Wide-angle lens edge image quality partition optimization method

Abstract

The application discloses a wide-angle lens edge image quality partition optimization method, and relates to the field of lens edge image quality optimization, and comprises the steps of establishing a reference degradation intensity map as a lens reference prior template; constructing a continuous field distortion mapping model to output a radial degradation gradient field and a radial chromatic dispersion offset field; constructing a lightweight U-Net variant network to generate a pixel-by-pixel correction matrix; performing resampling mapping and chromatic aberration decoupling transformation on the original wide-angle lens image by using the pixel-by-pixel correction matrix to output an optimized image; and analyzing the image quality improvement degree of the optimized image in the edge region to trigger a model correction mechanism to optimize the training effect of the lightweight U-Net variant network. The application has the advantages that by fusing optical measurement prior and adaptive soft partition learning and introducing residual feedback closed-loop optimization, abrupt change-free, high-precision and adaptive correction of the edge image quality of the wide-angle lens are realized.

Description

Technical Field

[0001] This invention relates to the field of edge image quality optimization for lenses, specifically to a method for edge image quality partitioning optimization for wide-angle lenses. Background Technology

[0002] Wide-angle lenses are widely used in image acquisition due to their large field of view. However, constrained by the physical limitations of optical design, wide-angle lenses generally suffer from problems such as reduced resolution, astigmatism, field curvature, and chromatic aberration at the edges of the image field, resulting in inferior image quality at the edges compared to the center area. Existing post-processing digital correction methods mainly include: static zonal correction based on lens contours, edge sharpening enhancement, and end-to-end image inpainting based on deep learning.

[0003] Existing methods suffer from the following common drawbacks: First, the static hard partitioning strategy divides the image field into a central region, a transition region, and an edge region, with each region using fixed correction parameters, leading to abrupt changes in image quality at the region boundaries and the formation of transition stripes that are sensitive to the human eye. Second, it does not consider that different aberrations (such as field curvature and chromatic aberration) have different spatial degradation distribution patterns, and a single correction parameter cannot compensate for multiple aberrations simultaneously. Third, it lacks the ability to adapt to shooting conditions (aperture, object distance), and the edge degradation characteristics of the same lens vary greatly under different parameters, resulting in poor generalization ability of the static partitioning model. Fourth, although the pure data-driven deep learning model can achieve a smooth transition, it ignores physical optics priors, leading to artifacts or detail distortion after correction. Summary of the Invention

[0004] To address the aforementioned technical issues, a method for optimizing edge image quality partitioning in wide-angle lenses is provided. This method constructs a five-stage architecture: "optical measurement—field mapping modeling—adaptive soft partitioning—cooperative correction—closed-loop feedback." This architecture enables a complete process from degradation measurement to pixel-by-pixel adaptive correction, effectively eliminating abrupt changes in partition boundaries, adapting to multiple aberration superpositions, and maintaining physical authenticity.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for optimizing edge image quality in wide-angle lenses, comprising:

[0007] Obtain the original edge degradation distribution map of the wide-angle lens at the reference aperture and reference object distance, and establish a reference degradation intensity map as the lens reference prior template;

[0008] Based on the radial basis function neural network, a continuous field distortion mapping model is constructed, which outputs the radial degradation gradient field and the radial dispersion offset field based on the baseline degradation intensity map.

[0009] A lightweight U-Net variant network is constructed, which takes the radially degraded gradient field as input to generate a multi-scale smooth partition confidence map. The radially dispersed offset field and the partition confidence map are combined to generate a pixel-wise correction matrix.

[0010] The original wide-angle lens image is resampled and decoupled by color difference using a pixel-by-pixel correction matrix, and the optimized image is output.

[0011] The degree of image quality improvement in the edge region of the optimized image is analyzed, triggering a model correction mechanism to optimize the training effect of the lightweight U-Net variant network.

[0012] Preferably, the step of obtaining the original edge degradation distribution map of the wide-angle lens at the reference aperture and reference object distance, and establishing a reference degradation intensity map as a reference prior template for the lens includes:

[0013] Multiple test target points are arranged in a plane perpendicular to the optical axis in front of the lens, according to the rule of equal solid angles;

[0014] By focusing the lens on the plane, each test target point is imaged by the lens, and multiple discrete measurement position points are formed in the image field.

[0015] Using an optical transfer function measurement platform, each measurement point in the image field is positioned sequentially. At each point, modulation transfer function, relative illumination, and magnification chromatic aberration data are collected. The modulation transfer function is taken as the value at the Nyquist frequency at that point, and the magnification chromatic aberration is converted into pixel offset.

[0016] Based on the modulation transfer function value, relative illumination, and magnification color difference, the degradation intensity value at the pixel coordinates is calculated using a weighted summation formula.

[0017] Centered on each discrete sampling point, its degradation intensity value is used as the known sample at that point. Bicubic interpolation is used to interpolate all pixel positions in the entire image to obtain the globally continuously differentiable degradation intensity distribution.

[0018] After interpolation, each pixel coordinate obtains a degradation intensity value in the range [0,1]. This value is stored in grayscale form to establish a baseline degradation intensity map, which serves as the baseline prior template for the lens.

[0019] Preferably, the step of constructing a continuous field distortion mapping model based on a radial basis function neural network, and outputting a radial degradation gradient field and a radial dispersion offset field based on a baseline degradation intensity map, includes:

[0020] Based on the polar coordinates of each pixel in the degradation intensity map and the current shooting conditions parameters, a four-dimensional input vector is constructed, where the polar coordinates include radial distance and polar angle, and the shooting conditions parameters include aperture value and object distance.

[0021] Using the degradation intensity values ​​measured under each working condition as the output target, we iterate through all pixels in the degradation intensity map under all working conditions to form a multi-working condition training sample set.

[0022] A continuous field distortion mapping model is constructed. The continuous field distortion mapping model is based on a radial basis function neural network structure and adopts a single hidden layer structure. The number of neurons in the hidden layer is determined by the modified Akaike information content criterion.

[0023] Based on the given model of wide-angle lens, set the reference width, reference aperture, and reference object distance;

[0024] Calculate the basis function width based on the reference width, reference aperture, and reference object distance;

[0025] The continuous field distortion mapping model sets up two independent parallel output branches: the first branch fits the continuous degradation intensity function and outputs the predicted degradation intensity value, and the second branch independently fits the radial dispersion offset vector and outputs the radial dispersion offset field.

[0026] The first branch uses the mean square error of the predicted degradation intensity as the loss, and the second branch uses the mean square error of the radial dispersion offset vector as the loss. The total loss is the sum of the two losses.

[0027] With the goal of minimizing the total loss, the Adam optimizer is used, and iterative training is performed until the loss function converges. During the training process, the center point remains unchanged, and only the weight parameters of the output layer are optimized.

[0028] After training, the input vectors of all pixels under any working condition are fed into the network pixel by pixel. The forward propagation is used to obtain the predicted value of the degradation intensity and the radial dispersion offset vector at each pixel. The radial degradation gradient value is obtained by taking the derivative of the predicted value of the degradation intensity.

[0029] The radial degradation gradient value and the radial dispersion offset vector are stored in image form to obtain the radial degradation gradient field map and the radial dispersion offset field map.

[0030] Preferably, the construction of the lightweight U-Net variant network, taking the radially degraded gradient field as input, generating a multi-scale smooth partition confidence map, and combining the radial dispersion offset field with the partition confidence map to generate a pixel-wise correction matrix includes:

[0031] A lightweight U-Net variant network is constructed, which consists of an encoder, a decoder, skip connections, and inputs and outputs;

[0032] The encoder employs multi-layer downsampling, with each layer consisting of a depthwise separable convolution. Each depthwise separable convolution is followed by batch normalization and a ReLU activation function.

[0033] The decoder employs multi-layer upsampling, with each layer consisting of a transposed convolution followed by batch normalization and a ReLU activation function.

[0034] The encoder output and the upsampled feature map of the decoder are concatenated by channel. After each skip connection, a channel attention mechanism is introduced to perform global average pooling on the concatenated feature map to obtain the channel description vector. Attention weights are generated through two fully connected layers and multiplied by the original feature map channel by channel.

[0035] The network input is a radially degraded gradient field, and the output channels correspond to the central region, transition region and edge region respectively. Each channel outputs the confidence probability map of the pixel belonging to the corresponding region.

[0036] The network training uses a combined loss function, which consists of three parts: smoothing L1 loss, total variational regularization term and image reconstruction loss. Among them, smoothing L1 loss is defined as the smoothing L1 distance per pixel and per channel, that is, calculating the smoothing L1 error between the network output confidence and the real confidence map per pixel and per channel.

[0037] Using optical design software, generate real partition confidence maps required for training offline;

[0038] A set of learnable correction parameter base values ​​is preset for each partition, including: radial stretch coefficient, tangential compression coefficient, chromatic aberration decoupling coefficient, and sharpening gain coefficient;

[0039] For each pixel, the baseline values ​​of the correction parameters for each partition are weighted and summed using the confidence probability as the weight to obtain the comprehensive correction parameter vector for that pixel.

[0040] The integrated correction parameter vector and the radial dispersion offset field are concatenated along the channel dimension to form the input feature vector. This vector is then input into a single-layer fully connected network, which outputs a pixel-wise correction matrix. The elements of the pixel-wise correction matrix are the radial stretching coefficient, the tangential compression coefficient, the chromatic aberration decoupling coefficient, and the sharpening gain coefficient.

[0041] Preferably, the step of performing resampling mapping and chromatic aberration decoupling transformation on the original wide-angle lens image using a pixel-by-pixel correction matrix, and outputting an optimized image, includes:

[0042] For each pixel in the original image, calculate its polar coordinates relative to the center of the image field;

[0043] Where the radial distance is equal to the Euclidean distance from the pixel to the center of the image field divided by the maximum image height, and the polar angle is equal to the angle of the line connecting the point and the center relative to the horizontal axis.

[0044] Based on the radial stretching coefficient in the pixel-by-pixel correction matrix, a resampling mapping rule is constructed. Specifically, the resampling mapping rule is: new radial distance = original radial distance × (1 + radial stretching coefficient), and new polar angle = original polar angle.

[0045] Based on the calculated polar coordinates, the mapped polar coordinates are obtained according to the resampling mapping rule, and the mapped polar coordinates are converted back to rectangular coordinates to obtain the target sampling position of each pixel;

[0046] Based on the four integer pixels surrounding the target sampling location, the brightness value of that point is calculated using bilinear interpolation to generate a corrected brightness channel image;

[0047] According to the color difference decoupling coefficient in the pixel-by-pixel correction matrix, the color difference decoupling coefficient is the distance that the red channel and the blue channel need to be offset from each other in the radial direction;

[0048] For each non-centered pixel, calculate the radial unit vector pointing from the center of the image field to the current pixel;

[0049] The offsets of the red and blue channels are calculated based on the radial unit vector. The offset is equal to the product of the chromatic aberration decoupling coefficient and the radial unit vector. The green channel is not offset. The direction of its offset is determined by the chromatic aberration characteristics of the given lens itself and is pre-calibrated offline.

[0050] Based on the obtained channel offset, bilinear interpolation sampling is performed on the red channel and blue channel at the offset coordinates to obtain the corrected red channel and blue channel values.

[0051] The green channel is kept unchanged and merged with the corrected red and blue channels to obtain a color image after color difference decoupling;

[0052] Based on the color image after color difference decoupling, the original grayscale image is obtained based on the grayscale value formula, and the original grayscale image is Gaussian blurred to obtain the blurred grayscale image.

[0053] For each pixel, calculate the ratio of the red channel to the original grayscale value, the ratio of the green channel to the original grayscale value, and the ratio of the blue channel to the original grayscale value.

[0054] Subtract the original grayscale image from the blurred grayscale image pixel by pixel to obtain the detail component. Then, multiply the detail component by the product of the sharpening gain coefficient and the sum of 1 and the tangential compression coefficient, and add it back to the original grayscale image to obtain the sharpened grayscale value.

[0055] Multiply the sharpened grayscale value by the corresponding red, green, and blue ratios respectively to obtain the values ​​of the red, green, and blue channels of each pixel after sharpening.

[0056] The values ​​of the red, green, and blue channels of each pixel after sharpening are limited to the range of 0 to 255, and then recombined into a color image to obtain the optimized output image.

[0057] Preferably, the step of analyzing the degree of image quality improvement in the edge region of the optimized image and triggering a model correction mechanism to optimize the training effect of the lightweight U-Net variant network includes:

[0058] Extract all pixels in the edge region of the optimized image, whereby the edge region is defined as the set of pixels with a polar radius greater than 0.8 times the maximum image height;

[0059] Calculate the no-reference quality evaluation index Q for the edge region, wherein the no-reference quality evaluation index adopts the average gradient magnitude;

[0060] The quality index Q is compared with the preset threshold T. If Q is less than the threshold T, it means that the image quality of the edge area does not meet the preset standard, and the model correction mechanism is triggered; otherwise, all network parameters remain unchanged.

[0061] Each time the model correction mechanism is triggered and training is completed, the optimized image is regenerated and the quality index Q is recalculated. The above comparison and correction process is repeated until Q≥T or the maximum number of iterations is reached.

[0062] The model correction mechanism specifically includes:

[0063] Keep all loss function hyperparameters unchanged;

[0064] Increase the sampling weight of pixels in the loss function for edge regions;

[0065] The encoder, decoder, channel attention mechanism, and fully connected layer parameters of the U-Net variant network were fine-tuned through a finite number of iterations.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] This invention provides a method for optimizing edge image quality in wide-angle lenses. It involves offline calibration of a given wide-angle lens model to establish a baseline degradation intensity map as a priori template. A continuous-field distortion mapping model with aperture and object distance adaptive correction is constructed using a radial basis function neural network, outputting a radial degradation gradient field and a radial dispersion offset field applicable to all conditions. A lightweight U-Net is then used to generate a smooth partition confidence map, and the correction parameters are fused using confidence weights and combined with the radial offset field to output a pixel-by-pixel correction matrix. Edge image quality optimization is achieved through resampling mapping, chromatic aberration decoupling, and adaptive sharpening. A closed-loop feedback mechanism driven by edge image quality analysis based on no-reference quality assessment is introduced to continuously optimize the U-Net training effect. This method effectively overcomes the boundary abruptness defect of traditional hard partition correction, and can simultaneously compensate for multiple aberrations such as field curvature, astigmatism, and magnification chromatic aberration. It possesses baseline prior and all-condition adaptive offline calibration and online real-time correction capabilities, improving the smoothness, sharpness, and overall image quality of wide-angle lenses at the edges. Attached Figure Description

[0068] Figure 1 This is a flowchart of a method for optimizing the edge image quality of a wide-angle lens according to the present invention;

[0069] Figure 2 The flowchart of the continuous field distortion mapping model based on radial basis function neural network of the present invention, and the output of radial deterioration gradient field and radial dispersion offset field is shown below.

[0070] Figure 3 This is a flowchart of the process for generating the pixel-by-pixel correction matrix according to the present invention;

[0071] Figure 4 The flowchart of the present invention describes the process of performing resampling mapping and chromatic aberration decoupling transformation on the original wide-angle lens image using a pixel-by-pixel correction matrix, and outputting the optimized image.

[0072] Figure 5 This is a structural diagram of the electronic device proposed in this invention;

[0073] Figure 6 This is a schematic diagram of the structure of the computer-readable storage medium proposed in this invention. Detailed Implementation

[0074] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0075] Example 1

[0076] Reference Figure 1 As shown, a method for optimizing edge image quality in a wide-angle lens includes:

[0077] Obtain the original edge degradation distribution map of the wide-angle lens at the preset aperture and object distance, and establish a standardized degradation intensity map;

[0078] Based on a radial basis function neural network, a continuous field distortion mapping model is constructed, which outputs a radially degraded gradient field and a radially dispersed offset field.

[0079] A lightweight U-Net variant network is constructed, which takes the radially degraded gradient field as input to generate a multi-scale smooth partition confidence map. The radially dispersed offset field and the partition confidence map are combined to generate a pixel-wise correction matrix.

[0080] The original wide-angle lens image is resampled and decoupled by color difference using a pixel-by-pixel correction matrix, and the optimized image is output.

[0081] The degree of image quality improvement in the edge region of the optimized image is analyzed, triggering a model correction mechanism to optimize the training effect of the lightweight U-Net variant network.

[0082] The process of obtaining the original edge degradation distribution map of the wide-angle lens at the reference aperture and reference object distance, and establishing a reference degradation intensity map as a reference prior template for the lens, includes:

[0083] Multiple test target points are arranged in a plane perpendicular to the optical axis in front of the lens, according to the rule of equal solid angles;

[0084] By focusing the lens on the plane, each test target point is imaged by the lens, and multiple discrete measurement position points are formed in the image field.

[0085] Using an optical transfer function measurement platform, each measurement point in the image field is positioned sequentially. At each point, modulation transfer function, relative illumination, and magnification chromatic aberration data are collected. The modulation transfer function is taken as the value at the Nyquist frequency at that point, and the magnification chromatic aberration is converted into pixel offset.

[0086] Based on the modulation transfer function value, relative illumination, and magnification color difference, the degradation intensity value at the pixel coordinates is calculated using a weighted summation formula.

[0087] Centered on each discrete sampling point, its degradation intensity value is used as the known sample at that point. Bicubic interpolation is used to interpolate all pixel positions in the entire image to obtain the globally continuously differentiable degradation intensity distribution.

[0088] After interpolation, each pixel coordinate obtains a degradation intensity value in the range [0,1]. This value is stored in grayscale form to establish a baseline degradation intensity map, which serves as the baseline prior template for the lens.

[0089] This approach does not require repeated execution for every actual optimization. Instead, it performs offline calibration only once for a given wide-angle lens model. The generated baseline degradation intensity map serves as a fixed prior template for all subsequent edge image quality optimization operations on that lens, thus avoiding repeated measurements on the test bench while ensuring optical prior accuracy. To achieve this, two key parameters need to be defined during data acquisition: first, the modulation transfer function (MTF), which is a value taken at the Nyquist frequency. The MTF varies with spatial frequency, and the Nyquist frequency is uniquely determined by the sensor pixel size. It is the highest spatial frequency at which the sensor can sample without aliasing, and reading the MTF value at this frequency best represents the edge-limit resolution capability; second, the position of the test target point, which refers to the discrete position coordinates on the lens's imaging plane. Each point corresponds to a specific image height and azimuth angle, used to quantitatively describe the continuous degradation pattern of image quality from the center to the edges.

[0090] The bicubic interpolation uses the degradation intensity values ​​of the surrounding sixteen sampling points for weighted fitting, which can ensure that the interpolated intensity map is spatially continuous and second-order differentiable, thus smoothly reflecting the natural gradual change of degradation intensity from the center of the image field to the edge.

[0091] The expression for calculating the degradation intensity value at pixel coordinates is:

[0092]

[0093] In the formula, For coordinates The degradation intensity value at a given location ranges from [0,1]. A larger value indicates more severe image quality degradation at that location. In order to be in The modulation transfer function value measured at the Nyquist frequency is taken, and its range is [0,1]. A higher value indicates better contrast preservation. In order to be in The relative illuminance at a point, ranging from [0,1], represents the proportion of the brightness at that point relative to the center of the image field. In order to be in The magnification color difference measured at this point is the offset in pixels, typically the offset of the red / blue channel relative to the green channel. This represents the maximum measured chromatic aberration offset across all sampling points within the entire image field, used to normalize the chromatic aberration to the [0,1] interval. , , These are the modulation transfer function, relative illumination, and magnification chromatic difference weighting coefficients, respectively. Their calibration method is as follows: based on the simulated image of the optical design software, with the peak signal-to-noise ratio as the objective function, a grid search method is used to search for the optimal weight combination of image quality indicators in the edge region within the interval [0,1] with a step size of 0.1.

[0094] Example 2

[0095] Based on Example 1, referring to Figure 2 As shown, the continuous field distortion mapping model constructed based on the radial basis function neural network, using the baseline degradation intensity map as a basis, outputs the radial degradation gradient field and the radial dispersion offset field, including:

[0096] Based on the polar coordinates of each pixel in the degradation intensity map and the current shooting conditions parameters, a four-dimensional input vector is constructed, where the polar coordinates include radial distance and polar angle, and the shooting conditions parameters include aperture value and object distance.

[0097] Using the degradation intensity values ​​measured under each working condition as the output target, we iterate through all pixels in the degradation intensity map under all working conditions to form a multi-working condition training sample set.

[0098] A continuous field distortion mapping model is constructed. The continuous field distortion mapping model is based on a radial basis function neural network structure and adopts a single hidden layer structure. The number of neurons in the hidden layer is determined by the modified Akaike information content criterion.

[0099] Based on the given model of wide-angle lens, set the reference width, reference aperture, and reference object distance;

[0100] Calculate the basis function width based on the reference width, reference aperture, and reference object distance;

[0101] The continuous field distortion mapping model sets up two independent parallel output branches: the first branch fits the continuous degradation intensity function and outputs the predicted degradation intensity value, and the second branch independently fits the radial dispersion offset vector and outputs the radial dispersion offset field.

[0102] The first branch uses the mean square error of the predicted degradation intensity as the loss, and the second branch uses the mean square error of the radial dispersion offset vector as the loss. The total loss is the sum of the two losses.

[0103] With the goal of minimizing the total loss, the Adam optimizer is used, and iterative training is performed until the loss function converges. During the training process, the center point remains unchanged, and only the weight parameters of the output layer are optimized.

[0104] After training, the input vectors of all pixels under any working condition are fed into the network pixel by pixel. The forward propagation is used to obtain the predicted value of the degradation intensity and the radial dispersion offset vector at each pixel. The radial degradation gradient value is obtained by taking the derivative of the predicted value of the degradation intensity.

[0105] The radial degradation gradient value and the radial dispersion offset vector are stored in image form to obtain the radial degradation gradient field map and the radial dispersion offset field map.

[0106] To further illustrate the approach of this solution, a simulation example is provided below to specifically demonstrate its implementation:

[0107] The radial distance is normalized to the interval [0,1] with the image field center as 0 and the maximum image height as 1. The polar angle ranges from 0 to 2π.

[0108] The method for determining the number of hidden layer neurons is as follows: set the value of k to be in the range of 5 to 100, train a radial basis function neural network for each candidate k, calculate its modified Akaike Information Criterion (AICc) value, and select the k corresponding to the smallest AICc as the final number of neurons. Each hidden layer neuron corresponds to a Gaussian radial basis function, and its center point is obtained by K-means clustering of the four-dimensional vectors of radial distance, polar angle, aperture and object distance in the input space. The number of clusters is k.

[0109] The basis function width is determined by the reference width, reference aperture, current aperture, reference object distance, and current object distance, using the following formula: ,in, The width of the basis functions. As the base width, As the reference aperture, For the current aperture, The current object distance, Reference object distance;

[0110] During training, the center point and basis function width remain fixed, and only the output layer weight parameters are optimized. Since the center point has covered the working space and the basis function width is dynamically adjusted with aperture and object distance, the model can respond to the degradation field changes under different working conditions through linear combination, and achieve full-condition adaptive mapping.

[0111] For a given model of wide-angle lens, a multi-condition training sample set is used. The degradation intensity map at a certain preset reference object distance is used as a reference. The grid search method is used to find the value that minimizes the mean of the radial basis function neural network fitting error under all test conditions within a set range. This value is the reference width of the lens, and the reference aperture is the aperture value most commonly used by the lens.

[0112] The continuous field distortion mapping model has two independent parallel output branches: The first branch takes radial distance, polar angle, aperture, and object distance as inputs and fits the continuous degradation intensity function under the given conditions. During training, the loss function of this branch is the mean square error between the predicted degradation intensity value and the baseline true value. The radial degradation gradient field is not supervised in the training loss, but is obtained after the network training is completed. The current conditions parameters, aperture and object distance, are fixed, and the partial derivative of the continuous degradation intensity function with respect to the radial distance is automatically calculated using automatic differentiation to obtain the radial degradation gradient value at each pixel. Its physical meaning is the rate of change of degradation intensity per unit length along the radial direction. The second branch independently fits the radial dispersion offset vector and outputs a two-dimensional vector. The first component represents the number of pixels offset in the horizontal direction of the red / blue channel relative to the green channel, and the second component represents the number of pixels offset in the vertical direction. The target value of the radial dispersion offset is calculated by the offset of the red / blue channel relative to the green channel in the calibrated image.

[0113] It should be noted that the number of neurons in the hidden layer controls the density of the basis function centers, and the width of the basis function controls the radius of influence of each center. During training, these are not learned parameters but are only used as hyperparameters to control the smoothness bias of the network. Their effects are reflected in two aspects: First, they directly affect the magnitude of the basis function values, thereby changing the contribution weight of each basis function to different samples in the linear combination; second, through the regularization effect of the basis function width, they constrain the range of variation of the first derivative of the output surface, which is equivalent to adding an implicit smoothing prior to the loss function. Therefore, although the basis function width does not participate in gradient updates, its value directly determines whether the network can achieve a low training error with limited weights.

[0114] Example 3

[0115] Based on Example 1, referring to Figure 3 As shown, the generation of the pixel-by-pixel correction matrix includes:

[0116] A lightweight U-Net variant network is constructed, which consists of an encoder, a decoder, skip connections, and inputs and outputs;

[0117] The encoder employs multi-layer downsampling, with each layer consisting of a depthwise separable convolution. Each depthwise separable convolution is followed by batch normalization and a ReLU activation function.

[0118] The decoder employs multi-layer upsampling, with each layer consisting of a transposed convolution followed by batch normalization and a ReLU activation function.

[0119] The encoder output and the upsampled feature map of the decoder are concatenated by channel. After each skip connection, a channel attention mechanism is introduced to perform global average pooling on the concatenated feature map to obtain the channel description vector. Attention weights are generated through two fully connected layers and multiplied by the original feature map channel by channel.

[0120] The network input is a radially degraded gradient field, and the output channels correspond to the central region, transition region and edge region respectively. Each channel outputs the confidence probability map of the pixel belonging to the corresponding region.

[0121] The network training uses a combined loss function, which consists of three parts: smoothing L1 loss, total variational regularization term and image reconstruction loss. Among them, smoothing L1 loss is defined as the smoothing L1 distance per pixel and per channel, that is, calculating the smoothing L1 error between the network output confidence and the real confidence map per pixel and per channel.

[0122] Using optical design software, generate real partition confidence maps required for training offline;

[0123] A set of learnable correction parameter base values ​​is preset for each partition, including: radial stretch coefficient, tangential compression coefficient, chromatic aberration decoupling coefficient, and sharpening gain coefficient;

[0124] For each pixel, the baseline values ​​of the correction parameters for each partition are weighted and summed using the confidence probability as the weight to obtain the comprehensive correction parameter vector for that pixel.

[0125] The integrated correction parameter vector and the radial dispersion offset field are concatenated along the channel dimension to form the input feature vector. This vector is then input into a single-layer fully connected network, which outputs a pixel-wise correction matrix. The elements of the pixel-wise correction matrix are the radial stretching coefficient, the tangential compression coefficient, the chromatic aberration decoupling coefficient, and the sharpening gain coefficient.

[0126] This scheme generates a spatially continuous and boundary-smooth soft partition confidence map by inputting the radially degraded gradient field into a lightweight U-Net variant network. Then, the preset correction parameter base values ​​of each partition are weighted and fused with confidence as the weight. Finally, it is input into a fully connected network together with the radial dispersion offset field to output a pixel-wise correction matrix that can simultaneously compensate for radial and tangential aberrations and has no boundary abrupt changes.

[0127] To further illustrate the approach of this solution, a simulation example is provided below to specifically demonstrate its implementation:

[0128] This embodiment constructs a lightweight U-Net variant network for semantic partitioning of the radially degraded gradient field. The network structure is as follows:

[0129] The encoder uses 4 layers of downsampling, each layer consisting of a depthwise separable convolution. The kernel size of each convolution is 3×3, the stride is 2, and the number of channels is 32, 64, 128, and 256 respectively. Each depthwise separable convolution is followed by batch normalization and ReLU activation function.

[0130] Decoder: It adopts 4 layers of upsampling, each layer consists of a transposed convolution with a kernel size of 3×3, a stride of 2, and channel numbers of 256, 128, 64, and 32 respectively, followed by batch normalization and ReLU activation function;

[0131] Skip connection: The corresponding layer output of the encoder is concatenated with the upsampled feature map of the decoder by channel. After each skip connection, a channel attention mechanism is introduced to perform global average pooling on the concatenated feature map to obtain the channel description vector. Through two fully connected layers, the channel reduction factor is 16 to generate attention weights. The weights are multiplied with the original feature map channel by channel to achieve feature recalibration.

[0132] Input and output: The network input is a radially degraded gradient field, which is a single-channel image. The output has 3 channels, corresponding to the central region, transition region and edge region respectively. After Softmax normalization, each channel outputs the confidence probability map of the pixel belonging to the corresponding region.

[0133] The network training uses a combined loss function, which consists of three parts: smoothing L1 loss, total variational regularization term and image reconstruction loss.

[0134] Smoothed L1 loss: used to measure the difference between the prediction confidence and the supervision label, defined as the smoothed L1 distance per pixel and per channel;

[0135] Total variational regularization term: Forces the confidence levels of adjacent pixels to change continuously and gradually, avoiding unnatural jumps at partition boundaries. Its calculation formula is:

[0136]

[0137] In the formula, To reduce the confidence probability of the output of the lightweight U-Net variant network, it is represented as... The confidence value belonging to the k-th partition, where k is the partition index, corresponding to the central region, transition region, and edge region, and the first item. The second term represents the absolute value of the confidence difference between adjacent pixels in the horizontal direction. It is the absolute value of the confidence difference between adjacent pixels in the vertical direction;

[0138] Image reconstruction loss is only used during offline training. Specifically, optical modeling with the same parameters as the wide-angle lens to be corrected is built using optical design software. Ray tracing is performed across the entire field of view to generate pairs of training data in batches. Each set of data includes an original wide-angle lens image generated by a lens model with actual aberrations (input any high-resolution test image, such as an ISO 12233 standard resolution board or a random texture image, which is then imaged by the lens model), and an ideal sharp image without distortion or aberrations generated by the same optical design software after all aberrations are turned off (setting the lens as an aberration-free ideal lens model) (using the same test image as input). For real-world images, image reconstruction loss does not need to be calculated during online inference; only the trained network is used for forward correction. The pixel-by-pixel correction matrix output by the fully connected network is applied to the original wide-angle lens image to obtain the corrected image. The mean absolute error between the corrected image and the corresponding ideal sharp image is calculated as the image reconstruction loss term. During the actual online inference stage, the trained network is directly applied to the real-world images, and this loss does not need to be calculated again.

[0139] The total loss function is obtained by weighted summation of these three terms. The weights are obtained by grid search method, using the image reconstruction error on the validation set and the peak signal-to-noise ratio as the evaluation index. The three weights are combined and searched to select the weight combination that optimizes the performance of the validation set.

[0140] By accumulating the absolute value of the confidence difference between adjacent pixels for all pixels and all partitions, and adding it as a regularization term to the loss function, the confidence values ​​of adjacent pixels are forced to change continuously, thereby eliminating hard partition boundaries and generating a soft partition confidence map with smooth transition.

[0141] The real-world confidence maps required for training are generated offline using optical design software. The generation method is as follows: A lens model with the same parameters as the wide-angle lens to be calibrated is established in Zemax optical design software. For multiple field-of-view points across the entire field of view, 50 points are uniformly sampled at radial distances of 0 to 1. Ray tracing is performed at 30° intervals between polar angles per revolution. The point spread function half-width at half-maximum (H) and the magnification chromatic aberration offset (C) at each field-of-view point are extracted to form a two-dimensional feature vector (H,C). After Min-Max linear normalization to the [0,1] interval, fuzzy C-means clustering (fuzzy exponential) is used. Using a resolution of 2.0, a maximum number of iterations of 100, and a convergence threshold of 1e-5, the number of clusters is 3 (corresponding to the empirically defined central region, transition region, and edge region). After clustering, the membership degree of each pixel to the three regions is obtained, which serves as the supervision label for the soft partition confidence map. By observing the image height distribution of the clustering results, it can be verified that: the central region corresponds to the region with a polar radius less than 0.3, the edge region corresponds to the region with a polar radius greater than 0.7, and the transition region is in between. At the same time, using the same optical design software, lens aberrations are turned off at the same field of view to generate the corresponding ideal sharp image, which serves as the reference ground value in the image reconstruction loss.

[0142] A set of learnable correction parameter base values ​​is preset for each partition, which is 4-dimensional in total, including: radial stretching coefficient, tangential compression coefficient, chromatic aberration decoupling coefficient and sharpening gain coefficient. The correction parameter base values ​​are stored as a 3×4 matrix B, with rows corresponding to 3 partitions and columns corresponding to 4 parameters. B is automatically optimized as network parameters during training.

[0143] For each pixel, the confidence probability output by the lightweight U-Net is used as the weight to sum the baseline values ​​of the correction parameters for each partition, resulting in a comprehensive correction parameter vector for that pixel. This vector contains four components: radial stretching coefficient, tangential compression coefficient, chromatic aberration decoupling coefficient, and sharpening gain coefficient. The 4-dimensional comprehensive correction parameter vector and the 2-dimensional radial dispersion offset field are concatenated along the channel dimension to form a 6-dimensional input feature vector. This 6-dimensional vector is then input into a single-layer fully connected network with no hidden layers, performing a direct linear mapping. This fully connected network contains a learnable weight matrix and bias terms, outputting a 4-dimensional pixel-wise correction matrix. The four dimensions correspond to the radial stretching coefficient, tangential compression coefficient, chromatic aberration decoupling coefficient, and sharpening gain coefficient, respectively.

[0144] During end-to-end training, the weights and biases of the fully connected network, the encoder and decoder parameters of the lightweight U-Net, the fully connected layer parameters in the channel attention mechanism, and the base value matrix of the correction parameters are jointly optimized. The Adam optimizer is used for training, with an initial learning rate of 0.001 and a batch size of 8. The training is conducted for a total of 200 epochs. After training, for any input image, the network can output a pixel-wise correction matrix in real time for subsequent image remapping and chromatic aberration correction, thereby achieving simultaneous compensation for radial and tangential aberrations without generating abrupt boundary changes.

[0145] Example 4

[0146] Based on Example 1, referring to Figure 4 As shown, the step of performing resampling mapping and chromatic aberration decoupling transformation on the original wide-angle lens image using a pixel-by-pixel correction matrix, and outputting an optimized image includes:

[0147] For each pixel in the original image, calculate its polar coordinates relative to the center of the image field;

[0148] Where the radial distance is equal to the Euclidean distance from the pixel to the center of the image field divided by the maximum image height, and the polar angle is equal to the angle of the line connecting the point and the center relative to the horizontal axis.

[0149] Based on the radial stretching coefficient in the pixel-by-pixel correction matrix, a resampling mapping rule is constructed. Specifically, the resampling mapping rule is: new radial distance = original radial distance × (1 + radial stretching coefficient), and new polar angle = original polar angle.

[0150] Based on the calculated polar coordinates, the mapped polar coordinates are obtained according to the resampling mapping rule, and the mapped polar coordinates are converted back to rectangular coordinates to obtain the target sampling position of each pixel;

[0151] Based on the four integer pixels surrounding the target sampling location, the brightness value of that point is calculated using bilinear interpolation to generate a corrected brightness channel image;

[0152] According to the color difference decoupling coefficient in the pixel-by-pixel correction matrix, the color difference decoupling coefficient is the distance that the red channel and the blue channel need to be offset from each other in the radial direction;

[0153] For each non-centered pixel, calculate the radial unit vector pointing from the center of the image field to the current pixel;

[0154] The offsets of the red and blue channels are calculated based on the radial unit vector. The offset is equal to the product of the chromatic aberration decoupling coefficient and the radial unit vector. The green channel is not offset. The direction of its offset is determined by the chromatic aberration characteristics of the given lens itself and is pre-calibrated offline.

[0155] Based on the obtained channel offset, bilinear interpolation sampling is performed on the red channel and blue channel at the offset coordinates to obtain the corrected red channel and blue channel values.

[0156] The green channel is kept unchanged and merged with the corrected red and blue channels to obtain a color image after color difference decoupling;

[0157] Based on the color image after color difference decoupling, the original grayscale image is obtained based on the grayscale value formula, and the original grayscale image is Gaussian blurred to obtain the blurred grayscale image.

[0158] For each pixel, calculate the ratio of the red channel to the original grayscale value, the ratio of the green channel to the original grayscale value, and the ratio of the blue channel to the original grayscale value.

[0159] Subtract the original grayscale image from the blurred grayscale image pixel by pixel to obtain the detail component. Then, multiply the detail component by the product of the sharpening gain coefficient and the sum of 1 and the tangential compression coefficient, and add it back to the original grayscale image to obtain the sharpened grayscale value.

[0160] Multiply the sharpened grayscale value by the corresponding red, green, and blue ratios respectively to obtain the values ​​of the red, green, and blue channels of each pixel after sharpening.

[0161] The values ​​of the red, green, and blue channels of each pixel after sharpening are limited to the range of 0 to 255, and then recombined into a color image to obtain the optimized output image.

[0162] This solution directly applies the obtained pixel-by-pixel correction matrix to the original wide-angle lens image, correcting the blur caused by geometric distortion and field curvature of the wide-angle lens through resampling mapping; aligning the red and blue channels through chromatic aberration decoupling correction to eliminate magnification chromatic aberration; and enhancing edge details through adaptive sharpening, ultimately outputting an optimized image. This method meticulously optimizes the edge image quality zones of the wide-angle lens, improving image quality.

[0163] The analysis of the image quality improvement in edge regions of the optimized image, triggering a model correction mechanism to optimize the training effect of the lightweight U-Net variant network, includes:

[0164] Extract all pixels in the edge region of the optimized image, whereby the edge region is defined as the set of pixels with a polar radius greater than 0.8 times the maximum image height;

[0165] Calculate the no-reference quality evaluation index Q for the edge region, wherein the no-reference quality evaluation index adopts the average gradient magnitude;

[0166] The quality index Q is compared with the preset threshold T. If Q is less than the threshold T, it means that the image quality of the edge area does not meet the preset standard, and the model correction mechanism is triggered; otherwise, all network parameters remain unchanged.

[0167] Each time the model correction mechanism is triggered and training is completed, the optimized image is regenerated and the quality index Q is recalculated. The above comparison and correction process is repeated until Q≥T or the maximum number of iterations is reached.

[0168] The model correction mechanism specifically includes:

[0169] Keep all loss function hyperparameters unchanged;

[0170] Increase the sampling weight of pixels in the loss function for edge regions;

[0171] The encoder, decoder, channel attention mechanism, and fully connected layer parameters of the U-Net variant network were fine-tuned through a finite number of iterations.

[0172] To further illustrate the model correction mechanism, the following details the mechanism in conjunction with the implementation of this solution:

[0173] During the model training phase, corrections are performed on all images in the validation set, the average gradient magnitude Q of the edge region of each image is calculated, these Q values ​​are sorted from smallest to largest, and the pth percentile (e.g., the 20th percentile) is taken as the preset threshold T.

[0174] The values ​​of all hyperparameters of the loss function during training, including the weight coefficients of the total variational regularization term, the weight coefficients of the smoothing L1 loss, and the weight coefficients of the image reconstruction loss, are maintained. The optimal configuration determined by the validation set during the offline training phase is used, and no dynamic adjustments are made during the closed-loop feedback process to avoid damaging the stability of the trained model.

[0175] In each loss calculation during the fine-tuning phase, pixels located in the image edge region are assigned a higher weight coefficient than those in non-edge regions. This is achieved by constructing a weight matrix with the same size as the input image, where the edge region is defined as a pixel with a polar radius greater than 0.8 times the maximum image height. The weight value corresponding to the edge region pixel is set to be a multiple greater than that of the central region pixel, such as 2 times. This weight matrix is ​​then multiplied pixel by pixel into the smoothing L1 loss and the image reconstruction loss. Through this weighting strategy, the network pays more attention to the correction error of the edge region when updating parameters, thereby guiding the optimization direction of the model parameters to tilt towards improving edge performance.

[0176] When fine-tuning, a lower learning rate than that in the initial training stage is adopted, such as 0.1 times the original learning rate, and only a limited number of rounds of parameter iterative updates are performed. This can not only perform directional optimization for edge quality problems, that is, the case where the average gradient magnitude Q does not reach the threshold, but also avoid excessive perturbation or catastrophic forgetting of the overall image distribution that the network has learned. After each fine-tuning, an optimized image is regenerated and the quality index Q of the edge region is evaluated. If the threshold requirement Q<T) is still not met, the above correction process is repeated until Q≥T or the preset maximum number of loop iterations is reached.

[0177] Furthermore, the method according to the embodiment of the present application can also be implemented by means of Figure 5 the architecture of the electronic device shown. As Figure 5 shown, the electronic device 500 may include a bus 501, one or more CPUs 502, a read-only memory (ROM) 503, a random access memory (RAM) 504, a communication port 505 connected to the network, an input / output component 506, a hard disk 507, etc. The storage device in the electronic device 500, such as the ROM 503 or the hard disk 507, can store a method for optimizing the edge image quality of a wide-angle lens provided by the present application. The electronic device 500 may further include a user interface 508. Of course, Figure 5 the architecture shown is only exemplary. When implementing different devices, one or more components in the Figure 5 shown electronic device may be omitted according to actual needs.

[0178] Figure 6 is a schematic diagram of the structure of a computer-readable storage medium provided by an embodiment of the present application. As Figure 6 shown, it is a computer-readable storage medium 600 according to an embodiment of the present application. Computer-readable instructions are stored on the computer-readable storage medium 600. When the computer-readable instructions are run by a processor, a method for optimizing the edge image quality of a wide-angle lens according to the embodiment of the present application described above with reference to the accompanying drawings can be executed. The storage medium 600 includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory, etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0179] In summary, the advantages of the present invention are as follows: By integrating optical measurement prior and adaptive soft partition learning and introducing residual feedback closed-loop optimization, the edge image quality of the wide-angle lens is realized without sudden change, high-precision and adaptive correction.

[0180] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for optimizing edge image quality in wide-angle lenses, comprising: Obtain the original edge degradation distribution map of the wide-angle lens at the reference aperture and reference object distance, and establish a reference degradation intensity map as the lens reference prior template; Based on the radial basis function neural network, a continuous field distortion mapping model is constructed, which outputs the radial degradation gradient field and the radial dispersion offset field based on the baseline degradation intensity map. A lightweight U-Net variant network is constructed, which takes the radially degraded gradient field as input to generate a multi-scale smooth partition confidence map. The radially dispersed offset field and the partition confidence map are combined to generate a pixel-wise correction matrix. The original wide-angle lens image is resampled and decoupled by color difference using a pixel-by-pixel correction matrix, and the optimized image is output. The degree of image quality improvement in the edge region of the optimized image is analyzed, triggering a model correction mechanism to optimize the training effect of the lightweight U-Net variant network.

2. The method for optimizing edge image quality zoning of a wide-angle lens according to claim 1, characterized in that, The process of obtaining the original edge degradation distribution map of the wide-angle lens at the reference aperture and reference object distance, and establishing a reference degradation intensity map as a reference prior template for the lens, includes: Multiple test target points are arranged in a plane perpendicular to the optical axis in front of the lens, according to the rule of equal solid angles; By focusing the lens on the plane, each test target point is imaged by the lens, and multiple discrete measurement position points are formed in the image field. Using an optical transfer function measurement platform, each measurement point in the image field is positioned sequentially. At each point, modulation transfer function, relative illumination, and magnification chromatic aberration data are collected. The modulation transfer function is taken as the value at the Nyquist frequency at that point, and the magnification chromatic aberration is converted into pixel offset. Based on the modulation transfer function value, relative illumination, and magnification color difference, the degradation intensity value at the pixel coordinates is calculated using a weighted summation formula. Centered on each discrete sampling point, its degradation intensity value is used as the known sample at that point. Bicubic interpolation is used to interpolate all pixel positions in the entire image to obtain the globally continuously differentiable degradation intensity distribution. After interpolation, each pixel coordinate obtains a degradation intensity value in the range [0,1]. This value is stored in grayscale form to establish a baseline degradation intensity map, which serves as the baseline prior template for the lens.

3. The method of claim 2, wherein the method further comprises: The continuous field distortion mapping model constructed based on the radial basis function neural network, using the baseline degradation intensity map as a basis, outputs the radial degradation gradient field and the radial dispersion offset field, including: Based on the polar coordinates of each pixel in the degradation intensity map and the current shooting conditions parameters, a four-dimensional input vector is constructed, where the polar coordinates include radial distance and polar angle, and the shooting conditions parameters include aperture value and object distance. Using the degradation intensity values ​​measured under each working condition as the output target, we iterate through all pixels in the degradation intensity map under all working conditions to form a multi-working condition training sample set. A continuous field distortion mapping model is constructed. The continuous field distortion mapping model is based on a radial basis function neural network structure and adopts a single hidden layer structure. The number of neurons in the hidden layer is determined by the modified Akaike information content criterion. Based on the given model of wide-angle lens, set the reference width, reference aperture, and reference object distance; Calculate the basis function width based on the reference width, reference aperture, and reference object distance; The continuous field distortion mapping model sets up two independent parallel output branches: the first branch fits the continuous degradation intensity function and outputs the predicted degradation intensity value, and the second branch independently fits the radial dispersion offset vector and outputs the radial dispersion offset field. The first branch uses the mean square error of the predicted degradation intensity as the loss, and the second branch uses the mean square error of the radial dispersion offset vector as the loss. The total loss is the sum of the two losses. With the goal of minimizing the total loss, the Adam optimizer is used, and iterative training is performed until the loss function converges. During the training process, the center point remains unchanged, and only the weight parameters of the output layer are optimized. After training, the input vectors of all pixels under any working condition are fed into the network pixel by pixel. The forward propagation is used to obtain the predicted value of the degradation intensity and the radial dispersion offset vector at each pixel. The radial degradation gradient value is obtained by taking the derivative of the predicted value of the degradation intensity. The radial degradation gradient value and the radial dispersion offset vector are stored as images to obtain the radial degradation gradient field map and the radial dispersion offset field map.

4. The method of claim 3, wherein the method further comprises: The construction of the lightweight U-Net variant network, using the radially degraded gradient field as input, generates a multi-scale smooth partitioned confidence map, and combines the radial dispersion offset field with the partitioned confidence map to generate a pixel-wise correction matrix, including: A lightweight U-Net variant network is constructed, which consists of an encoder, a decoder, skip connections, and inputs and outputs; The encoder employs multi-layer downsampling, with each layer consisting of a depthwise separable convolution. Each depthwise separable convolution is followed by batch normalization and a ReLU activation function. The decoder employs multi-layer upsampling, with each layer consisting of a transposed convolution followed by batch normalization and a ReLU activation function. The encoder output and the upsampled feature map of the decoder are concatenated by channel. After each skip connection, a channel attention mechanism is introduced to perform global average pooling on the concatenated feature map to obtain the channel description vector. Attention weights are generated through two fully connected layers and multiplied by the original feature map channel by channel. The network input is a radially degraded gradient field, and the output channels correspond to the central region, transition region and edge region respectively. Each channel outputs the confidence probability map of the pixel belonging to the corresponding region. The network training uses a combined loss function, which consists of three parts: smoothing L1 loss, total variational regularization term and image reconstruction loss. Among them, smoothing L1 loss is defined as the smoothing L1 distance per pixel and per channel, that is, calculating the smoothing L1 error between the network output confidence and the real confidence map per pixel and per channel. Using optical design software, generate real partition confidence maps required for training offline; A set of learnable correction parameter base values ​​is preset for each partition, including: radial stretch coefficient, tangential compression coefficient, chromatic aberration decoupling coefficient, and sharpening gain coefficient; For each pixel, the baseline values ​​of the correction parameters for each partition are weighted and summed using the confidence probability as the weight to obtain the comprehensive correction parameter vector for that pixel. The integrated correction parameter vector and the radial dispersion offset field are concatenated along the channel dimension to form the input feature vector. This vector is then input into a single-layer fully connected network, which outputs a pixel-wise correction matrix. The elements of the pixel-wise correction matrix are the radial stretching coefficient, the tangential compression coefficient, the chromatic aberration decoupling coefficient, and the sharpening gain coefficient.

5. The method of claim 4, wherein the method further comprises: The process of performing resampling mapping and chromatic aberration decoupling transformation on the original wide-angle lens image using a pixel-by-pixel correction matrix, and outputting the optimized image includes: For each pixel in the original image, calculate its polar coordinates relative to the center of the image field; Where the radial distance is equal to the Euclidean distance from the pixel to the center of the image field divided by the maximum image height, and the polar angle is equal to the angle of the line connecting the point and the center relative to the horizontal axis. Based on the radial stretching coefficient in the pixel-by-pixel correction matrix, a resampling mapping rule is constructed. Specifically, the resampling mapping rule is: new radial distance = original radial distance × (1 + radial stretching coefficient), and new polar angle = original polar angle. Based on the calculated polar coordinates, the mapped polar coordinates are obtained according to the resampling mapping rule, and the mapped polar coordinates are converted back to rectangular coordinates to obtain the target sampling position of each pixel; Based on the four integer pixels surrounding the target sampling location, the brightness value of that point is calculated using bilinear interpolation to generate a corrected brightness channel image. According to the color difference decoupling coefficient in the pixel-by-pixel correction matrix, the color difference decoupling coefficient is the distance that the red channel and the blue channel need to be offset from each other in the radial direction; For each non-centered pixel, calculate the radial unit vector pointing from the image field center to the current pixel; The offsets of the red and blue channels are calculated based on the radial unit vector. The offset is equal to the product of the chromatic aberration decoupling coefficient and the radial unit vector. The green channel is not offset. The direction of its offset is determined by the chromatic aberration characteristics of the given lens itself and is pre-calibrated offline. Based on the obtained channel offset, bilinear interpolation sampling is performed on the red channel and blue channel at the offset coordinates to obtain the corrected red channel and blue channel values. The green channel is kept unchanged and merged with the corrected red and blue channels to obtain a color image after color difference decoupling; Based on the color image after color difference decoupling, the original grayscale image is obtained based on the grayscale value formula, and the original grayscale image is Gaussian blurred to obtain the blurred grayscale image. For each pixel, calculate the ratio of the red channel to the original grayscale value, the ratio of the green channel to the original grayscale value, and the ratio of the blue channel to the original grayscale value. Subtract the original grayscale image from the blurred grayscale image pixel by pixel to obtain the detail component. Then, multiply the detail component by the product of the sharpening gain coefficient and the sum of 1 and the tangential compression coefficient, and add it back to the original grayscale image to obtain the sharpened grayscale value. Multiply the sharpened grayscale value by the corresponding red, green, and blue ratios respectively to obtain the values ​​of the red, green, and blue channels of each pixel after sharpening. The values ​​of the red, green, and blue channels of each pixel after sharpening are limited to the range of 0 to 255, and then recombined into a color image to obtain the optimized output image.

6. The method of claim 5, wherein the method further comprises: The analysis of the image quality improvement in edge regions of the optimized image, triggering a model correction mechanism to optimize the training effect of the lightweight U-Net variant network, includes: Extract all pixels in the edge region of the optimized image, whereby the edge region is defined as the set of pixels with a polar radius greater than 0.8 times the maximum image height; Calculate the no-reference quality evaluation index Q for the edge region, wherein the no-reference quality evaluation index adopts the average gradient magnitude; The quality index Q is compared with the preset threshold T. If Q is less than the threshold T, it means that the image quality of the edge area does not meet the preset standard, and the model correction mechanism is triggered; otherwise, all network parameters remain unchanged. Each time the model correction mechanism is triggered and training is completed, the optimized image is regenerated and the quality index Q is recalculated. The above comparison and correction process is repeated until Q≥T or the maximum number of iterations is reached. The model correction mechanism specifically includes: Keep all loss function hyperparameters unchanged; Increase the sampling weight of pixels in the loss function for edge regions; The encoder, decoder, channel attention mechanism, and fully connected layer parameters of the U-Net variant network were fine-tuned through a finite number of iterations.

7. An electronic device, comprising: include: At least one processor; And, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a wide-angle lens edge image quality partitioning optimization method as described in any one of claims 1-6.

8. A computer readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a method for optimizing the edge image quality of a wide-angle lens according to any one of claims 1-6.

Images