Adaptive Neural Network Non-Uniformity Correction Method, Terminal Device and Storage Medium

Abstract

The present application provides an adaptive neural network non-uniformity correction method, a terminal device, and a storage medium. The adaptive neural network non-uniformity correction method includes the following steps: collecting a plurality of images in real time; selecting valid images from the plurality of images; the valid images include a plurality of pixels distributed in a matrix; obtaining the pixels with neighborhoods in the valid images, that is, valid pixels; the neighborhood is a set of pixels distributed along the circumferential direction of the valid pixel and at a 45-degree angle to the valid pixel; obtaining the input irradiance X of the valid pixel i,j ; obtaining the output irradiance Y of the valid pixel through a formula i,j , and the parameters in the formula include the variance of the pixel neighborhood, the gain correction coefficient, and the bias correction coefficient. This method can obtain an ideal image without ghosting and without aliasing effects.

Description

Technical Field

[0001] This application relates to the technical field of scene non-uniformity correction, and specifically relates to an adaptive neural network non-uniformity correction method, a terminal device, and a storage medium. Background Art

[0002] The traditional non-uniformity correction method is the two-point correction method, which requires calibration with a standard reference source, has cumbersome operations, needs to be recalibrated regularly, and does not have the ability of adaptive correction. Therefore, it has gradually been replaced by the scene non-uniformity correction method.

[0003] The traditional adaptive neural network non-uniformity correction method belongs to the scene variable step-size correction method, and its step size changes with the eight-neighborhood variance value. On the one hand, it can determine the speed at which each point approaches the target expected value according to the eight-neighborhood variance information of each point, thus effectively suppressing the generation of ghost images. On the other hand, it may cause the algorithm to diverge due to improper step size adjustment, resulting in over-bright or over-dark noise points in the image, affecting the imaging quality. In addition, due to the existence of noise points, the traditional neural network may generate diffuse bright spots during the correction process. Summary of the Invention

[0004] In view of the above defects or deficiencies in the prior art, this application aims to provide an adaptive neural network non-uniformity correction method, a terminal device, and a storage medium.

[0005] In a first aspect, the adaptive neural network non-uniformity correction method includes the following steps:

[0006] Collect a number of images in real time;

[0007] Select valid images from the number of images, and the valid images include a plurality of pixels distributed in a matrix;

[0008] Obtain the pixels with neighborhoods in the valid images, that is, valid pixels; the neighborhood is a set of pixels distributed along the circumferential direction of the valid pixel and at a 45-degree angle to the valid pixel;

[0009] Obtain the input irradiance X of the valid pixel in the valid image i,j ;

[0010] Calculate the output irradiance Y of the valid pixel through the following formula i,j

[0011] Y i,j =K i,j X i,j +B i,j (1)

[0012] K i,j (n + 1)=K i,j(n) - 2η(n)X i,j (n)(Y i,j (n) - f i，j ) (2)

[0013] B i,j (n + 1) = B i,j (n) - 2η(n)(Y i,j (n) - f i，j ) (3)

[0014]

[0015] where i is the row number of the effective pixel in the effective image;

[0016] j is the column number of the effective pixel in the effective image;

[0017] K i,j is the gain correction coefficient;

[0018] B i,j is the bias correction coefficient;

[0019] f i，j is the neural network correction expected value;

[0020] η(n) is the update step of the effective pixel;

[0021] n is the processing sequence number;

[0022] is the neighborhood variance of the effective pixel;

[0023] step is the initial step;

[0024] According to the technical solution provided by the embodiments of the present application, the steps of selecting the effective image are as follows:

[0025] Obtain a number of input key frames from the number of images according to the first set rule;

[0026] Remove the bad pixels of the input key frames according to the second set rule;

[0027] Adopt the Sobel operator method to remove the edge pixels of the input key frames;

[0028] Obtain the effective key frames.

[0029] According to the technical solution provided by the embodiments of the present application, the first set rule is that when the displacement amount between the current frame and the previous frame is greater than the set value, it is the input key frame.

[0030] According to the technical solution provided by the embodiments of the present application, the second setting rule is that when the number of times the pixel in the input key frame is a candidate defective pixel is greater than the set value, the pixel is a defective pixel.

[0031] In a second aspect, the present application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned adaptive neural network non-uniformity correction method are implemented.

[0032] In a third aspect, the present application provides a computer-readable storage medium, which has a computer program. When the computer program is executed by a processor, the steps of the above-mentioned adaptive neural network non-uniformity correction method are implemented.

[0033] In summary, the present application proposes an adaptive neural network non-uniformity correction method. By selecting a plurality of pixels distributed in a matrix from several real-time collected images, valid pixels are selected from these pixels. The valid pixels have neighboring pixels distributed along their circumferences and at a 45-degree angle to them. Taking the input irradiance of the valid pixels as the input, and the neighborhood variance, gain correction coefficient, and bias correction coefficient as parameters, the output irradiance of the valid pixels is obtained, so that an ideal image without ghosting and sawtooth effect can be obtained. It realizes that the infrared imaging system does not need to plan the correction area and the scene movement mode, only needs to automatically extract image frames from the actual scene, and obtains an ideal image through this non-uniformity correction method. Description of the Drawings

[0034] Figure 1 It is a flowchart of the adaptive neural network non-uniformity correction method provided by the embodiments of the present application;

[0035] Figure 2 It is the neighborhood information of the vertical stripe image provided by the embodiments of the present application;

[0036] Figure 3 It is the comparison of the vertical stripe correction effects provided by the embodiments of the present application;

[0037] Figure 4 It is the convergence graph of the bias coefficient provided by the embodiments of the present application;

[0038] Figure 5 It is the adjacent frame images of relative motion provided by the embodiments of the present application;

[0039] Figure 6 It is the adjacent frame row vector projection information provided by the embodiments of the present application;

[0040] Figure 7 It is the comparison graph of the defective pixel repair effect provided by the embodiments of the present application;

[0041] Figure 8 The non-cooled infrared image provided by the embodiment of the present application;

[0042] Figure 9 The cooled infrared image provided by the embodiment of the present application;

[0043] Figure 10 The comparison of the images before and after the non-cooled image correction of the laboratory scene provided by the embodiment of the present application;

[0044] Figure 11 The comparison of the images before and after the non-cooled image correction of the outdoor scene provided by the embodiment of the present application;

[0045] Figure 12 The non-uniformity degraded image of the sky scene provided by the embodiment of the present application;

[0046] Figure 13 The non-uniformity degraded image of the building scene provided by the embodiment of the present application;

[0047] Figure 14 The non-uniformity degraded image of the downward-looking building scene provided by the embodiment of the present application;

[0048] Figure 15 The non-uniformity degraded image of the downward-looking airport scene provided by the embodiment of the present application;

[0049] Figure 16 The principle block diagram of Embodiment 3 of the present application. Detailed implementation manners

[0050] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the related invention, rather than limiting the invention. Additionally, it should be noted that for the convenience of description, only the parts related to the invention are shown in the drawings.

[0051] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the drawings and embodiments.

[0052] Embodiment 1

[0053] As mentioned in the background art, in view of the problems in the prior art, the present application proposes an adaptive neural network non-uniformity correction method, including the following steps:

[0054] Collect a plurality of images in real time;

[0055] Select valid images from the plurality of images, and the valid images include a plurality of pixels distributed in a matrix;

[0056] Obtain the pixels with neighborhoods in the effective image, which are the effective pixels; the neighborhood is a set of pixels distributed circumferentially along the effective pixel and at a 45-degree angle to the effective pixel;

[0057] Obtain the input irradiance X of the effective pixel in the effective image i,j ;

[0058] Calculate the output irradiance Y of the effective pixel through the following formula i,j :

[0059] Y i,j =K i,j X i,j +B i,j (1)

[0060] K i,j (n + 1)=K i,j (n)-2η(n)X i,j (n)(Y i,j (n)-f i，j ) (2)

[0061] B i,j (n + 1)=B i,j (n)-2η(n)(Y i,j (n)-f i，j ) (3)

[0062]

[0063] where i is the row number of the effective pixel in the effective image, j is the column number of the effective pixel in the effective image, K i,j is the gain correction coefficient, B i,j is the bias correction coefficient, f i，j is the neural network correction expected value, η(n) is the update step size of the effective pixel, n is the processing sequence number, is the neighborhood variance of the effective pixel; step is the initial step size;

[0064] Set the pixel to be linearly responsive within a certain temperature range. The sensor pixel in the i-th row and j-th column converts the irradiance X of the target scene i,j into a measured value Y i,j through linear response and outputs it. This linear model can be written as:

[0065] Y i,j =K i,j X i,j +B i,j (1)

[0066] The non-uniformity correction of the neural network adopts a linear fitting method to comprehensively consider the spatial domain noise of the scene environment. The algorithm consists of three parts: linear correction, linear smoothing, and correction parameter modification;

[0067] The expected value of the neural network correction is f i，j , and the error function E is expressed as:

[0068] E=(K i,j (n)X i,j (n)+B i,j (n)-f i,j ) 2 (5)

[0069] According to the steepest descent method, the gradient of E is:

[0070]

[0071]

[0072] Along the descent path that minimizes the error, the gain and bias are corrected to formulas (2) and (3); by continuously iterating and updating the K and B parameters, they finally converge.

[0073] The traditional neural network neighborhood is the "cross"-shaped four-neighborhood pixels above, below, left, and right of the pixel to be corrected. By analyzing the vertical stripe image, it is found that in the "cross"-shaped neighborhood, if the pixel to be corrected carries vertical stripe information, its upper and lower adjacent pixels must also carry vertical stripe information, and the left and right adjacent pixels are also likely to carry vertical stripe information. Therefore, in order to balance the accuracy of the neighborhood expected value and the vertical stripe suppression effect, the design idea of the "square"-shaped neighborhood is proposed. The effective pixel has a set of pixels distributed along its circumference and at a 45-degree angle to it, that is, the upper-left, upper-right, lower-left, and lower-right pixels of the effective pixel. The neighborhood information of the vertical stripe image is as Figure 2 shown; it can be seen from the figure that the central pixel to be processed is a dark stripe, and there are 2 pixels with dark stripe information in its "cross"-shaped neighborhood, while there is only 1 pixel with dark stripe information in the "square"-shaped neighborhood. As Figure 3 shown, the method of selecting the "square"-shaped neighborhood can effectively suppress the mutual transmission of vertical stripe information in the neighborhood and has a faster and better suppression effect on vertical stripe type non-uniformity.

[0074] Meanwhile, by analyzing Formulas (2) and (3), it is found that the selection of η(n) has a great influence on the convergence of the algorithm. To avoid the "ghosting" caused by the excessive update of the learning step size due to the appearance of high-brightness, over-dark, or texture-rich scenes, the neighborhood variance of each valid pixel can be used as the parameter for updating the correction step size of the valid pixel. When a high-brightness, over-dark, or texture-rich scene appears, the update step size of the valid pixel to be updated decreases due to the large neighborhood variance, thus effectively avoiding the appearance of "ghosting". The convergence speed of the algorithm is related to the severity of image non-uniformity. The original infrared image is used to evaluate the convergence performance of the algorithm, and the result is the maximum number of frames required for the algorithm to converge. As Figure 4 shown, for the multi-frame iteration of the original infrared image, the update result of the bias coefficient, the algorithm converges quickly within 50 frames and tends to be stable around 70 frames. The update step size η(n) of each valid pixel is determined by Formula (4):

[0075] After the above improvements, the parameter update formulas (2) and (3) of the traditional neural network correction are rewritten as:

[0076]

[0077]

[0078] In this method, several pixels distributed in a matrix are selected from the real-time acquired images, and valid pixels are selected from these pixels. The valid pixels have neighborhood pixels distributed along their circumferences and at a 45-degree angle to them. Taking the input irradiance of the valid pixels as the input and the neighborhood variance, gain correction coefficient, and bias correction coefficient as parameters, the output irradiance of the valid pixels is obtained, so that the final ideal image without ghosting and sawtooth effect can be obtained, realizing that the infrared imaging system does not need to plan the correction area and the scene movement mode, but only needs to automatically extract image frames from the actual scene and obtain the ideal image through this non-uniformity correction method.

[0079] Furthermore, the steps for selecting the valid images are as follows:

[0080] Several input key frames are obtained from the several images according to the first setting rule; the first setting rule is that when the displacement between the current frame and the previous frame is greater than the set value, it is the input key frame. The displacement is obtained through the motion estimation algorithm of gray projection. Specifically, it is assumed that there is sufficient motion between adjacent frame images, then there will be relative displacements in the horizontal and vertical directions. As Figure 5 shown, it is assumed that the area array of adjacent frame images is M×N, and let P be the projection vector in the row or column direction of the image (determined by the subscript), then the row and column projection vectors of the image can be obtained by the following formula:

[0081]

[0082]

[0083] For the scene information of motion, when using row-column projection vectors, the scene information newly entering the field of view and just exiting the field of view will greatly affect the accuracy of registration to a large extent because this part of the information has undergone a "qualitative change". Therefore, it is necessary to make certain corrections to the projection vectors to improve their robustness. In the present invention, the method of cosine filtering is used to correct the boundary data of the projection vectors, and the following formula is obtained:

[0084]

[0085]

[0086] where δ row and δ col are the maximum possible displacement numbers in the row and column directions. The corrected row-column projection vectors are calculated one by one using the Euclidean distance, and it is judged on which displacement the minimum value appears. The position value obtained at this time is the position where the adjacent frame images occur. The minimum value is obtained through the following formula:

[0087]

[0088]

[0089]

[0090]

[0091] As shown in 1) of Figure 6 , the abscissa is the number of rows, the ordinate is the projection value, and the two curves represent the row projection vectors of the adjacent frame images of the same scene. It can be seen that the part except for the head and tail maintains a large amount of consistency. As shown in 2) of Figure 6 , it is the set of adjacent frame row projection vectors, that is, the projection curves constructed by using the exhaustive method with all possible displacement values are compared with the No. 1 curve one by one, and the No. 2 curve that basically coincides is found. The displacement between the No. 2 curve and the No. 1 curve can be considered as the magnitude of the motion displacement of the adjacent frame in the row direction. Similarly, the column direction is obtained by a similar method. Thus, the relative displacement amounts in the row and column directions of the adjacent frames can be obtained through the present invention. When the relative displacement amount is greater than a certain threshold, it indicates that the scene is fully moving, and it can be selected as the key frame of the neural network algorithm.

[0092] Remove the bad pixels of the input key frame according to the second set rule; the second set rule is that when the number of times the pixel in the input key frame is determined to be a candidate bad pixel is greater than the set value, the pixel is a bad pixel. When the scene is relatively static or the displacement is very small, the scene information will be regarded as image non-uniformity or bad pixels. The correction parameters obtained from it will cause ghosting in the corrected image, and the bad pixels detected from it will cause misdetection and missed detection of bad pixels. Therefore, bad pixels need to be removed. The dynamic bad pixel detection algorithm based on the scene is used to detect bad pixels. Specifically, it includes the following steps:

[0093] 1) Establish a list of candidate bad pixels for the infrared detector. The list of candidate bad pixels stores the number of times each pixel of the detector is determined to be a bad pixel in the key frames extracted multiple times;

[0094] 2) Use the motion estimation algorithm based on gray projection to extract key frames. In the key frame, with the neighborhood as the template, compare the absolute value of the gray difference between the central pixel and each pixel in the neighborhood in turn. If the difference is greater than a certain threshold, the count value is incremented by 1. If the count value is greater than half of the total number of neighborhood pixels, this pixel is regarded as a candidate bad pixel, and the count value corresponding to this pixel in the list of candidate bad pixels is incremented by one;

[0095] 3) Repeat the process of step 2) to update the count value in the list of candidate bad pixels;

[0096] 4) Repeat step 3) 15 times. In these 15 statistics, if the count value of a certain pixel in the list of candidate bad pixels exceeds 12 times, this pixel is regarded as a bad pixel and added to the list of bad pixels. In this way, the entire dynamic bad pixel detection process is completed.

[0097] Since this bad pixel detection algorithm performs key frame extraction and time-domain statistics, it can effectively avoid misjudging small target points or bright or dark scenes in a static scene as bad pixels, thus improving the reliability of bad pixel detection. The comparison chart of bad pixel repair is as Figure 7 shown. The traditional bad pixel correction algorithm uses neighborhood information to replace bad pixels. This method is prone to the "jagged" effect at the image edge. The algorithm of this application makes the output pixel gradually approach the expected output value through multi-frame iteration, thus avoiding the generation of the "jagged" effect.

[0098] Use the Sobel operator method to remove the edge pixels of the input key frame; in an actual infrared imaging system, when the scene does not move or moves slowly for a long time, due to the large gray difference between the edge signal and the neighborhood expected value, the edge signal will "melt" into the background through the time-domain iteration process, resulting in the ghosting phenomenon. Use the Sobel operator for edge detection, and perform gray weighting operations within a 3×3 neighborhood centered on the pixel point to determine whether this point is an edge. The horizontal and vertical direction templates are as shown in the following formula:

[0099]

[0100] The x value of the nth frame image i,j (n) The x and y value edge images of the nth frame image are obtained by edge extraction using the Sobel operator.

[0101] The edge information is used to control the calculation of the expected output of each pixel and the update of the correction coefficient. The edge pixels of the binary edge image are represented by the set Φ, the non-edge pixels are represented by the set Θ, the defective pixels are represented by the set Ο, and the neighborhood expected value of the pixel x i,j (n) is represented by Ω. The expected output of each pixel is as follows:

[0102]

[0103] where <·> represents the neighborhood expected value selection strategy. If x i,j (n) is an edge and not a defective pixel, the expected output is x i,j (n). If x i,j (n) is an edge and a defective pixel, the expected output is the gray average of the non-edge points in the neighborhood. If x i,j (n) is a non-edge point, the expected output is the gray average of the non-edge points in the neighborhood. This selection strategy, on the one hand, based on the edge detection result, establishes an effective "isolation zone" at the edge of the scene, preventing the edge signal from spreading to the neighborhood and avoiding the occurrence of the ghosting problem; on the other hand, it overcomes the problem that the traditional neural network uses the four-neighborhood mean of pixels as the expectation. When there are fine grids and relatively dense defective points in the image, the blurring effect is greater than the detail recognition effect, and the image produces blurring and diffusion spots.

[0104] Obtain the effective key frame. After removing the defective and edge pixels, the effective key frame is obtained. Substitute formula (19) into formulas (8) and (9), and E n (i,j) replaces f i，j to obtain the gain and bias correction formulas as follows;

[0105]

[0106] Embodiment 2

[0107] On the basis of Embodiment 1, using an adaptive neural network non-uniformity correction method described in Embodiment 1, a non-cooled 384×288 core and a cooled long-wave 640×512 core are adopted to collect typical image sequences, and the algorithm is simulated and verified. The non-cooled infrared image is obtained as Figure 8 shown, and the cooled infrared image is obtained as Figure 9 shown, and from Figure 8 and Figure 9As shown, when there is a highlighted target in the scene, ghosts appear in the image corrected by the traditional scene non-uniformity correction algorithm. For example, Figure 8 on the right side of the water cup shown in the second figure of Figure 9 and on the upper and right edges of the leftmost building in the second figure of

[0108] The effects of non-uniformity correction of uncooled infrared images in different scenes are as shown in Figure 10 and Figure 11 As shown, as shown in Figure 10 the correction area is a laboratory scene with irregular scene motion (including motion modes such as scene stillness, rotation, horizontal or / and vertical translation, occlusion, etc.); as shown in Figure 11 the correction area is an outdoor scene with irregular scene motion (including motion modes such as scene stillness, rotation, horizontal or / and vertical translation, occlusion, etc.).

[0109] The effects of non-uniformity correction of cooled infrared images in different scenes are as shown in Figures 12 - 15 It can be seen that the method of the present application has obvious correction effects on non-uniformity and bad pixels under irregular scene motion.

[0110] Embodiment 3

[0111] Next, referring to Figure 16 FIG., there is shown a schematic structural diagram of a computer system 700 of a terminal device or a server suitable for implementing the embodiments of the present application.

[0112] As shown in Figure 16 FIG., the computer system 700 includes a central processing unit (CPU) 701, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 702 or the program loaded from the storage section 708 into the random access memory (RAM) 703. In the RAM 703, various programs and data required for the operation of the system 700 are also stored. The CPU 701, ROM 702, and RAM 703 are connected to each other through a bus 704. The input / output (I / O) interface 705 is also connected to the bus 704.

[0113] The following components are connected to the I / O interface 705: an input section 706 including a keyboard, a mouse, etc.; an output section 707 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc. and a speaker, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a LAN card, a modem, etc. The communication section 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to the I / O interface 705 as needed. A removable medium 711 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is mounted on the drive 710 as needed so that a computer program read therefrom is installed into the storage section 708 as needed.

[0114] Specifically, according to an embodiment of the present disclosure, the process described above with reference to Figure 1 can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product that includes a computer program tangibly embodied on a machine-readable medium, the computer program including program code for performing Figure 1 the method. In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 709, and / or installed from the removable medium 711.

[0115] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagram may represent a module, a program segment, or a portion of code that contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions noted in the blocks may occur in a different order than noted in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0116] Embodiment 3 of this application also provides a computer-readable storage medium, which may be the computer-readable storage medium included in the device described in the above embodiment; or it may exist separately and not be assembled into the device. The computer-readable storage medium stores one or more programs, and the one or more programs are used by one or more processors to execute the steps of the adaptive neural network non-uniformity correction method described in Embodiment 1.

[0117] In this text, specific examples are used to elaborate on the principles and implementation manners of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application. The above is only the preferred implementation manner of the present application. It should be noted that due to the limited nature of literal expression and the objectively infinite specific structures, for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements, refinements or changes can be made, or the above technical features can be combined in an appropriate manner; these improvements, refinements, changes or combinations, or directly applying the concept and technical solution of the invention to other occasions without improvement, shall all be regarded as the protection scope of the present application.

Claims

1. An adaptive neural network non-uniformity correction method, characterized in that It includes the following steps: Collect a number of images in real time; Select valid images from the number of images, where the valid images include a plurality of pixels distributed in a matrix; Obtain the pixels with a neighborhood in the valid images, which are valid pixels; the neighborhood is a set of pixels distributed circumferentially along the valid pixels and at a 45-degree angle to the valid pixels; Obtain the input irradiance X of the valid pixels in the valid image i,j ; The output irradiance Y of the effective pixel is calculated through the following formula i,j :[[]]END]] Y i,j = K i,j X i,j + B i,j (1) K i,j (n + 1) = K i,j (n) - 2η(n)X i,j (n)(Y i,j (n) - f i，j ) (2) B i,j (n + 1)= B i,j (n)-2η(n)(Y i,j (n)-f i，j ) (3) where i is the row number of the valid pixel in the valid image; j is the column number of the valid pixel in the valid image; K i,j is the gain correction coefficient; B i,j is the offset correction coefficient; f i，j correct the expected value for the neural network; η(n) is the update step size of the valid pixel; n is the processing sequence number; is the neighborhood variance of the effective pixel; step is the initial step size.

2. The adaptive neural network non-uniformity correction method according to claim 1, wherein: The steps of selecting the valid images are as follows: Obtain a number of input key frames from the number of images according to the first set rule; Remove the bad pixels of the input key frames according to the second set rule; Use the Sobel operator method to remove the edge pixels of the input key frames; Obtain valid key frames.

3. The adaptive neural network non-uniformity correction method according to claim 2, wherein: The first set rule is that when the displacement between the current frame and the previous frame is greater than the set value, it is the input key frame.

4. The adaptive neural network non-uniformity correction method according to claim 3, wherein: The second set rule is that when the number of times the pixel in the input key frame is a candidate bad pixel is greater than the set value, the pixel is a bad pixel.

5. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the adaptive neural network non-uniformity correction method according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the adaptive neural network non-uniformity correction method according to any one of claims 1 to 4.

Images