structure model, an adjacency matrix is constructed. The weight function of the adjacency matrix is: In the formula, loc i Let loc be the coordinates of pixel i in the image. j Let I be the coordinates of pixel j in the image. i Let I be the grayscale value of pixel i. j Let σ represent the grayscale value of pixel j, σ represent the distance between different pixel nodes, and γ represent the grayscale difference tolerance parameter. Based on the adjacency matrix, calculate the asymmetric normalized Laplacian matrix L of each level of the graph structure in the multi-layer non-Euclidean graph structure model. n =ID -1 A, where I is the identity matrix, D -1 Let A be the inverse of the degree matrix of the graph structure, and let A be the adjacency matrix of the graph structure.
9. The fruit detection image denoising method based on graph signal processing according to claim 1, characterized in that, Also includes: The peak signal-to-noise ratio (PSNR) is used to evaluate the denoising effect of the tightly supported bioorthogonal graph wavelet filter bank on the original fruit image. The PNR formula is as follows: Where MSE represents mean squared error and Lev is the maximum pixel value.