Infrared pot cover shape non-uniformity correction system based on scene gradient processing and correction algorithm thereof

Abstract

The invention relates to an infrared pot cover shape non-uniformity correction system based on scene gradient processing and a correction algorithm thereof. The system comprises an image input stream module, an initialization module, a down-sampling module, a median filtering module, a horizontal and vertical gradient processing module, an image reconstruction module, a bilinear interpolation module, an up-sampling module, an algorithm processing updating matrix module and an image output stream module which are connected in sequence. The median filtering module, the horizontal and vertical gradient processing module and the image reconstruction module are respectively connected with the RAM through a cache; the bilinear interpolation module and the up-sampling module are respectively connected with the FIFO through a cache, and the algorithm processing updating matrix module is connected with the DDR through a cache. The method has the advantages that the pot cover effect of an infrared image can be directly improved, system-level optimization of the infrared imaging quality is achieved, the method is adaptive to various types of infrared detectors after an input data protocol is configured, codes do not need to be changed one by one, and the method is low in hardware scheme resource consumption, high in convergence speed and high in transportability and universality.

Description

Technical Field

[0001] This invention belongs to the field of infrared image non-uniformity correction algorithms, and particularly relates to an infrared pot lid-shaped non-uniformity correction system and its correction algorithm based on scene gradient processing. Background Technology

[0002] Non-uniformity in infrared images refers to the phenomenon that different pixels in an infrared focal plane array have different output responses in a uniform infrared radiation scene. The main reasons for this are inconsistent pixel responsivity, unequal dark current, and different bias, which lead to fixed pattern noise in the infrared image and reduce the quality of infrared imaging.

[0003] A typical example of fixed-pattern noise is an image with an excessively large difference between the dark and bright areas around the edges and the center, forming a "pot lid" shape. This type belongs to the low-frequency non-uniformity problem of infrared images. It is mainly caused by the fact that during the operation of the infrared detector, the material temperature at the edge of the detector lens increases, resulting in a significantly higher temperature around the lens compared to the center; or thermal radiation from the lens edge material reaches the focal plane through the lens, producing a result where the gray value gradually increases from the center to the edges of the image. As the infrared imaging system continues to operate, the internal temperature of the detector increases, and the pot lid non-uniformity gradually worsens.

[0004] Methods for improving the non-uniformity of infrared images can be broadly classified into two categories: calibration-based methods (as described in this paper) and scene-based methods. Scene-based algorithms do not require a reference radiation source, do not undergo a calibration process, and do not need to be constantly recalibrated over time. They can adjust correction parameters according to scene information, adaptively transforming and making them suitable for observing distant targets, thus applicable in reconnaissance systems.

[0005] Currently, there are two main types of scene-based pot lid effect removal algorithms. One is a non-uniformity correction method based on singular value decomposition in the fuzzy domain. This method mainly suppresses pot lid-shaped non-uniformity by converting the image in the spatial and fuzzy domains to obtain the fundamental frequency information, and then subtracting it from the image before processing. However, this method involves singular value decomposition and eigenvector matrix transformation, making it extremely difficult to implement in hardware. The other type is a real-time pot lid effect suppression method proposed by some scholars. This method uses a complex mathematical model, requires multiple iterations, has a very high computational load, and is also difficult to implement in hardware. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing an infrared pot lid-shaped non-uniformity correction system and its correction algorithm based on scene gradient processing.

[0007] The invention aims to design an infrared pot-shaped non-uniformity correction system and its algorithm based on scene gradient processing. Implemented on an FPGA (Field-Programmable Gate Array), it provides a hardware storage solution for real-time data stream processing and output. Encapsulated as IP (Internet Protocol), it allows for the transmission, identification, and processing of data, information, or functions within an IP-based network environment. This IP can be integrated into infrared ISP (Image Signal Processing) algorithms to improve infrared imaging quality. The core of this invention is the design of a non-uniformity correction based on grayscale gradient processing to suppress pot-shaped image noise. The algorithm is implemented in hardware and packaged as IP. With a simple port, it only needs to meet the data input protocol to adapt to different detectors, thereby improving the imaging quality of infrared detectors.

[0008] The technical solution adopted in this invention is as follows:

[0009] An infrared pot lid-shaped non-uniformity correction system based on scene gradient processing comprises an image input stream module, an initialization module, a downsampling module, a median filtering module, a horizontal and vertical gradient processing module, an image reconstruction module, a bilinear interpolation module, an upsampling module, an algorithm processing and updating matrix module, and an image output stream module connected in sequence.

[0010] The median filtering module, the horizontal and vertical gradient processing module, and the image reconstruction module are all connected to RAM via caches.

[0011] The bilinear interpolation module and the upsampling module are connected to the FIFO via cache, and the algorithm processing and update matrix module is connected to the DDR via cache.

[0012] The image input stream module is the output video stream data processed by the ISP or the video stream data output by the infrared detector;

[0013] The initialization module performs mean preprocessing on a frame of image and provides an overall correction direction;

[0014] The downsampling module is used to smooth stripe noise in infrared images and reduce the amount of computation during processing;

[0015] The median filtering module is used to eliminate isolated noise points in an image;

[0016] The horizontal and vertical gradient processing modules are used to extract regions with excessive gradients in a frame of an image;

[0017] The image reconstruction module is used to restore image information;

[0018] The bilinear interpolation module is used to smooth images;

[0019] The upsampling module is used to restore the original size of the image;

[0020] The algorithm processing update matrix module is used to apply the calculated parameters to the pixel output data;

[0021] The image output stream module is used to output the processed video stream data.

[0022] To achieve real-time hardware processing, RAM is used for scenarios with small data volumes and low speed requirements, processing 5x5 window data. FIFO (First-In-First-Out) is used for buffering linear interpolation. DDR (High-Speed ​​DDR) is used for processing large frame data.

[0023] This invention also provides a correction algorithm for an infrared pot-shaped non-uniformity correction system based on scene gradient processing, comprising the following steps:

[0024] S1. Initialization process: Normalize the image pixel matrix to obtain a normalized matrix. To distribute the pixel values ​​between 0 and 1, the average value A of the entire normalized pixel matrix is ​​calculated, and then compressed after judgment. The process is shown in formula (1) to obtain the initialized image matrix. ;

[0025] (1)

[0026] S2. Next, the image matrix after initialization is downsampled, taking every other row and every other column. The remainder of the current frame number divided by 4 is taken, and the image size is changed from 2M rows and 2N columns to M rows and N columns, thus obtaining the downsampled image matrix.

[0027] S3. Perform median filtering on the downsampled image matrix with a 5*5 window size, cache the output of the 5*5 neighborhood window in RAM, and obtain the median filtering matrix;

[0028] S4. Perform horizontal gradient processing and vertical gradient processing on the median filter matrix respectively. The horizontal gradient processing process is shown in formula (2). Set all elements in the first row of the matrix to 0, and subtract the corresponding element of the previous row from each of the remaining rows. The difference obtained constitutes the horizontal gradient matrix. When implementing FPGA, the data of the previous row can be cached in RAM. The vertical gradient processing process is shown in formula (3). Set all elements in the first column of the matrix to 0, and subtract the corresponding element of the previous column from each of the remaining columns. The difference obtained constitutes the vertical gradient matrix.

[0029] (2)

[0030] (3)

[0031] in, The element in the i-th row and j-th column of the transverse gradient matrix. The element in the i-th row and j-th column of the ordinate gradient matrix;

[0032] Set threshold Find and The sum of squares D, as shown in formula (4), if D is greater than When , the elements at the corresponding positions in the horizontal and vertical gradient matrices are set to 0, that is =0, =0, otherwise remain unchanged.

[0033] (4);

[0034] S5. Reconstruct the image by processing the horizontal and vertical gradient matrices respectively. Then, calculate the mean of the first row of the median filter matrix to obtain... The values ​​are assigned to the first row of the transverse gradient reconstruction matrix. From the second row onwards, each element of the matrix is ​​equal to the gradient value of that row plus the reconstruction value of the first row, forming the transverse gradient reconstruction matrix. The process is shown in formula (5). In FPGA implementation, the data of the previous row can be cached in RAM. Then, the mean of the first column of the median filter matrix is ​​calculated to obtain the result. The values ​​are assigned to the first column of the vertical gradient reconstruction matrix. From the second column onwards, each column's element is equal to the gradient value of that column plus the reconstruction value of the next column, forming the vertical gradient reconstruction matrix. The process is shown in formula (6):

[0035] (5)

[0036] (6)

[0037] The obtained horizontal gradient reconstruction matrix and vertical gradient reconstruction matrix are combined to form a reconstruction matrix. The synthesis process is shown in formula (7):

[0038] (7)

[0039] Next, the image reconstruction matrix is ​​subjected to anchor point finding operation. Starting from the first row and first column element, an anchor point is determined every 8 rows and 8 columns. With the anchor point as the center, a 5*5 row and column window is selected as the mean filtering range for mean filtering. The mean value is assigned to the center anchor point, as shown in formula (8):

[0040] (8)

[0041] in, The value is 25. For all elements of a 5x5 row and column window, This is the anchor value for that row and column window;

[0042] S6. Perform quadratic linear interpolation on the matrix elements within the square formed by the four adjacent anchor points, as shown in formula (9), and update the image reconstruction matrix elements. ,

[0043] (9)

[0044] in, Let (x1, y1) be the anchor point. Let (x1, y2) be the anchor point. Let (x2, y1) be the anchor point. Let (x2, y2) be the anchor point.

[0045] Then perform a 5x5 row and column window mean filter again to remove the block effect. The block effect is a significant defect in the reconstructed image caused by discontinuities at the boundaries of blocks when quantization is coarse.

[0046] S7. Next, the elements of the obtained interpolation reconstruction matrix are divided by the mean and then upsampled. Two rows and two columns of the matrix are then used for bilinear interpolation to supplement elements in every other row and column. When the frame number is remainder 0 when divided by 4, odd rows and odd columns are supplemented; when the frame number is remainder 1 when divided by 4, even rows and odd columns are supplemented; when the frame number is remainder 2 when divided by 4, odd rows and even columns are supplemented; when the frame number is remainder 3 when divided by 4, even rows and even columns are supplemented. The matrix changes from the original M rows and N columns to 2M rows and 2N columns, thus obtaining the upsampled matrix. In FPGA implementation, a FIFO can be called to buffer the image matrix data.

[0047] S8. Obtain the gain matrix according to formula (10). In FPGA implementation, the parameters of the gain matrix can be called from the DDR memory. Then, the original image data is divided by the gain matrix at the corresponding position to obtain the corrected pixel matrix. The pot lid effect is removed based on scene correction.

[0048] ;

[0049] (10)

[0050] Where E is the identity matrix and rate is the correction parameter. Let be the upsampling matrix and 'a' be the gain matrix. The smallest element value is used to complete the correction process for the non-uniformity of the pot lid shape.

[0051] Because of the adoption of the above technical solution, the present invention has the following advantages: [The following text appears to be a separate, unrelated sentence fragment: "Utilizing the present invention..."] IP module It can directly improve the pot lid effect of infrared images, realize system-level optimization of infrared imaging quality, and provide real-time processing and output based on the scene. It can be applied in various practical scenarios. After configuring the input data protocol, it can be adapted to various types of infrared detectors without changing the code one by one. Moreover, the hardware solution implemented by this IP module consumes less resources, has a fast convergence speed, strong portability and versatility, and is easy to use in engineering. Attached Figure Description

[0052] Figure 1 This invention relates to the algorithm principle and FPGA system block diagram.

[0053] Figure 2 This is a partial schematic diagram of the algorithm anchor point determination method of the present invention;

[0054] Figure 3 The comparison diagrams show the correction effects of the algorithm of this invention, with the left side before correction and the right side after correction. Detailed Implementation

[0055] I. This invention provides an infrared pot lid-shaped non-uniformity correction system based on scene gradient processing, such as... Figure 1 As shown, its components include, in sequence, an image input stream module, an initialization module, a downsampling module, a median filtering module, a horizontal and vertical gradient processing module, an image reconstruction module, a bilinear interpolation module, an upsampling module, an algorithm processing and updating matrix module, and an image output stream module;

[0056] The median filtering module, the horizontal and vertical gradient processing module, and the image reconstruction module are all connected to RAM via caches.

[0057] The bilinear interpolation module and the upsampling module are connected to FIFT via cache, and the algorithm processing and update matrix module is connected to DDR via cache.

[0058] The image input stream module is the output video stream data processed by the ISP or the video stream data output by the infrared detector;

[0059] The initialization module performs mean preprocessing on a frame of image and provides an overall correction direction;

[0060] The downsampling module is used to smooth stripe noise in infrared images and reduce the amount of computation during processing;

[0061] The median filtering module is used to eliminate isolated noise points in an image;

[0062] The horizontal and vertical gradient processing modules are used to extract regions with excessive gradients in a frame of an image;

[0063] The image reconstruction module is used to restore image information;

[0064] The bilinear interpolation module is used to smooth images;

[0065] The upsampling module is used to restore the original size of the image;

[0066] The algorithm processing update matrix module is used to apply the calculated parameters to the pixel output data;

[0067] The image output stream module is used to output the processed video stream data.

[0068] To achieve real-time hardware processing, RAM is used for scenarios with small data volumes and low speed requirements, processing 5x5 window data. FIFO (First-In-First-Out) is used for buffering linear interpolation. DDR (High-Speed ​​DDR) is used for processing large frame data.

[0069] II. This invention also provides a correction algorithm for an infrared pot-shaped non-uniformity correction system based on scene gradient processing, comprising the following steps:

[0070] S1. First, perform initialization processing, normalizing the image pixel matrix to obtain a normalized matrix. To distribute the pixel values ​​between 0 and 1, the average value A of the entire normalized pixel matrix is ​​calculated, and then compressed after judgment. The process is shown in formula (1) to obtain the initialized image matrix. ;

[0071] (1)

[0072] S2. Next, the initialized image matrix is ​​downsampled, taking every other row and every other column. The frame number of the current frame is then divided by 4, and sampling is performed in four different ways:

[0073] (1) When the remainder is 0, the sampling method is odd row and odd column, a11, a13, ..., a31, a33, ..., a(m-1)(n-1);

[0074] (2) When the remainder is 1, the sampling method is even row and odd column, a21, a23, ..., a41, a43, ..., am(n-1);

[0075] (3) When the remainder is 2, the sampling method is odd row even column, a12, a14, ..., a31, a33, ..., a(m-1)n;

[0076] (4) When the remainder is 3, the sampling method is even row and even column, a22, a24, ..., a42, a44, ..., amn;

[0077] The image size is changed from 2M rows and 2N columns to M rows and N columns, resulting in a downsampled image matrix;

[0078] S3. Perform median filtering on the downsampled image matrix with a 5*5 window size, cache the output of the 5*5 neighborhood window in RAM, and obtain the median filtering matrix;

[0079] S4. Perform horizontal gradient processing and vertical gradient processing on the median filter matrix respectively. The horizontal gradient processing process is shown in formula (2). Set all elements in the first row of the matrix to 0, and subtract the corresponding element of the previous row from each of the remaining rows. The difference obtained constitutes the horizontal gradient matrix. When implementing FPGA, the data of the previous row can be cached in RAM. The vertical gradient processing process is shown in formula (3). Set all elements in the first column of the matrix to 0, and subtract the corresponding element of the previous column from each of the remaining columns. The difference obtained constitutes the vertical gradient matrix.

[0080] (2)

[0081] (3)

[0082] in, The element in the i-th row and j-th column of the transverse gradient matrix. The element in the i-th row and j-th column of the ordinate gradient matrix;

[0083] Set threshold Find and The sum of squares D, as shown in formula (4), if D is greater than When , the elements at the corresponding positions in the horizontal and vertical gradient matrices are set to 0, that is =0, =0, otherwise remain unchanged.

[0084] (4)

[0085] S5. Reconstruct the image by processing the horizontal and vertical gradient matrices respectively. Then, calculate the mean of the first row of the median filter matrix to obtain... The values ​​are assigned to the first row of the transverse gradient reconstruction matrix. Each subsequent row (starting from the second row) has its element equal to the gradient value of that row plus the reconstruction value of the previous row, forming the transverse gradient reconstruction matrix. The process is shown in formula (5). In the FPGA implementation, the RAM can be used to cache the data of the previous row. Then, the mean of the first column of the median filter matrix is ​​calculated to obtain... The values ​​are assigned to the first column of the vertical gradient reconstruction matrix. Then, the elements of each column (starting from the second column) are equal to the gradient value of that column plus the reconstruction value of the next column, forming the vertical gradient reconstruction matrix. The process is shown in formula (6):

[0086] (5)

[0087] (6)

[0088] The obtained horizontal gradient reconstruction matrix and vertical gradient reconstruction matrix are combined to form a reconstruction matrix. The synthesis process is shown in formula (7):

[0089] (7)

[0090] Next, anchor point finding is performed on the image reconstruction matrix. Starting from the first row and first column, an anchor point is determined every 8 rows and 8 columns. The anchor point determination method is as follows: Figure 2 As shown, the marked position 'a' is the anchor point.

[0091] Apply mean filtering centered on the anchor point, with a 5x5 window selected. Figure 2 The blue box represents the mean filtering range, and the mean is assigned to the center anchor point, as shown in formula (8):

[0092] (8)

[0093] in, The value is 25. For all elements of a 5x5 row and column window, This is the anchor value for that row and column window;

[0094] S6. Perform quadratic linear interpolation on the matrix elements within the square formed by the four anchor points, with the range as described above. Figure 2 As shown in the red box, the quadratic linear interpolation is performed as shown in formula (9) to update the elements of the image reconstruction matrix.

[0095] (9)

[0096] in, Let (x1, y1) be the anchor point. Let (x1, y2) be the anchor point. Let (x2, y1) be the anchor point. Let (x2, y2) be the anchor point.

[0097] A 5x5 mean filter is applied again to remove block artifacts. Block artifacts are a noticeable defect in the reconstructed image caused by discontinuities at block boundaries when quantization is coarse.

[0098] S7. Next, the elements of the obtained interpolation reconstruction matrix are divided by the mean area and then upsampled. Two rows and two columns of the matrix are then used for bilinear interpolation to supplement elements in every other row and column. When the frame number is remainder 0 when divided by 4, odd rows and odd columns are supplemented; when the frame number is remainder 1 when divided by 4, even rows and odd columns are supplemented; when the frame number is remainder 2 when divided by 4, odd rows and even columns are supplemented; and when the frame number is remainder 3 when divided by 4, even rows and even columns are supplemented. The matrix changes from the original M rows and N columns to 2M rows and 2N columns, thus obtaining the upsampled matrix. In FPGA implementation, a FIFO can be used to buffer the image matrix data.

[0099] S8. Obtain the gain matrix according to formula (10). In FPGA implementation, the parameters of the gain matrix can be called from the DDR memory. Then, the original image data is divided by the gain matrix at the corresponding position to obtain the corrected pixel matrix. The pot lid effect is removed based on scene correction.

[0100] ;

[0101] (10)

[0102] Where E is the identity matrix and rate is the correction parameter. Let be the upsampling matrix and 'a' be the gain matrix. The smallest element value is used to complete the correction process for the non-uniformity of the pot lid shape.

[0103] Figure 3 The present invention discloses an infrared pot lid-shaped non-uniformity correction system and its correction algorithm based on scene gradient processing. The comparison of the correction effect in a real scene of a parking lot shows that the phenomenon of "black corners and bright center" in the right image is significantly and effectively suppressed compared with the left image, and the details in the scene are restored.

[0104] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An infrared pot lid-shaped non-uniformity correction system based on scene gradient processing, characterized in that, Its components include, in sequence, an image input stream module, an initialization module, a downsampling module, a median filtering module, a horizontal and vertical gradient processing module, an image reconstruction module, a bilinear interpolation module, an upsampling module, an algorithm processing and updating matrix module, and an image output stream module; The median filtering module, the horizontal and vertical gradient processing module, and the image reconstruction module are all connected to RAM via caches. The bilinear interpolation module and the upsampling module are connected to the FIFO via cache, and the algorithm processing and update matrix module is connected to the DDR via cache. The image input stream module is the output video stream data processed by the ISP or the video stream data output by the infrared detector; The initialization module performs mean preprocessing on a frame of image and provides an overall correction direction; The downsampling module is used to smooth stripe noise in infrared images and reduce the amount of computation during processing; The median filtering module is used to eliminate isolated noise points in an image; The horizontal and vertical gradient processing modules are used to extract regions with excessive gradients in a frame of an image; The image reconstruction module is used to restore image information; The bilinear interpolation module is used to smooth images; The upsampling module is used to restore the original size of the image; The algorithm processing update matrix module is used to apply the calculated parameters to the pixel output data; The image output stream module is used to output the processed video stream data.

2. The correction algorithm for an infrared pot-shaped non-uniformity correction system based on scene gradient processing according to claim 1, comprising the following steps: S1. Initialization process: Normalize the image pixel matrix to obtain a normalized matrix. To distribute the pixel values ​​between 0 and 1, the average value A of the entire normalized pixel matrix is ​​calculated, and then compressed after judgment. The process is shown in formula (1) to obtain the initialized image matrix. ; （1） S2. Next, the image matrix after initialization is downsampled, taking every other row and every other column. The frame number of the current frame is divided by 4 and the remainder is taken. The image size is changed from 2M rows and 2N columns to M rows and N columns, resulting in the downsampled image matrix. S3. Perform median filtering on the downsampled image matrix with a 5*5 window size, cache the output of the 5*5 neighborhood window in RAM, and obtain the median filtering matrix; S4. Perform horizontal gradient processing and vertical gradient processing on the median filter matrix respectively. The horizontal gradient processing process is shown in formula (2). Set all elements in the first row of the matrix to 0, and subtract the corresponding element of the previous row from each of the remaining rows. The difference obtained constitutes the horizontal gradient matrix. When implementing FPGA, the data of the previous row can be cached in RAM. The vertical gradient processing process is shown in formula (3). Set all elements in the first column of the matrix to 0, and subtract the corresponding element of the previous column from each of the remaining columns. The difference obtained constitutes the vertical gradient matrix. （2） （3） in, The element in the i-th row and j-th column of the transverse gradient matrix. The element in the i-th row and j-th column of the ordinate gradient matrix; Set threshold Find and The sum of squares D, as shown in formula (4), if D is greater than When , the elements at the corresponding positions in the horizontal and vertical gradient matrices are set to 0, that is =0, =0, otherwise remain unchanged. （4）； S5. Reconstruct the image by processing the horizontal and vertical gradient matrices respectively. Then, calculate the mean of the first row of the median filter matrix to obtain... The values ​​are assigned to the first row of the transverse gradient reconstruction matrix. From the second row onwards, each element of the matrix is ​​equal to the gradient value of that row plus the reconstruction value of the first row, forming the transverse gradient reconstruction matrix. The process is shown in formula (5). In FPGA implementation, the data of the previous row can be cached in RAM. Then, the mean of the first column of the median filter matrix is ​​calculated to obtain the result. The values ​​are assigned to the first column of the vertical gradient reconstruction matrix. From the second column onwards, each column's element is equal to the gradient value of that column plus the reconstruction value of the next column, forming the vertical gradient reconstruction matrix. The process is shown in formula (6): （5） （6） The obtained horizontal gradient reconstruction matrix and vertical gradient reconstruction matrix are combined to form a reconstruction matrix. The synthesis process is shown in formula (7): （7） Next, the image reconstruction matrix is ​​subjected to anchor point finding operation. Starting from the first row and first column element, an anchor point is determined every 8 rows and 8 columns. With the anchor point as the center, a 5*5 row and column window is selected as the mean filtering range for mean filtering. The mean value is assigned to the center anchor point, as shown in formula (8): （8） in, The value is 25. For all elements of a 5x5 row and column window, This is the anchor value for that row and column window; S6. Perform quadratic linear interpolation on the matrix elements within the square formed by the four adjacent anchor points, as shown in formula (9), and update the image reconstruction matrix elements. , （9） in, Let (x1, y1) be the anchor point. Let (x1, y2) be the anchor point. Let (x2, y1) be the anchor point. Let (x2, y2) be the anchor point. Then perform a 5x5 row and column window mean filter again to remove the block effect. The block effect is a significant defect in the reconstructed image caused by discontinuities at the boundaries of blocks when quantization is coarse. S7. Next, the elements of the obtained interpolation reconstruction matrix are divided by the mean and then upsampled. Two rows and two columns of the matrix are then used for bilinear interpolation to supplement elements in every other row and column. When the frame number is remainder 0 when divided by 4, odd rows and odd columns are supplemented; when the frame number is remainder 1 when divided by 4, even rows and odd columns are supplemented; when the frame number is remainder 2 when divided by 4, odd rows and even columns are supplemented; when the frame number is remainder 3 when divided by 4, even rows and even columns are supplemented. The matrix changes from the original M rows and N columns to 2M rows and 2N columns, thus obtaining the upsampled matrix. In FPGA implementation, a FIFO can be called to buffer the image matrix data. S8. Obtain the gain matrix according to formula (10). In FPGA implementation, the parameters of the gain matrix can be called from the DDR memory. Then, the original image data is divided by the gain matrix at the corresponding position to obtain the corrected pixel matrix. The pot lid effect is removed based on scene correction. ; （10） Where E is the identity matrix and rate is the correction parameter. Let be the upsampling matrix and 'a' be the gain matrix. The smallest element value is used to complete the correction process for the non-uniformity of the pot lid shape.

Images