Single-frame vignetting correction method and program product based on material perception

Abstract

This invention discloses a material-aware single-frame vignetting correction method and program product. For single-frame microscopic images, this invention first establishes material proxy partitions based on color consistency, then obtains a global vignetting model under stable partitions through two-stage vignetting field estimation, and introduces a category response index based on the global vignetting model to perform differentiated correction for different material regions. This invention does not require additional reference flat-field images and multi-frame image sequences, and can alleviate the local undercorrection and overcorrection problems caused by a single global mask, improving the consistency, stability, and applicability of single-frame microscopic image vignetting correction.

Description

Technical Field

[0001] This invention belongs to the field of image processing and microscopic imaging, and specifically relates to an image vignetting correction technique. Background Technology

[0002] In microscopic imaging systems, due to factors such as objective and tube image aberrations, uneven illumination paths, field-of-view matching errors, and imaging link assembly deviations, acquired images often exhibit a gradual decrease in brightness, contrast, and even color uniformity from the center of the field of view towards the edges. This phenomenon typically manifests as vignetting or uneven illumination. For microscopic images, vignetting not only affects the visual experience but also directly impacts regional grayscale statistics, stitching consistency, quantitative analysis of cells or tissues, and subsequent model judgment results.

[0003] Currently, common vignetting or illumination inhomogeneity correction strategies in the field of microscopy can be broadly categorized into three types. The first type is the feedforward flat-field correction method based on a reference image. This involves pre-acquiring a blank field image, estimating the flat-field distribution based on it, and then normalizing and correcting the image under test. The second type is the retrospective correction method based on multiple frames or batches of images. This method uses a group of images acquired under the same microscopic setup to jointly estimate a shared illumination or vignetting field, and then applies this uniform correction to the entire batch of images. The third type is the single-image correction method, which is commonly found for natural images and photographic images. It typically uses methods such as background fitting, segmentation modeling, gradient statistics, or low-frequency illumination field estimation to infer the vignetting distribution from a single image.

[0004] Each of the aforementioned methods has its limitations. Reference image methods typically require additional acquisition of calibration images, and these calibration images must maintain a high degree of consistency with the actual sample in terms of objective magnification, illumination conditions, exposure parameters, and optical path configuration; otherwise, correction deviations are prone to occur. Furthermore, in practical experiments, many historical or one-time acquisition data sets do not possess complete reference images. Multi-frame or batch image joint correction methods usually rely on image sequences under the same settings. While they perform well under stable acquisition conditions, they are not suitable for microscopic snapshots or rapid detection scenarios where only a single frame image can be obtained.

[0005] Furthermore, existing publicly available solutions in the field of microscopy primarily focus on flat-field calibration and multi-image joint illumination correction, while vignetting correction solutions for single-frame microscopic images are relatively few. Most existing single-image vignetting correction methods are geared towards natural images or ordinary photographic images, and their modeling premise is usually closer to homogeneous regions or uniform material responses in natural scenes. They rarely address the specific characteristics of different tissues and stained regions in microscopic images, which exhibit varying brightness attenuation responses.

[0006] In microscopic images, although different regions are simultaneously affected by vignetting, due to differences in material, staining, and imaging response, directly using a single global correction mask can easily lead to insufficient or over-correction in local areas, thus affecting tissue boundaries and subsequent quantitative analysis results. Meanwhile, repeatedly performing free segmentation and global fitting within a single frame scene can easily cause partition drift and correction instability. Therefore, there is an urgent need for a single-frame vignetting correction method that is suitable for single-frame input conditions in microscopic images, requires no additional reference images or multi-frame sequences, can adaptively model the vignetting response based on material differences, and balances correction stability and result consistency. Summary of the Invention

[0007] To address the technical problems mentioned in the background section, this invention proposes a material-aware single-frame vignetting correction method and program product.

[0008] To achieve the above-mentioned technical objectives, the technical solution of the present invention is as follows:

[0009] The material-aware single-frame vignetting correction method includes the following steps:

[0010] (1) Acquire the single-frame color image to be corrected A normalized coordinate system is established for the image, and a certain color channel c1 is selected as the logarithmic domain fitting channel; where, The horizontal pixel coordinates are Vertical pixel coordinates For color channel indexes;

[0011] (2) Initialize the cumulative base vignetting mask Category response diagram and effective vignetting mask ; Record the first The working image of the wheel is ,when hour ,when hour This is the result of the previous round of correction; for the first... The working image of the wheel is downsampled and converted to the Lab color space, and then clustered based on color consistency to obtain a material proxy category label map:

[0012]

[0013] in, For the first Downsampling results of the working image. For downsampling operators, for Color clustering function, To convert to the Lab color space, Indicates the downsampling coordinates Category label map at the location;

[0014] (3) Record for In coordinates The color channel c1 value at that location is calculated, and its logarithm (luminance) is taken to obtain:

[0015]

[0016] Logarithmic brightness It is decomposed into a global vignetting term and a category bias term. The global vignetting term describes the smooth vignetting trend across the entire field of view, and the category bias term describes the average brightness shift within the same material proxy category. Let the first... The pixel set is Then we have:

[0017]

[0018]

[0019] in, For global shading, For the first Category bias of material proxy partition, These are predefined low-order two-dimensional polynomial basis functions used to describe the trend of the vignetting field smoothly changing with spatial position. For step (1), the horizontal and vertical normalized coordinates are obtained in the coordinate system. The corresponding form below, These are the weighting coefficients of the corresponding basis functions. Let be the number of terms in the polynomial. Indicates the first The basis functions are numbered, k∈[1,K];

[0020] Determine the fitting error term, the global shading regularization term, and the class bias regularization term, and construct the total cost function based on these three terms:

[0021]

[0022]

[0023]

[0024]

[0025] in, Represents the fitting error term. This indicates a global shading regularization term. This indicates a categorical bias regularization term. Represents the total cost function. The regularization weights for the polynomial coefficients. Regularization weights for class bias terms;

[0026] By minimizing the total cost function Find each and ;

[0027] (4) The result obtained in step (3) Substituting back into the full-resolution coordinate grid, we obtain the original global vignetting logarithmic field at full resolution. Then, we perform center normalization on this original global vignetting logarithmic field to obtain the center-normalized vignetting logarithmic field. Then, construct the incremental mask for this round and update the cumulative base vignetting mask:

[0028]

[0029] in, For the incremental mask in round t, Let be the cumulative base vignetting mask for round t, and min be the minimum operation.

[0030] (5) Accumulate the base vignetting mask downsampling And estimate the category response index within each material proxy category. For each category Establish the following fitting relationship:

[0031]

[0032]

[0033]

[0034] in, The c1 value is the color channel value after downsampling from the original single-frame color image. For the intercept term, For the residual term, for Within-category mean The coefficient of determination for linear fit within the category;

[0035] Categories Mapping the label map back to full resolution yields the category response map for round t. ;

[0036] (6) Based on cumulative basic vignetting mask and category response graph Construct the first Effective vignetting mask Based on this, differential correction is performed on the original single-frame color image to obtain the current corrected image. :

[0037]

[0038] in, To prevent extremely small positive numbers with a denominator of zero;

[0039] Current corrected image After being truncated to the [0,1] interval, it serves as the working image for the next iteration. ;

[0040] (7) At the end of each iteration, calculate the center-normalized halo logarithmic field. The root mean square index is used to determine whether to continue iterating. If the iteration termination condition is met, the effective vignetting mask corresponding to the last round is set. and correct image This serves as the final single-frame vignetting correction result.

[0041] Furthermore, in step (1), the operation of establishing a normalized coordinate system for the single-frame color image to be corrected is as follows:

[0042]

[0043] in, and These represent the normalized coordinates of the horizontal and vertical axes, respectively, where W is the image width and H is the image height.

[0044] The color channel c1 is the green channel.

[0045] Furthermore, in step (3), the basis functions of the low-order two-dimensional polynomial are removed. The constant terms in the text.

[0046] Furthermore, in step (4), ,in, med indicates the median operation. For the full-resolution original global vignetting logarithmic field, Indicates the central area. The threshold value for the center area radius. , and These are the normalized horizontal and vertical coordinates obtained in step (1), respectively.

[0047] Furthermore, in step (5), the number of valid pixels for each category, Variation variance and coefficient of determination within this category As a reliability index for the fit, if at least one of these three indices fails to meet the corresponding requirements, the fit reliability for that category is deemed insufficient, and then... Otherwise, the estimated It is limited to a preset range.

[0048] Furthermore, in step (7), the center-normalized halo logarithmic field The root mean square index is calculated as follows:

[0049]

[0050] in, is the root mean square index, W is the image width, and H is the image height.

[0051] Furthermore, in step (7), if the root mean square index of the current round is not higher than the root mean square value of the previous round, and the root mean square index of the current round is greater than the preset stop threshold, the iteration continues; otherwise, the iteration stops.

[0052] Furthermore, the entire iterative process is divided into two stages. The first stage of iteration ends when the iteration stopping condition is met for the first time. At this point, the category label map obtained in the current round is... Once the final category label map no longer changes, the second phase of iteration begins until the iteration stopping condition is met again, at which point the entire iteration process ends.

[0053] Further, in step (4), the median of the central region mask is compared with the median of the edge ring region mask to determine whether the vignetting direction is reasonable. If the median of the edge ring region mask is higher than the median of the central region mask, it means that the current vignetting direction is inconsistent with the expected attenuation direction, that is, the current vignetting direction is unreasonable. The current stage is terminated and the result of the previous iteration is used as the correction result of the current stage.

[0054] In addition, the present invention also designs a computer program product, including a computer program or instructions, which, when executed by a processor, implement the above-mentioned material-aware single-frame vignetting correction method.

[0055] The beneficial effects of adopting the above technical solution are as follows:

[0056] This invention uses color consistency as a material proxy, first partitioning and then fitting the data under single-frame conditions. It then uses a category response index to model the differentiated responses of different material regions to the same basic vignetting field. Compared to using a single global mask, this method better suppresses local undercorrection and overcorrection. The invention employs a two-stage iterative structure. The first stage allows for adaptive adjustment of partitions, while the second stage fixes the material proxy labels and only updates the vignetting model and category response parameters. This effectively reduces partition drift caused by repeated free segmentation in a single-frame scene, improving iterative stability and result consistency. This invention requires no additional reference flat-field image or multi-frame image sequences, making it suitable for applications with only a single-frame input, such as historical single-frame microscopic images and rapid snapshot detection. Detailed Implementation

[0057] Step 1: Acquire a single-frame microscopic color image to be corrected. ,in The horizontal pixel coordinates are Vertical pixel coordinates For color channel index, the image width is The height is A normalized coordinate system was established for the original image, and the green channel was preferentially selected as the logarithmic domain fitting channel to obtain a more stable brightness representation.

[0058]

[0059] in, and These represent the normalized coordinates of the horizontal and vertical axes, respectively, which are used to construct the two-dimensional polynomial halo basis functions.

[0060] Step 2: Initialize the cumulative base vignetting mask Category response diagram and effective vignetting mask Record the first The working image of the wheel is ,when hour ,when hour This is the result of the previous round of correction. For the... The working images are downsampled and converted to the Lab color space, then clustered based on color consistency to obtain a material proxy category label map. This process can be abstractly represented as:

[0061]

[0062] in, For the first Downsampling results of the working image. For downsampling operators, for Color-based clustering functions, such as those implemented using KMeans clustering, can be used. Indicates the downsampling coordinates Category label at the location.

[0063] Step 3, Note For downsampling working images In coordinates The green channel value at that location, and taking its logarithmic brightness, yields:

[0064]

[0065] The logarithmic brightness is decomposed into a global vignetting term and a category bias term. The global vignetting term describes the smooth vignetting trend across the entire field of view, while the category bias term describes the average brightness shift within the same material proxy category. Let the first... The pixel set is Then for any have:

[0066]

[0067]

[0068] In the formula, These are pre-defined low-order two-dimensional polynomial basis functions used to describe the trend of the vignetting field smoothly changing with spatial position; These are the weight coefficients of the corresponding basis functions, and the terms together constitute the global halo term. ; For the first Lightning bias term for material proxy partition; Let be the number of terms in the polynomial. Here... Indicates the first The numbering of the basis functions does not indicate the order of the polynomial.

[0069] Considering category bias The constant brightness offsets used to describe various categories are used to avoid redundancy between the global constant term and the category bias term. In a preferred embodiment, the basis function set of the global vignetting term does not include a constant term. For example, a second-order two-dimensional polynomial basis function set can be used. , , , , The corresponding number of polynomial terms at this time is If constant terms are temporarily retained in some implementations, these degrees of freedom can be effectively eliminated in subsequent center normalization steps.

[0070] To suppress overfitting and ensure that the solution closely matches the observed data without being overly complex, the cost function can be split into three parts: a fitting error term, a global shading regularization term, and a class bias regularization term.

[0071]

[0072]

[0073]

[0074]

[0075] in, This represents the fitting error term, which means the error in the model output. As close as possible to the actual observed logarithmic brightness ; This represents the global shading regularization term, used to restrict the coefficients of each polynomial. Too large a size may cause unnecessary fluctuations in the global vignetting surface; This represents a categorical bias regularization term, used to limit the bias amount for each category. Too large a value, to avoid misassigning excessive brightness variations to material category differences.

[0076] By minimizing the total cost function Find each and .in, The regularization weights for the polynomial coefficients. The regularization weights for the class bias term strike a balance between "fitting the image brightness as closely as possible" and "avoiding an overly complex model".

[0077] Step 4: Obtain the polynomial coefficients Substituting back into the full-resolution coordinate grid, we obtain the original global halosing logarithmic field at full resolution. To eliminate the global constant degrees of freedom and to make the vignetting field reference the center of the field of view, a center normalization is performed. The normalization radius and the central and edge ring regions are defined as follows:

[0078]

[0079]

[0080] in, The threshold value for the center area radius. and The threshold value for the radius of the edge ring region. This represents median statistics. Preferably, the reasonableness of the vignetting direction can be determined by comparing the median of the central region mask with the median of the edge ring region mask. If the median of the edge ring region mask is higher than the median of the central region mask, it is determined that the current vignetting direction is inconsistent with the expected attenuation direction, the current stage is terminated, and the stable result of the previous round is retained. Otherwise, the incremental mask for this round is constructed, and the cumulative basic vignetting mask is updated.

[0081]

[0082] Step 5: Apply the accumulated base vignetting mask downsampling And estimate the category response index within each material proxy category. .set up , For the green channel of the original single-frame image after downsampling, then for each category... Establish the following fitting relationship:

[0083]

[0084]

[0085]

[0086] in, For the intercept term, For the residual term, For the first Logarithmic brightness in class Within-category mean The coefficients of determination for linear fitting within each category. Preferably, the effective number of pixels per category can be used. Variation variance and coefficient of determination within this category As a reliability metric for fitting. In a preferred embodiment, when the number of effective pixels in a certain category is less than 200 pixels, or The variance of variation within this category is less than or coefficient of determination When the value is less than 0.10, the exponential fit reliability within that category is deemed insufficient, and then... Otherwise, the estimated Limit it to a preset range. Then classify the categories Mapping the label map back to full resolution yields the category response map. .

[0087] Step 6: Since vignetting primarily reflects illumination attenuation rather than color attributes, the cumulative base vignetting mask estimated from the green channel is used. and category response graph This serves as a shared calibration basis for the three RGB channels. Combining these two methods constructs the first... An effective vignetting mask is applied, and based on this, differential correction is performed on the original color image to obtain the current corrected image. :

[0088]

[0089] in, To prevent extremely small positive numbers with a denominator of zero; Located at the power exponent position, it indicates that the same base mask is used. Instead of simply multiplying with the base mask, class-related response intensity adjustment is performed. The current corrected image is truncated to the [0,1] interval and used as the input working image for the next iteration.

[0090] Step 7: At the end of each iteration, calculate the root mean square index of the center-normalized halo logarithmic field to determine whether to continue iteration in the current stage. Preferably, continuing iteration in the current stage requires simultaneously meeting the following conditions: First, the root mean square index of the current round is not higher than that of the previous round; Second, the root mean square index of the current round is still greater than a preset stopping threshold. The root mean square index... It can be represented as:

[0091]

[0092] Preferably, a two-stage iterative strategy is employed to improve single-frame correction stability. The first stage allows for label maps... The image is re-segmented as the corrected image is updated (returning to step two in each iteration) to obtain material proxy partitions that match the current sample. When the first stage reaches the stopping condition, its final label map is denoted as... In the second stage, the label map is fixed, and only the global vignetting field and category response index are updated, i.e.:

[0093]

[0094] in, This marks the end of the first iteration round. This refers to the number of iterations in the second stage. Through this two-stage design, the adaptive capability of material partitioning in the early stage is preserved, while avoiding partition drift caused by repeated free segmentation in the later stage.

[0095] If the fourth step triggers an early stop due to an incorrect vignetting direction, or if the root mean square index of the current round is greater than that of the previous round, or if the root mean square index of the current round is not greater than the preset stopping threshold, the current stage is terminated and the most recent stable result is retained. After the second stage iteration ends, the effective vignetting mask corresponding to the last round is output. and corrected images This serves as the final single-frame vignetting correction result.

[0096] This embodiment also relates to a computer program product, including a computer program or instructions, which, when executed by a processor, implement the above-mentioned content.

[0097] The embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A material-aware single-frame vignetting correction method, characterized in that, Includes the following steps: (1) Acquire the single-frame color image to be corrected A normalized coordinate system is established for the image, and a certain color channel c1 is selected as the logarithmic domain fitting channel; where, The horizontal pixel coordinates are... Vertical pixel coordinates For color channel indexes; (2) Initialize the cumulative base vignetting mask Category response diagram and effective vignetting mask ; Record the first The working image of the wheel is ,when hour ,when hour This is the result of the previous round of correction; for the first... The working image of the wheel is downsampled and converted to the Lab color space, and then clustered based on color consistency to obtain a material proxy category label map: , in, For the first Downsampling results of the working image. For downsampling operators, for Color clustering function, To convert to the Lab color space, Indicates the downsampling coordinates Category label map at the location; (3) Record for In coordinates The color channel c1 value at that location is calculated, and its logarithm (luminance) is taken to obtain: , Logarithmic brightness It is decomposed into a global vignetting term and a category bias term. The global vignetting term describes the smooth vignetting trend across the entire field of view, and the category bias term describes the average brightness shift within the same material proxy category. Let the first... The pixel set is Then we have: , , in, For global shading, For the first Category bias of material proxy partition, These are predefined low-order two-dimensional polynomial basis functions used to describe the trend of the vignetting field smoothly changing with spatial position. For step (1), the horizontal and vertical normalized coordinates are obtained in the coordinate system. The corresponding form below, These are the weighting coefficients of the corresponding basis functions. Let be the number of terms in the polynomial. Indicates the first The basis functions are numbered, k∈[1,K]; Determine the fitting error term, the global shading regularization term, and the class bias regularization term, and construct the total cost function based on these three terms: , , , , in, Represents the fitting error term. This indicates a global shading regularization term. This indicates a categorical bias regularization term. Represents the total cost function. The regularization weights for the polynomial coefficients. Regularization weights for class bias terms; By minimizing the total cost function Find each and ; (4) The result obtained in step (3) Substituting back into the full-resolution coordinate grid, we obtain the original global vignetting logarithmic field at full resolution. Then, we perform center normalization on this original global vignetting logarithmic field to obtain the center-normalized vignetting logarithmic field. Then, construct the incremental mask for this round and update the cumulative base vignetting mask: , in, For the incremental mask in round t, Let be the cumulative base vignetting mask for round t, and min be the minimum operation. (5) Accumulate the base vignetting mask downsampling And estimate the category response index within each material proxy category. For each category Establish the following fitting relationship: , , , in, The c1 value is the color channel value after downsampling from the original single-frame color image. For the intercept term, For the residual term, for Within-category mean The coefficient of determination for linear fit within the category; Categories Mapping the label map back to full resolution yields the category response map for round t. ; (6) Based on cumulative basic vignetting mask and category response graph Construct the first Effective vignetting mask Based on this, differential correction is performed on the original single-frame color image to obtain the current corrected image. : , in, To prevent extremely small positive numbers with a denominator of zero; Current corrected image After being truncated to the [0,1] interval, it serves as the working image for the next iteration. ; (7) At the end of each iteration, calculate the center-normalized halo logarithmic field. The root mean square index is used to determine whether to continue iterating. If the iteration termination condition is met, the effective vignetting mask corresponding to the last round is set. and correct image This serves as the final single-frame vignetting correction result.

2. The material-aware single-frame vignetting correction method according to claim 1, characterized in that, In step (1), the operation of establishing a normalized coordinate system for the single-frame color image to be corrected is as follows: , in, and These represent the normalized coordinates of the horizontal and vertical axes, respectively, where W is the image width and H is the image height. The color channel c1 is the green channel.

3. The material-aware single-frame vignetting correction method according to claim 1, characterized in that, In step (3), the basis functions of the low-order two-dimensional polynomial are removed. The constant terms in the text.

4. The material-aware single-frame vignetting correction method according to claim 1, characterized in that, In step (4), ,in, med indicates the median operation. For the full-resolution original global vignetting logarithmic field, Indicates the central area. The threshold value for the center area radius. , and These are the normalized horizontal and vertical coordinates obtained in step (1), respectively.

5. The material-aware single-frame vignetting correction method according to claim 1, characterized in that, In step (5), the number of valid pixels for each category, Variation variance and coefficient of determination within this category As a reliability index for the fit, if at least one of these three indices fails to meet the corresponding requirements, the fit reliability for that category is deemed insufficient, and then... Otherwise, the estimated It is limited to a preset range.

6. The material-aware single-frame vignetting correction method according to claim 1, characterized in that, In step (7), the center-normalized halo logarithmic field The root mean square index is calculated as follows: , in, is the root mean square index, W is the image width, and H is the image height.

7. The material-aware single-frame vignetting correction method according to claim 1, characterized in that, In step (7), if the root mean square index of the current round is not higher than the root mean square value of the previous round, and the root mean square index of the current round is greater than the preset stop threshold, the iteration continues; otherwise, the iteration stops.

8. The material-aware single-frame vignetting correction method according to claim 1, characterized in that, The entire iterative process is divided into two phases. The first phase of iteration ends when the iteration stopping condition is met for the first time. At this point, the category label map obtained in the current round is... Once the final category label map no longer changes, the second phase of iteration begins until the iteration stopping condition is met again, at which point the entire iteration process ends.

9. The material-aware single-frame vignetting correction method according to claim 1, characterized in that, In step (4), the median of the central region mask is compared with the median of the edge ring region mask to determine whether the vignetting direction is reasonable. If the median of the edge ring region mask is higher than the median of the central region mask, it means that the current vignetting direction is inconsistent with the expected attenuation direction, that is, the current vignetting direction is unreasonable. The current stage is terminated and the result of the previous iteration is used as the correction result of the current stage.

10. A computer program product, comprising a computer program or instructions, characterized in that: When the computer program or instructions are executed by the processor, they implement the material-aware single-frame vignetting correction method as described in any one of claims 1-9.

Images