A method, system, and medium for non-uniformity correction of large target area image sensors

By processing multiple frames of images from a large-area CMOS image sensor under non-uniform illumination conditions, separating optical vignetting and columnar noise, and combining this with seam neighborhood correction, the image quality and measurement accuracy problems of large-area sensors under non-uniform illumination are solved, achieving high-precision image correction.

CN123075293APending Publication Date: 2026-11-10HEFEI I TEK OPTOELECTRONICS CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611592614.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-15
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Large-area CMOS image sensors suffer from optical vignetting, aliasing of fixed-pattern noise and pixel response non-uniformity under non-uniform illumination conditions. Traditional correction methods cannot effectively distinguish and correct these issues, which affects image quality and measurement accuracy.

Method used

By acquiring multiple frames of images under non-uniform illumination, the column gain coefficient and sub-region gain coefficient of the readout area are calculated. Optical vignetting and column noise are separated by smoothing filtering and polynomial fitting. Combined with seam neighborhood correction, the three components are decoupled and accurately corrected at the seam.

Benefits of technology

Without increasing storage capacity, it effectively eliminates column stripes and stitching artifacts while preserving optical vignetting, achieving high-precision image correction and making it suitable for large-area sensors under strong vignetting conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123075293A_ABST
    Figure CN123075293A_ABST
Patent Text Reader

Abstract

This invention discloses a method, system, and medium for correcting non-uniformity in a large-area image sensor. The correction method includes: calculating the average brightness of each column of the image corresponding to all readout areas; analyzing and obtaining the average brightness of each sub-region; fitting the average brightness of all sub-regions to form a corresponding smooth surface, thereby obtaining the sub-region gain coefficient; acquiring the image data to be corrected, and determining whether each pixel is located in a sub-region adjacent to the seam: if so, the sub-region gain coefficient of the corresponding sub-region is used as the correction coefficient; if not, interpolation is performed based on the sub-region gain coefficients of adjacent sub-regions to obtain the correction coefficient; outputting the image data to be corrected after sequentially passing the corresponding column gain coefficient and the correction coefficient. This invention can simultaneously eliminate column stripes, block-level non-uniformity, and seam artifacts under strong vignetting conditions, while preserving true optical vignetting.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of machine vision, and particularly relates to a method, system and medium for non-uniformity correction of large target surface image sensors. Background Technology

[0002] With the increasing demands for imaging resolution and image size in industrial inspection, large-area CMOS image sensors are being applied to more and more scenarios. To achieve ultra-large arrays while maintaining readout speed, large-area CMOS image sensors generally adopt a multi-readout area stitching architecture: the pixel array is divided into several independent readout areas, each with its own independent readout circuitry (column readout channels, analog front-end, ADC group), and the physical boundary between adjacent readout areas is called the seam. Such sensors exhibit various sources of non-uniformity during imaging, which can be categorized into four main types: First, Column Fixed Pattern Noise (VFPN): Gain differences between the readout circuits of each column, manifesting as high-frequency stripes in the column direction. In practice, the shape remains unchanged, and the amplitude varies linearly with brightness within the range of 75–3077 DN. Second, Pixel Response Non-uniformity (PRNU): Differences in photoelectric conversion efficiency between pixels and pixel blocks, manifesting as block-level gain non-uniformity. Locally, in the vicinity of the seam, this manifests as response differences between odd and even columns. In practice, this difference can reach ±2%–4% at the image edges, and varies with row position and signal level. Third, Seam Steps: Average brightness steps on both sides of the seam (measured at 0.4%–0.7% for the lower half of the vertical seam and 0.1%–0.9% for the right half of the horizontal seam). Fourth, Optical Vignetting: Smooth two-dimensional brightness attenuation caused by the mismatch between the lens optical system and the principal ray angle (CRA) of the sensor microlens array. This is an inherent property of the optical system and must be preserved during calibration. The aforementioned non-uniformities overlap in actual images, severely affecting image quality and subsequent measurement accuracy.

[0003] Traditional non-uniformity correction has the following two problems: Problem 1: Physically uniform flat-field images cannot be obtained under non-uniform illumination. The parameter calculation method is as follows: acquire multiple frames of flat-field images under a uniform light source such as an integrating sphere, divide each frame by its own frame level for normalization, and then sum and average to directly obtain the correction coefficient field. This method relies on the following premise: the incident light illumination on the sensor target surface is physically uniform, that is, the intra-frame variation of the flat-field image is entirely caused by the non-uniformity of the sensor itself. However, for large-scale sensors, this assumption does not hold. Since the principal ray angle (CRA) changes continuously with the off-axis distance of the pixel, even with a uniform light source such as an integrating sphere, the actual illuminance received by the sensor still exhibits significant radial attenuation, and the ratio of corner to center illuminance can be as low as 0.485. The consequence of this is that optical vignetting and the sensor PRNU and VFPN are completely aliased in the flat-field image, making it impossible to distinguish their respective contributions, resulting in the correction coefficients incorrectly including the optical components that should have been retained.

[0004] Question 2: Block coefficient interpolation has a fundamental flaw at the seam. The conventional method for pixel-level correction is to store a gain coefficient for each pixel. For megapixel sensors, the storage size for pixel-level coefficients reaches hundreds of megabytes, making on-chip implementation impossible. Therefore, the commonly used architecture in engineering is block coefficient interpolation: the pixel array is divided into regular blocks (e.g., 64×64 pixels), and a gain coefficient is stored for each block; each block is assigned a representative position (usually the block center); for any pixel, bilinear interpolation is performed between the centers of adjacent blocks using its position to obtain the pixel's gain coefficient, which is then multiplied by the pixel value. This architecture achieves a smooth correction effect with spatial continuity and no blocky traces with minimal coefficient storage overhead, and is the mainstream practice in industrial cameras and consumer electronics ISPs. However, when the correction target is a seam within the sensor, this architecture has a fundamental flaw. The sensor response difference on both sides of the seam is a physical abrupt change located in a single column (or row), and the ideal correction amount should also be a gain step located in the same column. However, bilinear interpolation is designed to "smooth out" the gap between adjacent blocks. Its mathematical essence is a linear transition that unfolds over the block center distance (e.g., 64 pixels), which cannot correct for abrupt unevenness.

[0005] Therefore, this invention provides a method, system, and medium for non-uniformity correction of large-area image sensors. On the one hand, it aims to decouple the three components of optical vignetting, column-oriented fixed pattern noise, and pixel response non-uniformity from the original image sequence under conditions where physical uniform illumination is not possible. On the other hand, it aims to avoid the thinning and widening of the stitching correction step by interpolation operations in the real-time correction architecture using block coefficients and interpolation, thereby achieving residual correction of stitching artifacts without increasing the amount of coefficient storage. Summary of the Invention

[0006] The purpose of this invention is to overcome the above-mentioned problems in the prior art and to provide a method, system and medium for non-uniformity correction of large target surface image sensors.

[0007] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution: A method for non-uniformity correction of a large target surface image sensor includes: Acquire several frames of original images under non-uniform lighting, and after superimposing and averaging, calculate the average brightness of each column of the corresponding image for all readout areas; The ratio of the smooth curve formed by smoothing the average brightness of each column in each readout area to the average brightness of each column is calculated and used as the column gain coefficient. The image after column gain coefficient compensation is then divided into several sub-regions to obtain the average brightness of each sub-region. A two-dimensional fit is performed on the average brightness of all sub-regions based on a polynomial whose order does not exceed a preset value to obtain a smooth surface. The ratio of the smooth surface to the average brightness of the sub-region is used as the sub-region gain coefficient. Acquire the image data to be corrected and determine whether each pixel is located in a sub-region adjacent to the stitching seam: if so, use the sub-region gain coefficient of the corresponding sub-region as the correction coefficient; if not, interpolate based on the sub-region gain coefficient of the adjacent sub-region to obtain the correction coefficient. The output image data to be corrected is then processed sequentially by the corresponding column gain coefficients and correction coefficients to obtain the corrected image data. The smooth curves between different readout areas are independent of each other, and the seam is located between two adjacent readout areas. When a sub-region is traversed by the seam, the sub-region is split into two subordinate sub-regions along the seam, and the gain coefficient of each sub-region is determined.

[0008] Furthermore, before averaging, the mean or median brightness of all original images is calculated, and original images whose mean or median brightness deviates from the set threshold range are removed. Each image is then normalized by dividing it by its own mean or median brightness.

[0009] Furthermore, the sub-region gain coefficient is either the sub-region odd column gain coefficient or the sub-region even column gain coefficient. The sub-region odd column gain coefficient is the ratio of the smooth surface to the average brightness of the odd column of the sub-region, and is used for the correction of the odd column pixels within the sub-region. The sub-region even column gain coefficient is the ratio of the smooth surface to the average brightness of the even column of the sub-region, and is used for the correction of the even column pixels within the sub-region.

[0010] Furthermore, the interpolation is performed by bilinear interpolation on the odd-column gain coefficients and even-column gain coefficients of the sub-regions, respectively.

[0011] Furthermore, the two-dimensional fitting is a two-dimensional polynomial least squares fitting with a total order not exceeding 8, and the row and column indices of the sub-regions are normalized to [-1, 1].

[0012] Furthermore, the column gain coefficient ranges from 0.9 to 1.1.

[0013] Furthermore, the output image data is cropped to an upper limit, which is either the upper limit of the output bit width or the value corresponding to the preset maximum gain.

[0014] Furthermore, it also includes: multiplying the column gain coefficient and the sub-region gain coefficient by the same constant so that the resulting values ​​are not less than 1, and encoding them as integers according to a preset rounding rule for storage.

[0015] The present invention also provides a non-uniformity correction system for a large target surface image sensor, comprising: The initial acquisition module is used to acquire several frames of original images under non-uniform lighting, and after superimposing and averaging, calculate the average brightness of each column of the corresponding image of all readout areas; The smoothing analysis module is used to calculate the ratio of the smooth curve formed by the average brightness of each column in each readout area after smoothing filtering to the average brightness of each column, which is used as the column gain coefficient. The image after column gain coefficient compensation is divided into several sub-regions to obtain the average brightness of each sub-region. The smooth curves between different readout areas are independent of each other. The sub-region analysis module is used to perform two-dimensional fitting of the average brightness of all sub-regions based on a polynomial with an order not exceeding a preset value, to obtain a smooth surface. The ratio of the smooth surface to the average brightness of the sub-region is used as the sub-region gain coefficient. The edge analysis module is used to acquire the image data to be corrected and determine whether each pixel is located in a sub-region adjacent to the stitching seam. If so, the sub-region gain coefficient of the corresponding sub-region is used as the correction coefficient. If not, the correction coefficient is obtained by interpolation based on the sub-region gain coefficients of the adjacent sub-regions. The stitching seam is located between two adjacent readout areas. When a sub-region is crossed by the stitching seam, the sub-region is split into two subordinate sub-regions along the stitching seam, and the gain coefficient of each sub-region is determined. The correction output module is used to output the image data after the image data to be corrected has been successively corrected by the corresponding column gain coefficient and correction coefficient.

[0016] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described correction method.

[0017] The beneficial effects of this invention are: This invention addresses the problems of optical vignetting under non-uniform illumination, the overlapping of column-oriented fixed pattern noise and pixel response non-uniformity, and the thinning of correction steps at the seams by block interpolation. It achieves joint decoupling of three components. By independently smoothing column gain in the readout area, the seam step is preserved; by replicating seam neighborhood coefficients and interpolating internal regions, the correction step is precisely aligned with the physical seam. Thus, column fringes, block-level non-uniformity, and seam artifacts can be eliminated simultaneously under strong vignetting conditions, preserving true optical vignetting without increasing coefficient storage, resulting in low hardware costs, and supporting adaptive ROI and binning. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart of the correction method in this invention; Figure 2This is a block diagram of the correction system structure in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Existing technologies for non-uniformity correction suffer from two interconnected shortcomings: first, under non-uniform illumination conditions, it is impossible to obtain the correct block correction coefficients through conventional flat-field calibration; second, even if the correct block coefficients are obtained, conventional block interpolation correction architectures cannot correctly apply these coefficients at the seams. Both of these factors contribute to the difficulty in eliminating seam and block-level non-uniformity in large-area spliced ​​sensors under strong vignetting conditions.

[0021] To solve the above problems, such as Figure 1 As shown, this embodiment first provides a method for correcting the non-uniformity of a large target surface image sensor, including: Step 1: Acquire several frames of original images under non-uniform lighting, and after superimposing and averaging, calculate the average brightness of each column of the corresponding image for all readout areas.

[0022] Multiple frames of raw images with different exposure times were acquired under non-uniform illumination conditions, with brightness values ​​ranging from approximately 100 DN to 3200 DN. The average brightness values ​​at the same pixel location in all raw images were calculated and summed to obtain a pre-processed image after overlay averaging. At this point, the average brightness value of each column was calculated separately for each readout area. Overlay averaging across multiple frames suppresses random noise and temporal fluctuations. Since physical uniform illumination is not required, calibration can be performed directly under actual optical systems and lighting conditions. Calculating column averages separately for each readout area ensures that column-wise statistics for each readout area do not contaminate each other, especially avoiding data mixing on both sides of the seam. The column brightness average simultaneously includes high-frequency fixed pattern noise and low-frequency optical vignetting components in the column direction, providing a basis for subsequent separation. This step eliminates the dependence on a uniform light source, making it suitable for sensors with strong vignetting and large target surfaces; it reduces the impact of random noise on coefficient calculations; and it preserves the independent characteristics of each readout area, avoiding artifacts caused by cross-seam statistics.

[0023] In a specific embodiment of the present invention, before superimposing and averaging, the mean or median brightness of all original images is calculated. Original images whose mean or median brightness deviates from a set threshold range are removed, and each image is normalized by dividing it by its own mean or median brightness. Removing original images whose mean or median brightness deviates from the set threshold range means removing original image data that is below the lower limit or above the upper limit, retaining only original image data that falls within the preset range.

[0024] Step 2: Calculate the ratio of the smooth curve formed by smoothing the average brightness of each column in each readout area to the average brightness of each column, and use it as the column gain coefficient. Then, divide the image after column gain coefficient compensation into several sub-regions and obtain the average brightness of each sub-region. The smooth curves between different readout areas are independent of each other.

[0025] This step utilizes the frequency difference between column-directed fixed pattern noise (high-frequency signal) and optical vignetting (low-frequency signal) to smooth the contours of the average brightness of each column, extracting the low-frequency component of the optical vignetting in the column direction. The column gain coefficient is then determined by the ratio of this low-frequency component to the average brightness of each column. This smoothing filter is performed side-by-side with the vertical seam as the boundary: if overall smoothing across the seam is used, the seam step will be "sloped" into a gradually changing slope with a width equal to the width of the smoothing window, thus mismatching with the subsequent block-level correction applied by readout area isolation, resulting in systematic undercorrection. Side-by-side smoothing ensures that the seam step passes through this step intact, allowing it to be processed by subsequent block-level correction. This embodiment takes four readout areas, i.e., four quadrants, as an example. The column brightness mean profile P(x) is calculated for the upper half (rows 0-4767) and the lower half (rows 4768-9535), where x=0,1,...,13439. Using the vertical seam column (x=6720) as a hard boundary, a 128-pixel moving average is applied to P(x) on both the left and right sides to obtain the smooth curve T(x) for each readout area. Then, the column gain coefficient k(x) is calculated column by column, where k(x)=clip(T(x) / P(x),0.9,1.1), indicating that the result of T(x) / P(x) is limited to between 0.9 and 1.1. This definition performs a low-pass segmentation of the one-dimensional spectrum of the column mean at 1 / 128 (period / pixel): the high-frequency component (column-directed fixed pattern noise) is removed, while the low-frequency component (including the column-directed components of optical vignetting and the seam steps themselves) is retained. Its magnitude does not participate in the final vignetting localization; it is only used to remove high-frequency column-directed fringes. Side-by-side smoothing ensures that the seam steps are not sloped. Actual measurements show that this step introduces only approximately +0.017% vignetting at the edges and approximately 0.008% RMS across the full width, which is negligible.

[0026] After analyzing the column gain coefficients of each readout region, compensation is first applied using these coefficients. The compensated image is then divided into several sub-regions, which are rectangular array regions. Low-frequency smoothing curves are extracted using smoothing filtering. The ratio of the low-frequency target to the original column mean is then used as the column gain coefficient. This removes only high-frequency column fringes while preserving low-frequency optical vignetting and seam steps. Side-by-side smoothing is performed using the seam as the boundary, extracting the column-oriented low-frequency vignetting contour within each readout region as the smoothing curve. Seam steps, as transitions between readout regions, are preserved and not included in this elimination step.

[0027] Step 3: Perform a two-dimensional fit on the average brightness of all sub-regions based on a polynomial with an order not exceeding a preset value to obtain a smooth surface. Use the ratio of the smooth surface to the average brightness of the sub-region as the sub-region gain coefficient.

[0028] Calculate the average brightness of each sub-region and fit a two-dimensional low-order surface (in this embodiment, a polynomial with a total order of no more than 8). The fitted surface is smooth and cannot express block noise and seam steps, so it approximately represents a pure optical vignetting surface. Once the average brightness of all sub-regions is known, a smooth surface can be fitted.

[0029] As a specific embodiment of the present invention, the two-dimensional fitting is a two-dimensional polynomial least squares fitting with a total order not exceeding 8, and the row and column indices of the sub-regions are normalized to [-1, 1]. The fitting formula is:

[0030] Where F(u,v) is the fitted smooth surface, u and v are the normalized x and y coordinates, and a ij Let i be the polynomial coefficients, i be the power of u, j be the power of v, and i+j≤8 means that the total order of the polynomial does not exceed 8.

[0031] To further address the image odd-even column defect, as a specific embodiment of the present invention, the sub-region gain coefficient is either the sub-region odd-column gain coefficient or the sub-region even-column gain coefficient. The sub-region odd-column gain coefficient is the ratio of the smooth surface to the average brightness of the odd-column pixels in the sub-region, used for the correction of odd-column pixels within the sub-region. The sub-region even-column gain coefficient is the ratio of the smooth surface to the average brightness of the even-column pixels in the sub-region, used for the correction of even-column pixels within the sub-region. The average brightness of the odd-column pixels and the average brightness of the even-column pixels in each sub-region are statistically analyzed to obtain the odd-column block statistical field and the even-column block statistical field. Taking a sub-region size of 64 pixels as an example, the image is divided into 149×210 sub-regions. For each sub-region, the average brightness of the odd-column pixels (column numbers 1, 3, 5, ...) and the even-column pixels (column numbers 0, 2, 4, ...) are calculated to obtain the odd-column block statistical field and the even-column block statistical field, both with a size of 149×210. The middle block row (rows 4736–4799) located at the horizontal seam is split into an upper half (rows 4736–4767) and a lower half (rows 4769–4799, excluding defective rows) for separate statistical analysis. The fitted smooth surface, i.e., the optical vignetting surface, is used as the common correction target for odd and even columns. The gain coefficients for odd and even columns in the sub-regions are calculated separately to ensure that the average brightness values ​​of odd and even columns, after correction by the odd and even column gain coefficients, are consistent with the corresponding smooth surface. A pure optical vignetting surface can still be extracted under non-uniform illumination; block-level correction of pixel response non-uniformity is achieved; the difference between odd and even responses is fully modeled, and artifacts caused by odd-even transitions in the vertical seam can be eliminated along with block-level non-uniformity; no additional brightness or shadow distortion is introduced. Each sub-region contains only the gain coefficients of its corresponding readout area, and there are no shared blocks between sub-regions. For cameras with local odd-even column non-uniformity, each sub-region is further divided into two sets: odd-column sub-regions and even-column sub-regions. By dividing the image into blocks and calculating the odd-column and even-column mean values ​​respectively, block-level differences in pixel response non-uniformity and odd-even column response differences in the stitching neighborhood can be captured.

[0032] As a specific embodiment of the present invention, it further includes: multiplying the column gain coefficient and the sub-region gain coefficient by the same constant to ensure that the resulting values ​​are all not less than 1, and encoding them as integers according to a preset rounding rule for storage. All the above coefficients are normalized according to a global scale (ensuring that all gains are not less than 1 after encoding), encoded as a fixed-point integer array, divided into multiple subfields according to the readout area, and then written into the camera's non-volatile memory.

[0033] For example, the column gain coefficient is denoted as k, and the sub-region gain coefficient (either the odd-column gain coefficient or the even-column gain coefficient) is denoted as c. The encoded value of the column gain coefficient is round(k × global scale × 2048), cropped to [0, 4095]; the encoded value of the sub-region gain coefficient (either the odd-column gain coefficient or the even-column gain coefficient) is round((c × global scale - 1) × 2048), cropped to [0, 8191]. "round" represents rounding down. The global scale is the constant mentioned above. Since the coefficient array is based on gain 1 and the global scale does not change the spatial distribution of the coefficients, this normalization will not introduce new stitching steps, only slightly increasing the overall image brightness, which can be compensated for by exposure.

[0034] The encoded coefficients are divided into multiple subfields according to the readout area (4 quadrants × 2 sets of odd and even = 8 subfields). After correcting the correspondence between the upper and lower halves according to the FPGA hardware routing characteristics, they are written to the camera Flash through a high-speed interface. After writing is completed, the parameter submission and activation commands are triggered.

[0035] Step 4: Obtain the image data to be corrected and determine whether each pixel is located in a sub-region adjacent to the seam. If yes, use the sub-region gain coefficient of the corresponding sub-region as the correction coefficient. If no, interpolate based on the sub-region gain coefficient of the adjacent sub-region to obtain the correction coefficient. The seam is located between two adjacent readout areas. When a sub-region is crossed by the seam, the sub-region is split into two subordinate sub-regions along the seam, and the sub-region gain coefficient of each is determined.

[0036] For real-time input image data to be corrected, it is necessary to determine the readout area, sub-region, and position of each pixel within its respective sub-region in real time. Different coefficient determination rules are used based on the region where the pixel is located: Internal region (conventional interpolation): When a pixel is located inside the readout area, the sub-region gain coefficients of multiple adjacent sub-regions (preferably 2×2) are taken, and bilinear interpolation is performed according to the pixel position to obtain the correction coefficient of the pixel. Seam Neighborhood (Coefficient Copying, Interpolation Disabled): When a pixel is located in a sub-region adjacent to the seam, its correction coefficient is directly taken as the sub-region gain coefficient of the sub-region closest to the seam. Specifically, as shown above, this embodiment divides the image into 149×210 sub-regions, with the vertical seam located between the 105th and 106th column sub-regions. For pixels to the left of the vertical seam (left quadrant sub-field): when the pixel is located between the center of the 105th sub-region and the seam, its correction coefficient is taken as the sub-region gain coefficient of the 105th sub-region and maintained up to the seam; for pixels to the right of the vertical seam (right quadrant sub-field): when the pixel is located between the seam and the center of the 106th sub-region, its correction coefficient is taken as the sub-region gain coefficient of the 106th sub-region and maintained up to the seam; the coefficients on the left and right sides are not mixed, and interpolation operations are not performed across the stitching boundary. Thus, the correction coefficient field forms a gain step between columns 6719 and 6720, with an amplitude equal to the difference between the gain coefficients of the 105th and 106th sub-regions. This step corresponds in position and amplitude to the abrupt change in the sensor's response at the vertical seam, thereby achieving residue-free correction at the seam. The horizontal seam passes through the center of the 75th row sub-region, which can be divided into two half-rows: the upper half (row 32) and the lower half (row 32), and each half is assigned an independent coefficient, i.e., divided into two subordinate sub-regions. When calculating the sub-region gain coefficient of one of the subordinate sub-regions, the ratio of the smooth surface to the average brightness of the current subordinate sub-region is used as the sub-region gain coefficient of that subordinate sub-region.

[0037] If bilinear interpolation is still performed in the seam neighborhood, the step will be thinned and widened. Therefore, edge sub-region corresponding coefficients are used in the seam neighborhood to make the correction coefficients form a step aligned with the physical seam from one side of the seam to the other; bilinear interpolation is retained in the readout region to ensure smoothness without blocky traces. Selecting the odd column gain coefficient or even column gain coefficient of the sub-region according to the odd and even columns can simultaneously correct for odd and even response differences.

[0038] Step 5: Output the image data to be corrected after sequentially passing through the corresponding column gain coefficients and correction coefficients.

[0039] Each pixel value of the image data to be corrected is first multiplied by a column gain coefficient obtained by looking up a table of column coefficients (the column gain coefficient is obtained directly from the column coefficient table according to the column number of the pixel, without interpolation in the column direction), and then multiplied by the sub-region gain coefficient determined above, and the corrected image data is output. In specific implementation, since the corrected pixel values ​​may exceed the sensor bit width or output range, in order to prevent overflow, the output image data is cropped at an upper limit, which is the upper limit of the output bit width or the preset maximum gain value.

[0040] When the readout window (ROI), window position, or binning mode changes, the position of the stitching seam relative to the readout window changes accordingly. At this time, the sub-region index where the stitching seam is located is recalculated in real time based on the window offset and updated to the configuration register of the interpolation unit, so that the switching boundary between copying and interpolation is always aligned with the physical stitching seam.

[0041] In actual testing, under strong vignetting conditions with a 13440×9536 sensor and a corner / center illuminance ratio of approximately 0.485, the above method resulted in all indicators falling within the range of 0.01%–0.13%. Specifically, the parity difference on the left and right edges converged from ±1.5%–2.5% to within ±0.04%. Since the main component of the vertical seam step is the parity jump, the corrected residual decreased to below 0.03%. Two invariants were also verified: the mean value of the entire image changed by no more than 0.006% before and after correction, and the change in optical vignetting shape (corner / center ratio) did not exceed 6×10⁻⁶. -5 This means that the correction method eliminates seams and block-level non-uniformity without introducing additional brightness or shadow distortion.

[0042] Conventional flat-field calibration methods acquire images under uniform light sources and directly normalize them to obtain correction coefficients. This relies on the premise that "intra-frame variations in the flat-field image are entirely caused by sensor non-uniformity." This invention does not depend on this premise. Instead, it acknowledges that the illumination and optical systems themselves are non-uniform, and decouples the three types of physical components through differences in spatial frequency characteristics. Therefore, it is suitable for large-area target sensors where CRAs vary significantly with off-axis distance and cannot achieve physically uniform illumination.

[0043] The present invention achieves the following effects through the above-described correction method: First, the three components are jointly decoupled: optical vignetting, column-oriented fixed pattern noise and pixel response non-uniformity are jointly separated by spatial frequency feature differences, which avoids the problem of mutual contamination of components in traditional methods and allows optical vignetting to be preserved as is.

[0044] Second, the parity difference is fully modeled: the same optical vignetting surface is used as the common target for both the odd and even columns, so that the parity response difference and the vertical stitching artifacts of the parity jump form can be eliminated by the same set of coefficients.

[0045] Third, the correction step amplitude at the seam is complete: near the seam, bilinear interpolation is no longer performed on the correction coefficients of adjacent blocks, so that the correction amount reaches the complete amplitude at the seam, avoiding the approximately 50% undercorrection caused by bilinear interpolation.

[0046] Fourth, eliminate the gradual artifacts caused by interpolation tails: the correction step is confined to the neighborhood of the stitching seam and will not spread to both sides as a ramp tens of pixels wide, thus avoiding the wide-scale gradual bright bands that are sensitive to the human eye.

[0047] Fifth, it does not increase storage and has extremely low hardware costs: it still uses block coefficient storage; the stitching process only requires adding a block index comparison and a multiplexer in the interpolation unit, without additional multipliers or storage.

[0048] Sixth, adaptive ROI and binning: The seam block index is dynamically updated with window parameters, and it can work correctly under any readout window.

[0049] Seventh, brightness and shadow invariants: the change in the mean of the entire image before and after correction is no greater than 0.006%, and the change in the shape of the optical vignetting is no greater than 6×10. -5 Correction does not introduce additional brightness or shadow distortion.

[0050] The present invention also provides a non-uniformity correction system for a large target surface image sensor, comprising: The initial acquisition module is used to acquire several frames of original images under non-uniform lighting conditions, and after superimposing and averaging, calculates the average brightness of each column of the corresponding image of all readout areas.

[0051] The smoothing analysis module is used to calculate the ratio of the smooth curve formed by the average brightness of each column in each readout area after smoothing filtering to the average brightness of each column, which is used as the column gain coefficient. The image after column gain coefficient compensation is divided into several sub-regions to obtain the average brightness of each sub-region. The smooth curves between different readout areas are independent of each other.

[0052] The sub-region analysis module is used to perform two-dimensional fitting of the average brightness of all sub-regions based on a polynomial of order not exceeding a preset value, to obtain a smooth surface. The ratio of the smooth surface to the average brightness of the sub-region is used as the sub-region gain coefficient.

[0053] The edge analysis module is used to acquire the image data to be corrected and determine whether each pixel is located in a sub-region adjacent to the stitching seam. If so, the sub-region gain coefficient of the corresponding sub-region is used as the correction coefficient. If not, the correction coefficient is obtained by interpolation based on the sub-region gain coefficients of the adjacent sub-regions. The stitching seam is located between two adjacent readout areas. When a sub-region is crossed by the stitching seam, the sub-region is split into two subordinate sub-regions along the stitching seam, and the gain coefficient of each sub-region is determined.

[0054] The correction output module is used to output the image data after the image data to be corrected has been successively corrected by the corresponding column gain coefficient and correction coefficient.

[0055] The specific operation method of the calibration system of the present invention can be referred to the above calibration method, and will not be repeated here.

[0056] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described correction method.

[0057] In practical applications, a computer-readable storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this embodiment, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0058] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0059] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, or any suitable combination thereof.

[0060] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via an Internet service provider).

[0061] In the description of this specification, references to terms such as "an embodiment," "example," and "specific example" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0062] 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 illustrative of the 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.

Claims

1. A method for correcting non-uniformity in a large target surface image sensor, characterized in that, include: Acquire several frames of original images under non-uniform lighting, and after superimposing and averaging, calculate the average brightness of each column of the corresponding image for all readout areas; The ratio of the smooth curve formed by smoothing the average brightness of each column in each readout area to the average brightness of each column is calculated and used as the column gain coefficient. The image after column gain coefficient compensation is then divided into several sub-regions to obtain the average brightness of each sub-region. A two-dimensional fit is performed on the average brightness of all sub-regions based on a polynomial whose order does not exceed a preset value to obtain a smooth surface. The ratio of the smooth surface to the average brightness of the sub-region is used as the sub-region gain coefficient. Acquire the image data to be corrected and determine whether each pixel is located in a sub-region adjacent to the stitching seam: if so, use the sub-region gain coefficient of the corresponding sub-region as the correction coefficient; if not, interpolate based on the sub-region gain coefficient of the adjacent sub-region to obtain the correction coefficient. The output image data to be corrected is then processed sequentially by the corresponding column gain coefficients and correction coefficients to obtain the corrected image data. The smooth curves between different readout areas are independent of each other, and the seam is located between two adjacent readout areas. When a sub-region is traversed by the seam, the sub-region is split into two subordinate sub-regions along the seam, and the gain coefficient of each sub-region is determined.

2. The method for non-uniformity correction of a large target surface image sensor according to claim 1, characterized in that, Before averaging, calculate the mean or median brightness of all original images, remove original images whose mean or median brightness deviates from the set threshold range, and normalize them by dividing them by their own mean or median brightness.

3. The non-uniformity correction method for a large target surface image sensor according to claim 1, characterized in that, The sub-region gain coefficient is either the sub-region odd column gain coefficient or the sub-region even column gain coefficient. The sub-region odd column gain coefficient is the ratio of the smooth surface to the average brightness of the odd column of the sub-region, and is used for the correction of the odd column pixels within the sub-region. The sub-region even column gain coefficient is the ratio of the smooth surface to the average brightness of the even column of the sub-region, and is used for the correction of the even column pixels within the sub-region.

4. The non-uniformity correction method for a large target surface image sensor according to claim 3, characterized in that, The interpolation is performed by bilinear interpolation on the odd-column gain coefficients and even-column gain coefficients of the sub-region, respectively.

5. The non-uniformity correction method for a large target surface image sensor according to claim 1, characterized in that, The two-dimensional fitting is a two-dimensional polynomial least squares fitting with a total order not exceeding 8, and the row and column indices of the sub-regions are normalized to [-1, 1].

6. The method for non-uniformity correction of a large target surface image sensor according to claim 1, characterized in that, The column gain coefficient ranges from 0.9 to 1.

1.

7. The non-uniformity correction method for a large target surface image sensor according to claim 1, characterized in that, The output image data is cropped to an upper limit, which is either the upper limit of the output bit width or the value corresponding to the preset maximum gain.

8. A method for non-uniformity correction of a large target surface image sensor according to any one of claims 1-7, characterized in that, Also includes: The column gain coefficient and the sub-region gain coefficient are multiplied by the same constant to ensure that the resulting values ​​are not less than 1, and then encoded as integers according to the preset rounding rules for storage.

9. A non-uniformity correction system for a large target surface image sensor, characterized in that, include: The initial acquisition module is used to acquire several frames of original images under non-uniform lighting, and after superimposing and averaging, calculate the average brightness of each column of the corresponding image of all readout areas; The smoothing analysis module is used to calculate the ratio of the smooth curve formed by the average brightness of each column in each readout area after smoothing filtering to the average brightness of each column, which is used as the column gain coefficient. The image after column gain coefficient compensation is divided into several sub-regions to obtain the average brightness of each sub-region. The smooth curves between different readout areas are independent of each other. The sub-region analysis module is used to perform two-dimensional fitting of the average brightness of all sub-regions based on a polynomial with an order not exceeding a preset value, to obtain a smooth surface. The ratio of the smooth surface to the average brightness of the sub-region is used as the sub-region gain coefficient. The edge analysis module is used to acquire the image data to be corrected and determine whether each pixel is located in a sub-region adjacent to the stitching seam. If so, the sub-region gain coefficient of the corresponding sub-region is used as the correction coefficient. If not, the correction coefficient is obtained by interpolation based on the sub-region gain coefficients of the adjacent sub-regions. The stitching seam is located between two adjacent readout areas. When a sub-region is crossed by the stitching seam, the sub-region is split into two subordinate sub-regions along the stitching seam, and the gain coefficient of each sub-region is determined. The correction output module is used to output the image data after the image data to be corrected has been successively corrected by the corresponding column gain coefficient and correction coefficient.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the correction method as described in any one of claims 1-8.