Noise suppression optimization method for high-sensitivity camera module

Abstract

The invention discloses a noise suppression optimization method for a high-sensitivity camera module. The method comprises the following steps: S1, acquiring an original image and preprocessing the original image; s2, executing discrete Fourier transform, and converting the image into a frequency domain image; s3, inputting the frequency domain image into a Fourier neural operator model, and extracting frequency domain features; s4, inverse Fourier transform is executed, and a preliminary noise reduction image is generated to serve as a reference image; s5, adopting an improved jellyfish group optimization algorithm to perform global search and optimization on model structure parameters and hyper-parameters; s6, updating the Fourier neural operator model according to an optimization result, and performing closed-loop iterative optimization; and S7, performing image quality evaluation on the noise-reduced image and the reference image until performance indexes are met. According to the method, the Fourier neural operator and the improved jellyfish group optimization algorithm are fused, so that efficient noise suppression and image quality improvement of the high-sensitivity camera module in a low-light environment are realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular, to a method for optimizing noise suppression of a high-sensitivity camera module. Background Art

[0002] With the rapid development of image processing, optical imaging, and intelligent terminal technologies, the application of high-sensitivity camera modules in low-light environments has become increasingly widespread, such as in the fields of security monitoring, night driving, drone reconnaissance, and smartphone night photography. The requirements for image quality are also increasing day by day. However, traditional image denoising methods often struggle to balance detail preservation and noise suppression in low-light environments, resulting in obvious noise interference in the image, reducing the overall clarity and realism of the image. Currently, the existing technologies mainly include hardware filtering, circuit noise suppression, digital signal processing, and image denoising methods based on convolutional neural networks (CNNs). Although hardware filtering and circuit noise suppression can reduce noise to a certain extent, due to the limitations of the performance and design parameters of physical devices, their noise reduction effects often fail to meet the high-quality requirements under variable lighting conditions. Digital signal processing methods, although having a high processing speed, in a complex noise environment, especially in the high-sensitivity mode, their filtering algorithms are prone to problems such as image detail loss and edge blurring. In recent years, deep learning methods have made some progress in the field of image denoising. However, denoising models based on convolutional neural networks usually rely on a large amount of training data and empirical parameter tuning, and are easily affected by fixed model structures, local optima, and insufficient real-time performance in practical applications.

[0003] In the process of applying Fourier transform technology to image denoising, traditional discrete Fourier transform methods can convert image information in the spatial domain into frequency domain information, thereby separating noise and signals. However, due to the lack of means for adaptively processing frequency domain data, it is often difficult to make full use of spectral information to improve the denoising effect. At the same time, Fourier transform methods are prone to be affected by noise accumulation during the inverse transform process, resulting in unsatisfactory image restoration effects. To solve the above problems, in recent years, Fourier Neural Operator (FNO) technology has emerged. By performing deep feature mapping and data fusion in the frequency domain, it can more efficiently capture global image information and noise characteristics, thereby achieving the purpose of image denoising. However, the performance of the FNO model largely depends on the setting of its structural parameters and training hyperparameters. Traditional methods often use fixed or manually adjusted parameter combinations, which are difficult to adapt to the dynamic changes of complex noise distributions in low-light environments, are prone to falling into local optimal solutions, and there is still much room for improvement in the robustness and real-time performance of the model.

[0004] On the other hand, in recent years, meta-heuristic optimization algorithms have been widely used in the parameter tuning of deep learning models. Among them, the jellyfish swarm optimization algorithm, as a new global optimization method, has attracted attention due to its good balance ability between global exploration and local exploitation. The traditional jellyfish swarm optimization algorithm simulates the behavior of jellyfish randomly drifting and actively searching in the ocean, and finds the optimal solution through a dynamic motion strategy. However, in dealing with high-dimensional parameter spaces and complex noise environments, its performance is still limited by insufficient utilization of local information and a single motion strategy, making it difficult to achieve fine control of model parameters.

[0005] Therefore, how to provide an optimization method for noise suppression of high-sensitivity camera modules is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] An object of the present invention is to propose an optimization method for noise suppression of high-sensitivity camera modules. The present invention makes full use of the ability of the Fourier neural operator to perform feature mapping, data fusion, and adaptive enhancement on images in the frequency domain. At the same time, an improved jellyfish swarm optimization algorithm is introduced to globally search and intelligently optimize the structural parameters and training hyperparameters of the Fourier neural operator model, and a closed-loop iterative optimization process for image noise suppression and detail restoration in low-light environments is described in detail. This method adopts a multi-scale residual spectrum attention mechanism and an adaptive second-order statistical feature fusion strategy to effectively extract and strengthen useful frequency-domain features in the image; at the same time, through the improved jellyfish swarm optimization algorithm, on the basis of the original global exploration and local exploitation, local information fusion and rotational motion patterns are further introduced to dynamically balance the search process, realizing fine adjustment between the global and local optima of candidate solutions, thereby improving the efficiency and robustness of parameter optimization. Through this solution, various noises generated by high-sensitivity cameras in low-light environments can be significantly reduced, the signal-to-noise ratio and structural similarity of images can be improved, the integrity of image details can be guaranteed, and real-time processing can be achieved, with strong adaptability, high model robustness, significant noise reduction effect, and fast operation speed.

[0007] A method for optimizing noise suppression of a high-sensitivity camera module according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect an original image through a high-sensitivity camera module and preprocess the original image;

[0009] S2. Perform a discrete Fourier transform on the preprocessed original image to convert it into a frequency-domain image;

[0010] S3. Input the frequency-domain image into the Fourier neural operator model, perform multi-level feature mapping and data fusion processing on the frequency-domain image, and form an intermediate spectrum representation for noise reduction processing;

[0011] S4. Perform an inverse Fourier transform on the intermediate spectral representation, convert it into a spatial domain image, and obtain the preliminarily denoised image as the reference image;

[0012] S5. Use the jellyfish swarm optimization algorithm to globally search for and optimize the key structural parameters and training hyperparameters in the Fourier neural operator model;

[0013] S6. Update the Fourier neural operator model according to the parameter combination optimized by the jellyfish swarm optimization algorithm, and apply the updated Fourier neural operator model to the closed-loop iterative optimization process;

[0014] S7. During the closed-loop iterative optimization process, perform quality assessment on the denoised image output by the updated Fourier neural operator model and the reference image until the preset denoising performance index is reached, and output the image after deep fusion denoising processing.

[0015] Optionally, the S2 specifically includes:

[0016] S21. Perform size detection on the preprocessed original image to obtain the number of rows, columns, and channels of the image;

[0017] S22. Convert the pixel matrix of the image into a data format suitable for Fourier transform operations and format the image data;

[0018] S23. Initialize the calculation region of the two-dimensional discrete Fourier transform, set the range of frequency distribution in the frequency domain coordinate system according to the image size, and determine the appropriate sampling interval and frequency resolution;

[0019] S24. Call the Fourier transform function in the image processing module to perform the conversion operation from the spatial domain to the frequency domain on the formatted image data:

[0020]

[0021] where f(x, y) represents the pixel value of the image at the coordinate (x, y), M and N respectively represent the width and height of the image, u and v are the frequency domain coordinates, j is the imaginary unit, exp is the exponential function, and F(u, v) represents the frequency domain image obtained through the discrete Fourier transform;

[0022] S25. Perform separation processing on the obtained frequency domain image F(u, v) for the amplitude spectrum and the phase spectrum, and store the energy information and phase information in the frequency domain image data in independent matrices respectively;

[0023] S26. Perform energy normalization processing on the separated frequency domain image data, adjust the energy distribution of the spectral representation through standardization operations, and finally output the standardized frequency domain image as the input of the Fourier neural operator model.

[0024] Optionally, S3 specifically includes:

[0025] S31. Receive the standardized frequency-domain image data F(u, v) output;

[0026] S32. Perform a preliminary linear mapping on the frequency-domain image data F(u, v) to form a preliminary feature representation T(u, v);

[0027] S33. Introduce a multi-scale residual spectral attention mechanism to apply a dynamic frequency-domain adaptive transformation to the preliminary feature representation T(u, v) to obtain an enhanced frequency-domain feature expression:

[0028]

[0029] where represents the Fourier transform operation, Θ(T(u, v)) is the frequency-domain adaptive weight function, A is the learnable weight matrix, C is the learnable bias, λ o is the residual modulation coefficient, P(u, v) represents the enhanced frequency-domain feature expression at the frequency coordinates (u, v), ⊙ represents the element-wise multiplication operation, and μ represents the local mean;

[0030] S34. Apply an adaptive second-order statistical feature fusion operation to the enhanced frequency-domain feature expression P(u, v) to form an intermediate feature representation Q(u, v):

[0031] Q(u, v) = W · {P(u, v) + η o · [P(u, v) - Avg(P(u, v))] 2} + b;

[0032] where W represents the weight matrix, b represents the bias term, η o is the learnable second-order modulation coefficient, and Avg(P(u, v)) represents the operation of calculating the mean of the enhanced frequency-domain feature expression within a local region;

[0033] S35. Perform a multi-level feature fusion process on the intermediate feature representation Q(u, v), and form a denoised intermediate spectral representation H(u, v) by layer-by-layer stacking and fusion;

[0034] S36. Output the intermediate spectral representation H(u, v) as the final output of the Fourier neural operator model.

[0035] Optionally, S4 specifically includes:

[0036] S41. Receive the output intermediate spectral representation H(u, v);

[0037] S42. Perform an inverse discrete Fourier transform on the intermediate frequency spectrum representation H(u, v) to convert it into a spatial domain image g(x, y):

[0038]

[0039] where g(x, y) represents the pixel value of the spatial domain image at the coordinate (x, y), M and N represent the width and height of the image respectively, u and v represent the horizontal and vertical coordinates in the frequency domain respectively, j is the imaginary unit, and exp is the exponential function;

[0040] S43. Perform amplitude correction processing on the obtained spatial domain image g(x, y) to adjust the amplitude distribution of each pixel value in the spatial domain image;

[0041] S44. Perform brightness and contrast normalization operations on the spatial domain image g(x, y) so that the overall brightness and contrast of the spatial domain image meet the preset standards;

[0042] S45. Apply edge detail enhancement processing to the normalized spatial domain image g(x, y) to strengthen the edge and texture information in the spatial domain image;

[0043] S46. Determine the spatially domain image g(x, y) after amplitude correction, brightness and contrast normalization, and edge detail enhancement processing as the preliminarily denoised image and output it as a reference image.

[0044] Optionally, the S5 specifically includes:

[0045] S51. Determine the search space of the key structural parameters and training hyperparameters in the Fourier neural operator model and construct a candidate parameter vector:

[0046] X = [α, β, γ, δ, η, B, λ, d, ξ];

[0047] where α represents the number of frequency domain dimension truncations, β represents the width of the projection layer, γ represents the number of layers, δ represents the skip connection design parameter, η represents the learning rate, B represents the batch size, λ represents the regularization parameter, d represents the Dropout ratio, and ξ represents the rotational motion adjustment coefficient;

[0048] S52. Randomly initialize the candidate solution set where N represents the number of candidate solutions, represents the parameter combination of the i-th candidate solution at the initial time t = 0;

[0049] S53. Define a fitness function f(X) to evaluate the quality of candidate solutions:

[0050] f(X) = ω1·PSNR(X) + ω2·SSIM(X);

[0051] Among them, PSNR(X) represents the peak signal-to-noise ratio, SSIM(X) represents the structural similarity index, and ω1 and ω2 are set weight coefficients;

[0052] S54. For each candidate solution where t represents the number of iterations, and a local neighborhood set is defined as all candidate solutions whose Euclidean distance is less than the preset threshold δ th and calculate the local information fusion vector:

[0053]

[0054] Among them, represents the Euclidean distance between candidate solutions and , σ ′ is the local diffusion adjustment parameter, and exp is the exponential function;

[0055] S55. For each candidate solution calculate the rotational motion vector The rotational motion vector is constructed based on the angular difference between the candidate solution and the local neighborhood center C i :

[0056]

[0057] Among them, C i is the local neighborhood center, θ i represents the phase angle difference between candidate solution and the local neighborhood center C i , represents performing a rotation operation on vector with θ i as the rotation angle, and ξ is the rotational motion adjustment coefficient;

[0058] S56. Update the position of each candidate solution using an improved hybrid motion strategy:

[0059]

[0060] Among them, represents the candidate solution with the highest fitness in the current iteration, is a random vector within the interval [-1, 1], ρ is the global exploration coefficient, σ is the random perturbation coefficient, τ is the local information fusion coefficient, ψ is the rotational motion adjustment coefficient, represents the updated position of the i-th candidate solution at the (t + 1)-th iteration;

[0061] S57. For each updated candidate solution Calculate fitness

[0062] S58. Determine whether the preset termination condition is satisfied, i.e., the maximum number of iterations T max or the fitness convergence threshold is reached. If not, let t = t + 1 and return to S54. If so, proceed to the next step;

[0063] S59. Output the candidate solution X for which the fitness function f(X) achieves the optimal value in the last iteration opt as the optimized parameter combination of the Fourier neural operator model.

[0064] Optionally, the specific steps of S6 are as follows:

[0065] S61. Receive the output optimal parameter combination, where the optimal parameter combination includes the structural parameters and training hyperparameters of the Fourier neural operator model;

[0066] S62. Apply the optimal parameter combination to initialize the Fourier neural operator model to complete the reconstruction of the Fourier neural operator model structure and the update of the training strategy;

[0067] S63. Based on the updated Fourier neural operator model structure, perform feature mapping and noise reduction processing on the frequency-domain image again to output a new intermediate frequency spectrum representation;

[0068] S64. Perform an inverse Fourier transform on the new intermediate frequency spectrum representation to generate the noise-reduced image output for the current iteration round;

[0069] S65. Evaluate the image quality of the current noise-reduced image and the obtained reference image, and obtain the performance results in terms of peak signal-to-noise ratio and structural similarity index;

[0070] S66. Determine whether the output of the current Fourier neural operator model meets the set noise reduction performance standard according to the evaluation results. If not, use the current model parameters as the input for the next round of closed-loop iteration, and continue to execute the parameter optimization and Fourier neural operator model update process. If so, complete the optimization of the Fourier neural operator model and perform the output.

[0071] The beneficial effects of the present invention are as follows:

[0072] The present invention realizes the efficient noise reduction and detail restoration of images of high-sensitivity camera modules in low-light environments by deeply integrating the Fourier neural operator with an improved jellyfish swarm optimization algorithm. By using the Fourier neural operator to perform multi-level feature mapping and data fusion in the frequency domain, the present invention can fully extract the global spectral features in the image, and adopts a multi-scale residual spectral attention mechanism and an adaptive second-order statistical feature fusion strategy, significantly improving the discrimination ability between image noise and useful signals, and effectively retaining image details.

[0073] At the same time, the improved jellyfish swarm optimization algorithm introduces local information fusion and rotational motion patterns, which not only realizes the global search and intelligent optimization of the key structural parameters and training hyperparameters of the Fourier neural operator model, but also makes the parameter optimization process more efficient and robust by dynamically balancing global exploration and local exploitation. After the closed-loop iterative optimization process, the model can automatically adjust to the optimal state, thereby outputting high-quality denoised images, significantly improving the peak signal-to-noise ratio and structural similarity of the images.

[0074] Generally speaking, while solving the defects such as serious noise interference, obvious detail loss, and parameter tuning relying on manual experience in traditional low-light noise reduction, the present invention realizes a significant improvement in real-time performance and adaptability, providing a high-precision, robust and efficient noise suppression optimization scheme for high-sensitivity camera modules in various complex low-light environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0076] Figure 1 is a flowchart of an optimization method for noise suppression of a high-sensitivity camera module proposed by the present invention;

[0077] Figure 2 is a schematic structural diagram of a Fourier neural operator model of an optimization method for noise suppression of a high-sensitivity camera module proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0078] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0079] Refer to Figure 1 and Figure 2 , an optimization method for noise suppression of a high-sensitivity camera module, includes the following steps:

[0080] S1. Collect the original image through a high-sensitivity camera module and preprocess the original image;

[0081] S2. Perform a discrete Fourier transform on the preprocessed original image to convert it into a frequency-domain image;

[0082] S3. Input the frequency-domain image into the Fourier neural operator model, perform multi-level feature mapping and data fusion processing on the frequency-domain image to form an intermediate spectrum representation after noise reduction processing;

[0083] S4. Perform an inverse Fourier transform on the intermediate spectrum representation to convert it into a spatial-domain image, and obtain the preliminarily noise-reduced image as a reference image;

[0084] S5. Use the jellyfish swarm optimization algorithm to globally search for and optimize the key structural parameters and training hyperparameters in the Fourier neural operator model;

[0085] S6. Update the Fourier neural operator model according to the parameter combination optimized by the jellyfish swarm optimization algorithm, and apply the updated Fourier neural operator model to the closed-loop iterative optimization process;

[0086] S7. During the closed-loop iterative optimization process, perform quality evaluation on the noise-reduced image output by the updated Fourier neural operator model and the reference image until the preset noise reduction performance index is reached, and output the image after deep fusion noise reduction processing.

[0087] In the present invention, by performing feature mapping and deep fusion on the image in the frequency domain, the separability between noise and the image signal is effectively improved, so that the effective structures and details in the image are retained. The jellyfish swarm optimization algorithm is introduced to globally search for the structural parameters and training hyperparameters of the Fourier neural operator model, which not only improves the convergence speed and generalization ability of the model, but also avoids the problems of low efficiency and unstable performance caused by manual parameter tuning. In addition, the present invention constructs a closed-loop iterative optimization process, and dynamically optimizes the noise reduction process through continuous quality evaluation with the reference image, so that the output image achieves the optimal effect in terms of peak signal-to-noise ratio, structural similarity, and detail fidelity. The overall method has the characteristics of good noise reduction effect, clear image details, high optimization efficiency, and strong adaptability, and is particularly suitable for the image enhancement requirements of high-sensitivity camera modules in complex low-light scenarios, improving the imaging quality and the intelligent perception ability of the system.

[0088] In this embodiment, the S2 specifically includes:

[0089] S21. Detect the size of the preprocessed original image to obtain the number of rows, columns, and channels of the image;

[0090] S22. Convert the pixel matrix of the image into a data format suitable for Fourier transform operations, and format the image data;

[0091] S23. Initialize the calculation region of the two-dimensional discrete Fourier transform, set the range of frequency distribution in the frequency domain coordinate system according to the image size, and determine the appropriate sampling interval and frequency resolution;

[0092] S24. Call the Fourier transform function in the image processing module to perform the conversion operation from the spatial domain to the frequency domain on the formatted image data:

[0093]

[0094] where f(x, y) represents the pixel value of the image at the coordinate (x, y), M and N respectively represent the width and height of the image, u and v are the frequency domain coordinates, j is the imaginary unit, exp is the exponential function, and F(u, v) represents the frequency domain image obtained through the discrete Fourier transform;

[0095] S25. Perform the separation process of the amplitude spectrum and the phase spectrum on the obtained frequency domain image F(u, v), and store the energy information and the phase information in the frequency domain image data in independent matrices respectively;

[0096] S26. Perform energy normalization processing on the separated frequency domain image data, adjust the energy distribution of the spectrum representation through standardization operations, and finally output the standardized frequency domain image as the input of the Fourier neural operator model.

[0097] Through the refined processing of the discrete Fourier transform process, the present invention significantly improves the expression ability of the frequency domain image and the subsequent modeling quality, and has many beneficial effects. First, by introducing the size detection and formatting conversion steps after image preprocessing, it ensures that the input data structure of the Fourier transform is complete and the numerical types are consistent, laying a high-precision input foundation for frequency domain modeling. Second, by initializing the frequency domain coordinate system and sampling parameters, a frequency distribution framework suitable for the current image size is adaptively constructed, effectively improving the resolution and accuracy of spectrum calculation. After the frequency domain image is transformed by the Fourier transform, it is further separated into two parts: the amplitude spectrum and the phase spectrum, and stored separately, which is beneficial for the subsequent model to independently extract and jointly model the energy features and structural information. Especially in the energy normalization processing link, by standardizing and adjusting the energy distribution of different frequency components, local high-frequency interference is suppressed, and the model's perception ability of key frequency components is enhanced. The finally output standardized frequency domain image not only enhances the stability and expressiveness of the features, but also significantly improves the input consistency and training efficiency of the Fourier neural operator model, providing a solid data foundation for high-quality noise reduction processing.

[0098] In this embodiment, the specific content of S3 includes:

[0099] S31. Receive the standardized frequency-domain image data F(u, v) output;

[0100] S32. Perform a preliminary linear mapping on the frequency-domain image data F(u, v) to form a preliminary feature representation T(u, v);

[0101] S33. Introduce a multi-scale residual spectral attention mechanism to apply a dynamic frequency-domain adaptive transformation to the preliminary feature representation T(u, v) to obtain an enhanced frequency-domain feature expression:

[0102]

[0103] where represents the Fourier transform operation, Θ(T(u, v)) is the frequency-domain adaptive weight function, A is the learnable weight matrix, C is the learnable bias, λ o is the residual modulation coefficient, P(u, v) represents the enhanced frequency-domain feature expression at the frequency coordinates (u, v), ⊙ represents the element-wise multiplication operation, and μ represents the local mean;

[0104] S34. Apply an adaptive second-order statistical feature fusion operation to the enhanced frequency-domain feature expression P(u, v) to form an intermediate feature representation Q(u, v):

[0105] Q(u, v) = W · {P(u, v) + η o · [P(u, v) - Avg(P(u, v))] 2} + b;

[0106] where W represents the weight matrix, b represents the bias term, η o is the learnable second-order modulation coefficient, and Avg(P(u, v)) represents the operation of calculating the mean of the enhanced frequency-domain feature expression within a local region;

[0107] S35. Perform a multi-level feature fusion process on the intermediate feature representation Q(u, v), and form a denoised intermediate spectral representation H(u, v) by layer-by-layer stacking and fusion;

[0108] S36. Output the intermediate spectral representation H(u, v) as the final output of the Fourier neural operator model.

[0109] Through the improved design of the structure of the Fourier neural operator model, the present invention has achieved significant performance improvement in frequency-domain feature modeling, with clear technical advantages and application values. First, for the standardized frequency-domain image data, the present invention constructs a feature basis through preliminary linear mapping, effectively reducing the complexity of the original spectrum distribution and providing a structured input for subsequent modeling. By introducing a multi-scale residual spectrum attention mechanism, the model can adaptively adjust the feature weights according to the importance of different frequency components, enhancing the model's ability to identify noise and detail regions in the image, thereby improving the suppression effect on structural noise in the image. In particular, this mechanism combines residual information with a local mean adjustment term, strengthening the model's collaborative modeling ability for local and global frequency patterns. Further, in the intermediate feature processing stage, the present invention innovatively introduces an adaptive second-order statistical feature fusion operation. By introducing local mean and variance offsets, it realizes the refined enhancement and robustness improvement of frequency-domain features, effectively alleviating the distortion problem of image details due to frequency filtering. Finally, through a multi-level feature fusion structure, it realizes the unified modeling of deep semantic information and shallow edge textures, and outputs a more discriminative intermediate spectrum representation. This design significantly improves the model's ability to understand images under complex noise backgrounds, laying a solid foundation for subsequent image reconstruction and noise reduction effect optimization.

[0110] In this embodiment, step S4 specifically includes:

[0111] S41. Receive the output intermediate spectrum representation H(u, v);

[0112] S42. Perform an inverse discrete Fourier transform on the intermediate spectrum representation H(u, v) to convert it into a spatial-domain image g(x, y):

[0113]

[0114] where g(x, y) represents the pixel value of the spatial-domain image at coordinates (x, y), M and N respectively represent the width and height of the image, u and v respectively represent the horizontal and vertical coordinates in the frequency domain, j is the imaginary unit, and exp is the exponential function;

[0115] S43. Perform amplitude correction processing on the obtained spatial-domain image g(x, y) to adjust the amplitude distribution of each pixel value in the spatial-domain image;

[0116] S44. Perform brightness and contrast normalization operations on the spatial-domain image g(x, y) so that the overall brightness and contrast of the spatial-domain image meet the preset standards;

[0117] S45. Apply edge detail enhancement processing to the normalized spatial-domain image g(x, y) to strengthen the edge and texture information in the spatial-domain image;

[0118] S46. Determine the spatial domain image g(x,y) that has undergone amplitude correction, brightness and contrast normalization, and edge detail enhancement as the pre-denoised image and output it as the reference image.

[0119] Through performing the inverse Fourier transform on the intermediate frequency spectrum representation and cooperating with a series of spatial domain image enhancement processes, the present invention constructs a high-quality image reconstruction path from the frequency domain to the spatial domain, having significant technical advantages and application effects. First, during the inverse Fourier transform process, the frequency domain features are accurately restored to the spatial domain image, realizing the effective mapping of the frequency domain noise reduction information and ensuring the structural integrity of the image during the conversion process. Subsequently, the amplitude correction step proposed by the present invention can automatically adjust the dynamic range of the image pixels, solve the problem of local gray scale offset caused by frequency domain reconstruction, and improve the uniformity of the image brightness distribution. The brightness and contrast normalization operation further enhances the visual consistency and viewing comfort of the overall picture, enabling the image to have good visual effects in different scenarios. To prevent the problems of image detail blurring and edge weakening in traditional noise reduction methods, the present invention introduces edge detail enhancement processing after image reconstruction, effectively strengthening the texture, contour, and structural information in the image, so that the image still has high clarity and recognition while maintaining low noise. The finally output reference image serves as an important basis for subsequent closed-loop iterative optimization, providing a highly reliable reference sample for image quality evaluation and model update. In summary, the present invention integrates multiple spatial domain enhancement strategies based on frequency domain reconstruction, not only significantly improving the image restoration quality, but also enhancing the adaptability of the model to low-light and high-noise conditions.

[0120] In this embodiment, the specific content of S5 includes:

[0121] S51. Determine the search space of the key structural parameters and training hyperparameters in the Fourier neural operator model, and construct a candidate parameter vector:

[0122] X = [α, β, γ, δ, η, B, λ, d, ξ];

[0123] where α represents the number of frequency domain dimension truncations, β represents the width of the projection layer, γ represents the number of layers, δ represents the skip connection design parameter, η represents the learning rate, B represents the batch size, λ represents the regularization parameter, d represents the Dropout ratio, and ξ represents the rotational motion adjustment coefficient;

[0124] S52. Randomly initialize the candidate solution set where N represents the number of candidate solutions, represents the parameter combination of the i-th candidate solution at the initial time t = 0;

[0125] S53. Define a fitness function f(X) to evaluate the quality of the candidate solutions:

[0126] f(X) = ω1·PSNR(X) + ω2·SSIM(X);

[0127] Wherein, PSNR(X) represents the peak signal-to-noise ratio, SSIM(X) represents the structural similarity index, and ω1 and ω2 are set weight coefficients;

[0128] S54. For each candidate solution where t represents the number of iterations, define the local neighborhood set as all candidate solutions with Euclidean distance less than the preset threshold δ th and calculate the local information fusion vector:

[0129]

[0130] Wherein, represents the Euclidean distance between candidate solutions and , σ ′ is the local diffusion adjustment parameter, and exp is the exponential function;

[0131] S55. For each candidate solution calculate the rotational motion vector The rotational motion vector is constructed based on the angular difference between the candidate solution and the local neighborhood center C i :

[0132]

[0133] Wherein, C i is the local neighborhood center, θ i represents the phase angle difference between candidate solution and the local neighborhood center C i , represents the rotation operation of vector with θ i as the rotation angle, and ξ is the rotational motion adjustment coefficient;

[0134] S56. For each candidate solution perform position update, adopting an improved hybrid motion strategy:

[0135]

[0136] Wherein, represents the candidate solution with the highest fitness in the current iteration, is a random vector within the interval [-1, 1], ρ is the global exploration coefficient, σ is the random perturbation coefficient, τ is the local information fusion coefficient, ψ is the rotational motion adjustment coefficient, represents the updated position of the i-th candidate solution at the (t + 1)-th iteration;

[0137] S57. For each updated candidate solution Calculate the fitness

[0138] S58. Determine whether the preset termination condition is satisfied, that is, reaching the maximum number of iterations T max or the fitness convergence threshold. If not satisfied, let t = t + 1 and return to S54. If satisfied, proceed to the next step;

[0139] S59. Output the candidate solution X for which the fitness function f(X) achieves the optimal value in the last iteration opt as the optimized parameter combination of the Fourier neural operator model.

[0140] In view of the problems of slow convergence, easy to fall into local optimum, and insufficient information utilization of the traditional jellyfish swarm optimization algorithm in the high-dimensional complex parameter space, the present invention proposes an improved jellyfish swarm optimization algorithm that integrates local information perception and rotational motion mechanism, and applies it to the global search and optimization of the key structural parameters and training hyperparameters of the Fourier neural operator model, with significant technical progress and practical effects. The first core of the improvement lies in the introduction of the local information fusion mechanism: by constructing a local neighborhood set of candidate solutions and calculating the weighted fusion vector, each individual not only considers the global optimal guidance during update, but also fuses the distribution characteristics of excellent solutions in the neighborhood, improving the distribution diversity and stability of the solutions, and significantly enhancing the fine-grained exploration ability of the algorithm for the local space. The second innovation is the proposed rotational motion strategy as an important part of the hybrid motion mechanism: simulating the phase rotation of candidate solutions around their local centers, enhancing the diversity of the search path, avoiding the premature trap caused by linear convergence, and effectively expanding the search space. The finally formed hybrid motion update strategy integrates four update modes: global guidance, local perturbation, information fusion, and rotational adjustment, achieving a dynamic balance between exploration and exploitation at different stages. Through the above improvements, the algorithm shows stronger global search ability, faster convergence speed, and better optimal solution quality in high-dimensional non-linear optimization scenarios, providing a stable, efficient, and automated parameter optimization scheme for the Fourier neural operator model, and significantly improving the noise reduction performance and robustness of the model.

[0141] In this embodiment, the S6 specifically includes:

[0142] S61. Receive the output optimal parameter combination, where the optimal parameter combination includes the structural parameters and training hyperparameters of the Fourier neural operator model;

[0143] S62. Apply the optimal parameter combination to initialize the Fourier neural operator model, completing the reconstruction of the Fourier neural operator model structure and the update of the training strategy;

[0144] S63. Based on the updated Fourier neural operator model structure, re - perform feature mapping and noise reduction processing on the frequency - domain image, and output a new intermediate spectrum representation;

[0145] S64. Perform inverse Fourier transform on the new intermediate spectrum representation to generate the denoised image output for the current iteration round;

[0146] S65. Perform image quality assessment on the current denoised image and the obtained reference image, and obtain the performance results in terms of peak signal - to - noise ratio and structural similarity index;

[0147] S66. Judge whether the output of the current Fourier neural operator model reaches the set denoising performance standard according to the evaluation results. If not, use the current model parameters as the input for the next round of closed - loop iteration, and continue to execute the parameter optimization and Fourier neural operator model update process. If so, complete the optimization of the Fourier neural operator model and perform the output.

[0148] The present invention realizes the dynamic tuning and performance self - adaptation of the Fourier neural operator model in the image denoising task by constructing a closed - loop iterative optimization mechanism driven by optimal parameters, and has the beneficial effects of high intelligence, refinement, and stabilization. First, by receiving the optimal parameter combination output by the jellyfish swarm optimization algorithm, the model can be globally updated at both the structural parameter and training strategy levels, avoiding the problem of insufficient adaptability of the fixed model structure in different image scenarios. Second, the updated Fourier neural operator model has stronger expressiveness in frequency - domain feature extraction and denoising expression, can quickly output a new intermediate spectrum representation according to the optimization goal, and obtain the denoised image for the current iteration round through inverse Fourier transform. By introducing a reference image comparison mechanism and quality evaluation indexes (PSNR and SSIM), the performance feedback and quantitative evaluation of each round of output images are realized, effectively supporting the self - supervised optimization process of the system. This closed - loop iterative mechanism enables the model to continuously receive performance feedback, adjust parameters, and update again, effectively avoiding over - fitting or under - fitting problems, and ensuring that the final output image reaches the optimal level in terms of detail retention, noise suppression, and structural similarity. This method not only improves the adaptive ability and optimization accuracy of the model, but also enhances its stability and robustness in various complex image scenarios.

[0149] Example 1:

[0150] To verify the feasibility of the present invention in implementation, the present invention is applied to the night security monitoring system of a large city. Due to insufficient light and environmental complexity, there are often serious noise interferences in the images collected by the camera module in the high-sensitivity mode, resulting in the loss of details in the monitoring images and blurred pictures, thus affecting the timely detection and response of the monitoring system to abnormal behaviors. The traditional hardware filtering and traditional noise reduction algorithms have poor effects in this scenario. They can neither achieve the ideal noise reduction effect in a short time nor will they lose some key image details, bringing potential hazards to practical applications.

[0151] To solve this problem, the present invention adopts a noise suppression optimization method for high-sensitivity camera modules based on the deep fusion of Fourier neural operators and an improved jellyfish swarm optimization algorithm. Specifically, in an experiment in the monitoring center of a certain city's public security bureau, an improved high-sensitivity camera module was used to collect images of a key area at night. The test site is located in a bustling area in the city center, and the test time is from 22:00 at night on November 15, 2023 to 02:00 in the early morning of the next day. The system first preprocesses the collected original images, converts the images to the frequency domain through discrete Fourier transform, and then uses Fourier neural operators to perform multi-level feature mapping and adaptive data fusion in the frequency domain to generate an intermediate frequency spectrum representation. Subsequently, this intermediate frequency spectrum representation is restored to a spatial domain image through inverse Fourier transform to obtain a preliminary noise-reduced image, which serves as a reference image for the subsequent closed-loop iterative optimization process.

[0152] On this basis, the system uses the improved jellyfish swarm optimization algorithm to globally search and optimize the key structural parameters and training hyperparameters in the Fourier neural operator model. This improved algorithm introduces local information fusion and rotational motion strategies into the global search and local exploitation mechanisms of the traditional jellyfish swarm optimization algorithm, dynamically balancing the global optimal and local optimal information, thereby making the parameter tuning more refined. During the closed-loop iterative process, the system evaluates the image quality of the noise-reduced image output in each round and the preliminary noise-reduced image, and conducts performance comparison based on the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) indicators, so as to continuously optimize the model parameters until the preset noise reduction performance indicators are achieved. After repeated iterative optimization, the finally generated noise-reduced image shows obvious noise reduction and detail enhancement effects, and both the image clarity and contrast have been significantly improved.

[0153] In this embodiment, the comparative experiment shows that under the same acquisition conditions, the average PSNR of the image without being processed by the method of the present invention is only 22.5 dB, and the SSIM is 0.80. After being processed by the method of the present invention, the average PSNR of the image is increased to 26.8 dB, and the SSIM is increased to 0.87. In addition, the processing delay of the system is controlled within 50 milliseconds, meeting the requirements of real-time monitoring. After testing, the edge details in the denoised image are clearer, the texture information is better retained, the noise interference is greatly reduced, and the visual effect of the image is significantly better than that of the traditional denoising method. At the same time, the method of the present invention shows good robustness and adaptability in different low-light scenarios, and can stably output high-quality images in complex environments such as rainy days and foggy days, effectively improving the judgment accuracy of the monitoring system for abnormal situations.

[0154] Table 1 Comparative experimental data of denoising performance

[0155]

[0156] In the comparative experiment on the image denoising performance carried out by the present invention, we selected five common image processing methods and comprehensively evaluated them from four aspects: peak signal-to-noise ratio (PSNR), structural similarity (SSIM), average processing time, and image detail recovery score. It can be seen from the experiment that the present invention is superior to the traditional method in multiple key indicators, showing good image quality improvement effect and practicability.

[0157] In terms of image quality, the traditional hardware filtering method is only 20.8 dB and 0.75 in terms of PSNR and SSIM indicators respectively, and there is more noise residue in the processed image. The traditional convolutional neural network (CNN) method has improved, with PSNR reaching 22.5 dB and SSIM being 0.80, but there are still details missing. The basic method based on Fourier transform has a small improvement, with PSNR being 23.0 dB and SSIM being 0.82, and the effect is limited. In three independent experiments of the present invention, the average PSNR is increased to more than 26.5 dB, up to 27.1 dB at most, and the SSIM is above 0.86 in all cases, up to 0.88 at most, indicating that the clarity and structure restoration ability of the image are significantly enhanced.

[0158] In terms of the image detail recovery ability, the scores of the traditional methods are between 5.2 and 6.4, and the scores of the method of the present invention all exceed 8 points, up to 8.4 at most, reflecting the outstanding performance of the present invention in retaining key image details such as edges and textures. Especially in low-light and high-sensitivity imaging scenarios, this method effectively avoids the common image blurring and edge drift problems in traditional algorithms.

[0159] In terms of processing efficiency, the average processing time of the traditional CNN algorithm is 80 milliseconds, which is difficult to meet the real-time requirements. The average processing time of the method of the present invention is controlled between 47 and 50 milliseconds, meeting the requirements of high-frame-rate monitoring and real-time image processing systems, and demonstrating excellent algorithm calculation efficiency.

[0160] Overall, the present invention extracts image features in the frequency domain through Fourier neural operators and dynamically adjusts model parameters by combining an improved jellyfish swarm optimization algorithm, successfully achieving high-quality image noise reduction and detail retention. Compared with traditional methods, the present invention has significant advantages in terms of image quality, calculation efficiency, and model robustness, and is particularly suitable for practical scenarios such as night monitoring, low-light environment imaging, and intelligent transportation, providing an efficient, stable, and popularizable noise suppression optimization solution for high-sensitivity camera modules.

[0161] The above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. An optimization method for noise suppression of a high-sensitivity camera module, characterized in that, It includes the following steps: S1. Collect the original image through a high-sensitivity camera module and preprocess the original image; S2. Perform discrete Fourier transform on the preprocessed original image to convert it into a frequency-domain image; S3. Input the frequency-domain image into the Fourier neural operator model, perform multi-level feature mapping and data fusion processing on the frequency-domain image, and form an intermediate spectrum representation after noise reduction processing; S4. Perform inverse Fourier transform on the intermediate spectrum representation to convert it into a spatial-domain image, and obtain the preliminarily noise-reduced image as a reference image; S5. Use the jellyfish swarm optimization algorithm to globally search for and optimize the key structural parameters and training hyperparameters in the Fourier neural operator model; S6. Update the Fourier neural operator model according to the parameter combination optimized by the jellyfish swarm optimization algorithm, and apply the updated Fourier neural operator model to the closed-loop iterative optimization process; S7. During the closed-loop iterative optimization process, perform quality evaluation on the noise-reduced image output by the updated Fourier neural operator model and the reference image until the preset noise reduction performance index is reached, and output the image after deep fusion noise reduction processing.

2. The noise suppression optimization method for a high-sensitivity camera module according to claim 1, wherein The specific content of S2 includes: S21. Detect the size of the preprocessed original image, and obtain the number of rows, columns and channels of the image; S22. Convert the pixel matrix of the image into a data format suitable for Fourier transform operation, and format the image data; S23. Initialize the calculation area of the two-dimensional discrete Fourier transform, set the range of frequency distribution in the frequency-domain coordinate system according to the image size, and determine the appropriate sampling interval and frequency resolution; S24. Call the Fourier transform function in the image processing module to perform the conversion operation from the spatial domain to the frequency domain on the formatted image data: Among them, f(x, y) represents the pixel value of the image at the coordinate (x, y), M and N respectively represent the width and height of the image, u and v are the frequency-domain coordinates, j is the imaginary unit, exp is the exponential function, and F(u, v) represents the frequency-domain image obtained through discrete Fourier transform; S25. Perform separation processing on the obtained frequency-domain image F(u, v) for amplitude spectrum and phase spectrum, and store the energy information and phase information in the frequency-domain image data in independent matrices respectively; S26. Perform energy normalization processing on the separated frequency-domain image data, adjust the energy distribution of the spectrum representation through standardization operation, and finally output the standardized frequency-domain image as the input of the Fourier neural operator model.

3. A method for optimizing noise suppression of a high-sensitivity camera module according to claim 1, characterized in that, The specific content of S3 includes: S31. Receive the output standardized frequency-domain image data F(u, v); S32. Perform preliminary linear mapping on the frequency-domain image data F(u, v) to form a preliminary feature representation T(u, v); S33. Introduce a multi-scale residual spectrum attention mechanism, apply a dynamic frequency-domain adaptive transformation to the preliminary feature representation T(u, v), and obtain an enhanced frequency-domain feature expression; Among them, represents the Fourier transform operation, Θ(T(u,v)) is the frequency-domain adaptive weight function, A is the learnable weight matrix, C is the learnable bias, and λ o is the residual modulation coefficient, P(u,v) represents the enhanced frequency-domain feature expression at the frequency coordinates (u,v), ⊙ represents the element-wise multiplication operation, and μ represents the local mean; S34. Apply an adaptive second-order statistical feature fusion operation to the enhanced frequency-domain feature expression P(u, v) to form an intermediate feature representation Q(u, v): Q(u, v) = W · {P(u, v) + η o · [P(u, v) - Avg(P(u, v))] 2}+ b; where W represents the weight matrix, b represents the bias term, and η o is a learnable second-order modulation coefficient, and Avg(P(u, v)) represents an operation of calculating the mean value of the enhanced frequency-domain feature representation within a local region; S35. Perform multi-level feature fusion processing on the intermediate feature representation Q(u, v), and form the denoised intermediate spectrum representation H(u, v) through layer-by-layer stacking and fusion; S36. Output the intermediate spectrum representation H(u, v) as the final output of the Fourier neural operator model.

4. A method for optimizing noise suppression of a high-sensitivity camera module according to claim 1, characterized in that, The specific steps of S4 are as follows: S41. Receive the output intermediate spectrum representation H(u, v); S42. Perform an inverse discrete Fourier transform on the intermediate spectrum representation H(u, v) to convert it into a spatial domain image g(x, y): where g(x, y) represents the pixel value of the spatial domain image at coordinates (x, y), M and N respectively represent the width and height of the image, u and v respectively represent the horizontal and vertical coordinates in the frequency domain, j is the imaginary unit, and exp is the exponential function; S43. Perform amplitude correction processing on the obtained spatial domain image g(x, y) to adjust the amplitude distribution of each pixel value in the spatial domain image; S44. Perform brightness and contrast normalization operations on the spatial domain image g(x, y) so that the overall brightness and contrast of the spatial domain image meet the preset standards; S45. Apply edge detail enhancement processing to the normalized spatial domain image g(x, y) to strengthen the edge and texture information in the spatial domain image; S46. Determine the spatially domain image g(x, y) after amplitude correction, brightness and contrast normalization, and edge detail enhancement processing as the preliminarily denoised image and output it as a reference image.

5. A method for optimizing noise suppression of a high-sensitivity camera module according to claim 1, characterized in that The specific steps of S5 are as follows: S51. Determine the search space of the key structural parameters and training hyperparameters in the Fourier neural operator model, and construct a candidate parameter vector: X = [α, β, γ, δ, η, B, λ, d, ξ]; where α represents the number of frequency domain dimension truncations, β represents the width of the projection layer, γ represents the number of layers, δ represents the skip connection design parameter, η represents the learning rate, B represents the batch size, λ represents the regularization parameter, d represents the Dropout ratio, and ξ represents the rotational motion adjustment coefficient; S52. Randomly initialize the candidate solution set where N represents the number of candidate solutions, represents the parameter combination of the i-th candidate solution at the initial time t = 0; S53. Define a fitness function f(X) to evaluate the quality of candidate solutions: f(X) = ω1·PSNR(X) + ω2·SSIM(X); where PSNR(X) represents the peak signal-to-noise ratio, SSIM(X) represents the structural similarity index, and ω1 and ω2 are set weight coefficients; S54. For each candidate solution where t represents the iteration number, define the local neighborhood set as the set of all candidate solutions with Euclidean distance less than the preset threshold δ th and calculate the local information fusion vector: Among them, represents the Euclidean distance between and σ ′ is the local diffusion adjustment parameter, and exp is the exponential function; S55. For each candidate solution Calculate the rotational motion vector The rotational motion vector is constructed based on the angular difference between the candidate solution and the local neighborhood center C i as follows: Among them, C i is the local neighborhood center, θ i represents the candidate solution and the phase angle difference between the local neighborhood center C i ; represents the rotation operation of the vector with θ i as the rotation angle, and ξ is the rotation motion adjustment coefficient; S56. For each candidate solution perform position update, adopting an improved hybrid motion strategy: Among them, represents the candidate solution with the highest fitness in the current iteration, is a random vector within the interval [-1, 1], ρ is the global exploration coefficient, σ is the random perturbation coefficient, τ is the local information fusion coefficient, and ψ is the rotational motion adjustment coefficient. represents the updated position of the i-th candidate solution at the (t + 1)-th iteration; S57. For each updated candidate solution Calculate the fitness S58. Determine whether the preset termination condition is met, i.e., the maximum number of iterations T max or the fitness convergence threshold is reached. If not, set t = t + 1 and return to S54. If so, proceed to the next step; S59. Output the candidate solution X for which the fitness function f(X) attains the optimal value in the last iteration. opt As the optimized parameter combination of the Fourier neural operator model.

6. The noise suppression optimization method for a high-sensitivity camera module according to claim 1, characterized in that The specific steps of S6 are as follows: S61. Receive the output optimal parameter combination, and the optimal parameter combination includes the structural parameters and training hyperparameters of the Fourier neural operator model; S62. Apply the optimal parameter combination to initialize the Fourier neural operator model to complete the reconstruction of the Fourier neural operator model structure and the update of the training strategy; S63. Based on the updated Fourier neural operator model structure, re-perform feature mapping and denoising processing on the frequency domain image and output a new intermediate spectrum representation; S64. Perform an inverse Fourier transform on the new intermediate spectrum representation to generate the denoised image output in the current iteration round; S65. Perform image quality evaluation on the current denoised image and the obtained reference image, and obtain the performance results in terms of the peak signal-to-noise ratio and structural similarity index; S66. Determine whether the output of the current Fourier neural operator model meets the set noise reduction performance standard according to the evaluation results. If not, use the current model parameters as the input for the next round of closed-loop iteration, and continue to execute the parameter optimization and Fourier neural operator model update process. If so, complete the optimization of the Fourier neural operator model and perform the output.

Images