Feature point extraction method for dark light images based on machine learning

Abstract

The feature point extraction method for dark light images based on machine learning solves the problem that existing technologies have poor image processing effects and complex algorithms for images actually captured by cameras, and belongs to the field of image processing technology. The present invention includes the following steps: S1, removing background noise from N frames of adjacent or similar images; S2, performing wavelet denoising on the N frames of adjacent or similar images after background noise removal, and removing pixels below a set threshold to obtain N frames of adjacent or similar images after denoising; S3, using the optical flow method to remove pixels whose optical flow variance exceeds the mean of the optical flow variance matrix in the N frames of adjacent or similar images after denoising, to obtain N frames of adjacent or similar images with large jitter removed; S4, using the moving Gaussian convolution algorithm to extract feature points from the N frames of adjacent or similar images with large jitter removed to obtain a feature image. The present invention achieves high-precision and high-robustness feature extraction.

Description

Technical Field

[0001] The present invention relates to a method for extracting feature points of dark-light images based on machine learning, and belongs to the technical field of image processing. Background Art

[0002] Useful information in the image can be extracted, and these feature points with effective information can be used for orientation correction, target recognition, and real-time positioning and mapping (SLAM) processing. There are many existing image feature extraction methods, including corner detection algorithms such as SIFT algorithm, SURF algorithm, BRIEF algorithm, and many other feature extraction methods to obtain image feature points. Most of these methods are for extracting feature points of normal light images. However, in the fields of space optical communication, autonomous driving, real-time positioning, etc., the images actually obtained by the camera will have many problems such as dark images and low signal-to-noise ratio, and the existing feature extraction methods for normal light images are not good when applied to dark light image feature extraction. However, in the fields of autonomous driving, drone night flight, robot night work, etc., there is a strong practical demand and market demand for high-precision dark light image feature point extraction.

[0003] Currently, most existing technologies for processing dark-light images use various deep learning algorithms to brighten dark-light images, improve image contrast, and restore image colors. This method is mainly aimed at restoring the image as well as possible and obtaining a visually appealing image. However, in the current application field of image processing, we only need to find certain high-precision feature points in the image that can be tracked and easily compared, without requiring all the information of the entire image. Moreover, the algorithms for brightening dark-light images are relatively complex, and the training data sets mostly use images of relatively good quality, resulting in poor processing effects and reduced accuracy on the actual images. Therefore, computer vision technology still faces difficulties and challenges in processing dark-light images or images with low signal-to-noise ratios.

[0004] In summary, existing technologies are ineffective for processing images captured by cameras, and the algorithms are complex. Images captured by cameras are often low in brightness and noisy. A reliable approach to address this situation is image accumulation, which produces an effect somewhat similar to long exposure. However, due to the presence of strong random noise, the result of image stacking after long exposure alone is still blurry, so a simultaneous denoising and accumulation process is employed. Summary of the Invention

[0005] In response to the problems that the existing technologies have poor image processing effects on images actually captured by cameras and the algorithms are relatively complex, the present invention provides a high-precision and high-robustness feature point extraction method for dark-light images based on machine learning.

[0006] The present invention provides a method for extracting feature points from dark-light images based on machine learning, comprising:

[0007] S1. Remove background noise from N adjacent or similar frames of images;

[0008] S2, performing wavelet denoising on N frames of adjacent or similar images after background noise removal, and removing pixels below a set threshold to obtain N frames of adjacent or similar images after denoising;

[0009] S3, using the optical flow method to remove pixels whose optical flow variance exceeds the mean value of the optical flow variance matrix in N adjacent or similar frames after noise reduction, to obtain N adjacent or similar frames with large jitter removed;

[0010] S4. Using a moving Gaussian convolution algorithm to extract feature points from N frames of adjacent or similar images with large jitter removed, including:

[0011] S41, Gaussian convolution G1, G2, G3, ..., G N Act on N adjacent or similar images I1, I2, I3, ..., I N , get the result images IG1, IG2, IG3, ..., IG after N frames of Gaussian convolution N :

[0012]

[0013] S42. Calculate the mean of N frames of result images:

[0014]

[0015] S43. Establishing the objective function A deep learning framework is used for optimization. When the objective function converges, the N frames of result images tend to be consistent. These N frames of result images are then superimposed and synthesized using Gaussian weights to obtain a feature image.

[0016] As a preference, Gaussian convolution G1, G2, G3, ..., G N for:

[0017]

[0018] Among them, x j 、y j Represent the two-dimensional coordinates of the j-th convolution kernel, j = 1, 2, ..., N; (A1, A2, A3, ..., A N )、(δx1,δx2,δx3,...,δx N )、(δy1,δy2,δy3,...,δy N)、(Dx1,Dx2,Dx3,...,Dx N ) and (Dy1,Dy2,Dy3,...,Dy N ) is the variable to be optimized;

[0019] (A1,A2,A3,...,A N ) is used to adjust the overall amplitude of each frame image; (δx1,δx2,δx3,...,δx N )、(δy1,δy2,δy3,...,δy N ) are used to adjust the pixel movement distance of each frame image in the x and y directions; (Dx1, Dx2, Dx3, ..., Dx N )、(Dy1,Dy2,Dy3,...,Dy N ) are used to adjust the slope of Gaussian convolution in the x and y directions of each frame image respectively.

[0020] Preferably, S3 includes:

[0021] S31, assuming that N frames of adjacent or similar images I1, I2, I3, ..., I N Take any image I k , image I k Use dense optical flow method with the remaining N-1 frames, k∈{1,2,...,N}, to get the optical flow matrix O1,O2,...,O N-1 ; S32, calculate the optical flow matrix O1, O2, ..., O N-1 The variance matrix var(O1),var(O2),...,var(O N-1 ), and calculate the mean of each variance matrix mean(var(O1)),mean(var(O2)),...,mean(var(O N-1 ));

[0022] S33, extract the variance matrix var(O1),var(O2),...,var(O N-1 ) is greater than mean(var(O1)),mean(var(O2)),...,mean(var(O N-1 ))'s element coordinates, and the image grayscale value at the corresponding element coordinates in the N-1 frame image is taken as 0.

[0023] Preferably, in S1, a pulse coupled neural network is used to remove background noise in N adjacent or similar images.

[0024] This embodiment also provides a method for correcting the posture of an optical communication terminal, including:

[0025] Establish a regression equation for the difference between the azimuth and elevation angles at any two adjacent or similar moments;

[0026] Using the above-mentioned dark light image feature point extraction method based on machine learning, a total of M feature images extracted under adjacent or similar dark light conditions of the satellite optical communication terminal are obtained;

[0027] Comparing two feature images from the M adjacent or similar feature images extracted under dark light conditions, calculating pixel deviations in the x and y directions of the two images, substituting the pixel deviations into a regression equation for the difference between the azimuth and elevation angles, and obtaining constraint equations corresponding to the two feature images;

[0028] M feature images extracted under adjacent or similar dark light conditions are compared with each other to obtain constraint equations;

[0029] right The nonlinear least square method is applied to the constraint equation to obtain the optimal value of the azimuth error angle and complete the azimuth error correction.

[0030] Preferably, the optical communication terminal posture correction method further includes:

[0031] Establish a regression equation for the deviation between the standard position and the actual position;

[0032] Sending the M feature images extracted under adjacent or similar dark light conditions to a ground station;

[0033] The ground station obtains multiple adjacent or similar cloud images at the corresponding time, and processes the cloud images using the feature point extraction method for dark light images based on machine learning to obtain M feature images at the corresponding time as standard feature images;

[0034] The received M feature images extracted under adjacent or similar dark light conditions are matched and compared with the standard feature images at the corresponding moment, and the pixel deviation of the two compared feature images in the x and y directions is calculated. The pixel deviation is substituted into the regression equation of the deviation between the standard position and the actual position to obtain the corresponding constraint equation. A total of M constraint equations are obtained, and then the least squares method is used to obtain the zero-bit error vector to complete the zero-bit error correction.

[0035] The present invention proposes a moving Gaussian convolution method for extracting image feature points. This method processes multiple consecutive frames of images captured by a camera. It removes as much noise as possible while retaining common information between adjacent frames. It also removes points in the synthetic enhanced frame that may interfere with matching and correction, achieving high-precision and robust feature extraction. This allows the present invention to obtain stable feature point information even when applied to other low-light images. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a surface cloud map;

[0037] Figure 2 for Figure 1 Local dark light image;

[0038] Figure 3 To directly synthesize the effect of multi-frame surface cloud images;

[0039] Figure 4 Schematic diagram of moving Gaussian convolution;

[0040] Figure 5 for Figure 3 through Figure 4 The synthetic effect of multiple frames of surface cloud images after moving Gaussian convolution processing;

[0041] Figure 6 This is the result of synthesizing multiple frames of surface cloud images after optical flow method and moving Gaussian convolution processing;

[0042] Figure 7 are 128 characteristic highlights in wavelet;

[0043] Figure 8 This is a diagram of the standard position and the actual position;

[0044] Figure 9 This is a normal light image of the lunar surface;

[0045] Figure 10 for Figure 9 Local dark light image;

[0046] Figure 11 This is the result of synthesizing multiple frames of lunar surface images after optical flow method and moving Gaussian convolution processing. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0048] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0049] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0050] The method for extracting feature points from a dark-light image based on machine learning in this embodiment includes:

[0051] Step 1: Remove background noise, such as dark current noise and thermal noise, from N adjacent or similar frames of images. This can be achieved using a pulse coupled neural network (PCNN).

[0052] Step 2: performing wavelet denoising on the N frames of adjacent or similar images after background noise removal, and removing pixels below a set threshold to obtain N frames of adjacent or similar images after denoising;

[0053] Step 3: Use the optical flow method to remove pixels whose optical flow variance exceeds the mean value of the optical flow variance matrix in N adjacent or similar frames after noise reduction, and obtain N adjacent or similar frames with large jitter removed. Specifically:

[0054] Step 31: Assume that N frames of adjacent or similar images I1, I2, I3, ..., I N Take any image I k (You can also use multiple images to improve accuracy, but it requires multiple times the computing power), and image I k Use dense optical flow method with the remaining N-1 frames, k∈{1,2,...,N}, to get the optical flow matrix O1,O2,...,O N-1 ;

[0055] Step 32: Calculate the optical flow matrix O1, O2, ..., O N-1 The variance matrix var(O1),var(O2),...,var(O N-1 ), and calculate the mean of each variance matrix mean(var(O1)),mean(var(O2)),...,mean(var(O N-1 ));

[0056] Step 33, extract the variance matrix var(O1),var(O2),...,var(O N-1 ) is greater than mean(var(O1)),mean(var(O2)),...,mean(var(O N-1 ))'s element coordinates, and the image grayscale value at the corresponding element coordinates in the N-1 frame image is taken as 0.

[0057] Step 4: After removing pixels with large jitter caused by motion, etc. using the optical flow method, the moving Gaussian convolution algorithm is used to extract feature points from N adjacent or similar frames of images without large jitter to reduce the blurring effect after image superposition. Specifically, the following steps are performed:

[0058] Step 41, Gaussian convolution G1, G2, G3, ..., GN Act on N adjacent or similar image slices I1, I2, I3, ..., I N , get the result images IG1, IG2, IG3, ..., IG after N frames of Gaussian convolution N :

[0059]

[0060] Step 42: Calculate the mean of the N frames of result images:

[0061]

[0062] Step 43: Establish the objective function A deep learning framework is used for optimization. When the objective function converges, the N frames of result images tend to be consistent. These N frames of result images are then superimposed and synthesized using Gaussian weights to obtain a feature image.

[0063] In this embodiment, Gaussian convolution G1, G2, G3, ..., G N for:

[0064]

[0065] Among them, x j 、y j Represent the two-dimensional coordinates of the j-th convolution kernel, j = 1, ... N; (A1, A2, A3, ..., A N )、(δx1,δx2,δx3,...,δx N )、(δy1,δy2,δy3,...,δy N )、(Dx1,Dx2,Dx3,...,Dx N ) and (Dy1,Dy2,Dy3,...,Dy N ) is the variable to be optimized;

[0066] (A1,A2,A3,...,A N ) is used to adjust the overall amplitude of each frame image to adjust the possible light intensity flicker; (δx1,δx2,δx3,...,δx N )、(δy1,δy2,δy3,...,δy N ) are used to adjust the pixel movement distance of each frame image in the two-dimensional direction of the convolution kernel; (Dx1, Dx2, Dx3, ..., Dx N )、(Dy1,Dy2,Dy3,...,Dy N ) are used to adjust the slope of the Gaussian convolution of each frame image in the two-dimensional direction of the convolution kernel. If a sharper image is required, a smaller value should be used.

[0067] Embodiment 1: A method for correcting the posture of an optical communication terminal, comprising:

[0068] The optical communication terminal posture correction is performed using a multi-frame image synthesis image. According to the current correction method, it is assumed that the accurate direction vector of the current terminal is However, due to the installation azimuth error angle (α, β, γ) T The existence of , causes the actual direction vector to become And the relationship between the two satisfies Correspondingly, the azimuth and pitch angle vectors with deviations are Among them, (δAz, δEl) T is the zero error vector of the Az and El axes. In the process of applying machine learning to approximate the error, It is known that the azimuth error angle (α, β, γ) T and the zero error vector (δAz, δEl) T is the error to be optimized. The following corrections are made to the azimuth error angle and the zero position error vector.

[0069] According to the direction vector and azimuth pitch angle vector of any two adjacent or similar times t1 and t2:

[0070] Time t1: direction vector Azimuth and elevation angles

[0071] Time t2: direction vector Azimuth and elevation angles

[0072] Calculate the difference between the azimuth and elevation angles at two moments:

[0073]

[0074] During the calibration process, the method of this embodiment is used to obtain a total of M adjacent or similar characteristic images of the satellite optical communication terminal under low-light conditions. Assuming that 2000 frames of images can be extracted from the video, N in this embodiment refers to 64 frames or another number of frames. After processing N frames, they are superimposed to obtain a composite image, i.e., a characteristic image. Subsequently, the first frame of these N frames is removed, leaving N-1 frames. The next frame is read from the video and added to the end of the N-1 frame, still N frames. The extraction method of this embodiment is applied to these frames to obtain a composite image. This process is repeated in this way, and a characteristic image is synthesized after every N frames are processed. Ultimately, after all images in the video are processed, M characteristic images are obtained.

[0075] by Figure 6As shown in the figure, the difference between the azimuth and elevation angles in two adjacent or similar feature images under dark light conditions is calculated. From the two adjacent composite images, the pixel deviations in the x and y directions of the feature images are 1.216 and 0.834 respectively. From this, the following regression equation can be obtained:

[0076]

[0077] Assume that a total of M adjacent or similar effective synthetic images are obtained during a calibration process, and then a pairwise comparison can be performed to obtain a total of A constraint equation of the form above; By applying the nonlinear least squares method to the constraint equations, (α, β, γ) can be calculated on the satellite. T Calculate the pixel deviation in the x and y directions of the two feature images under dark light conditions, and substitute the obtained pixel deviation into the above regression equation as the constraint equation corresponding to the two feature images;

[0078] right The nonlinear least squares method is applied to the constraint equation to obtain the optimal value of the azimuth error angle and complete the azimuth error correction. The above completes the onboard terminal attitude correction;

[0079] Next, we use the synthetic image to extract feature points and transmit them to realize zero-bit error correction, including:

[0080] Matching adjacent or similar synthetic images on a satellite can only achieve relative error correction, while the zero error is an absolute error, which requires an absolute reference or finding the corresponding image in an absolute coordinate system to calculate the absolute error. Figure 6 It can be seen that there are still more than 100,000 pixel values, and it is unrealistic to download the image. During the attitude correction of the satellite terminal, M feature images extracted under adjacent or similar dark light conditions are obtained. 128 feature highlights are extracted from each feature image, such as Figure 7 Shown are the feature maps of two adjacent or similar images, representing Figure 6 characteristics.

[0081] The 128 points have 128 coordinates, a total of 256 values, which can be transmitted back to the ground station via telemetry. Based on these returned coordinates, the following can be regenerated: Figure 7 Similar feature point images are shown.

[0082] At the ground station, multiple adjacent or similar cloud images at the corresponding time are obtained, and dimmed. At the same time, the cloud images are processed using the feature point extraction method of this embodiment to obtain M feature images at the corresponding time as standard feature images;

[0083] Match and compare the received M feature images extracted under adjacent or similar dark light conditions with the standard feature image at the corresponding moment, and calculate the pixel deviation (dx, dy) of the two compared feature images in the x and y directions. T .like Figure 8 As shown, it is assumed that there is a deviation (dx, dy) between the standard position and the actual position at a certain moment T , thus obtaining the regression equation of the deviation between the standard position and the actual position:

[0084]

[0085] The pixel deviation (dx, dy) T By bringing it into the regression equation of the deviation between the standard position and the actual position, after M times of similar sampling and processing, M constraint equations can be obtained, and the parameters to be optimized (δAz, δEl) can be obtained using the least squares method. T .

[0086] Example 2:

[0087] Satellite-photographed surface cloud images Figure 1 and Figure 2 As shown, Figure 1 is a normal light image, Figure 2 The completely black image is a dark light image taken by a simulated satellite under dark light conditions. Figure 1 yes Figure 2 The comparison diagram shows that the image processed by this embodiment is Figure 1 A small area in . Figure 2 is a dark light image captured from a video, and the method of this embodiment is used to Figure 2 The feature point extraction process is performed on the continuous multi-frame images under this extremely dark condition. 64 images are obtained by sampling, and after denoising, the multi-frame image synthesis effect is directly performed as shown in the figure. Figure 3 As shown in the figure, the image is very blurry, which directly affects the accuracy and reliability of matching and correction based on the image. In order to suppress the blur effect, moving Gaussian convolution is used to process the image. Figure 4 The convolution kernel is a schematic diagram. The specific convolution kernel is obtained through optimization of the deep learning algorithm. After a series of processing such as denoising and moving Gaussian convolution, the result of synthesizing 64 images is shown in the figure below. Figure 5 As shown, Figure 3 The comparison shows that the image noise is reduced and the blur effect is improved.

[0088] The blur caused by large pixel jumps due to camera rotation and other reasons still exists in the image, so the optical flow method is used to remove pixels with large jumps before performing moving Gaussian convolution. Figure 6The image shown is the result after a series of processing, including image denoising, optical flow method and moving Gaussian convolution followed by image synthesis. Figure 3 、 Figure 5 In comparison, the noise removal effect is better and the blurring effect is suppressed, resulting in a fairly stable pixel cluster retained in the image.

[0089] Example 3:

[0090] The above algorithm can also be used to extract image feature points from the near-moon surface images. Figure 9 This is a normal light lunar surface image. Figure 10 The lunar surface image is taken under dark light conditions. After a series of operations such as image denoising, optical flow method to remove large noise points, and moving Gaussian convolution to suppress image blur, the following is obtained. Figure 11 Compared with the original dark light image, the image feature points obtained by the algorithm have higher clarity, improved stability and higher accuracy.

[0091] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the invention. It should be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be employed in conjunction with other described embodiments.

Claims

1. A feature point extraction method for dark light images based on machine learning, characterized in that: The method comprises: S1. Remove background noise from N adjacent or similar frames of images; S2, performing wavelet denoising on N frames of adjacent or similar images after background noise removal, and removing pixels below a set threshold to obtain N frames of adjacent or similar images after denoising; S3, using the optical flow method to remove pixels whose optical flow variance exceeds the mean value of the optical flow variance matrix in N adjacent or similar frames after noise reduction, to obtain N adjacent or similar frames with large jitter removed; S4. Using a moving Gaussian convolution algorithm to extract feature points from N frames of adjacent or similar images with large jitter removed, including: S41, Gaussian convolution G1, G2, G3, ..., G N Act on N adjacent or similar images I1, I2, I3, ..., I N , get the result images IG1, IG2, IG3, ..., IG after N frames of Gaussian convolution N : S42. Calculate the mean of N frames of result images: S43. Establishing the objective function A deep learning framework is used for optimization. When the objective function converges, the N frames of result images tend to be consistent. These N frames of result images are then superimposed and synthesized using Gaussian weights to obtain a feature image.

2. The method for extracting feature points from dark-light images based on machine learning according to claim 1, wherein: Gaussian convolution G1, G2, G3, ..., G N for: Among them, x j 、y j Represent the two-dimensional coordinates of the j-th convolution kernel, j = 1, 2, ..., N; (A1, A2, A3, ..., A N )、(δx1,δx2,δx3,...,δx N )、(δy1,δy2,δy3,...,δy N )、(Dx1,Dx2,Dx3,...,Dx N ) and (Dy1,Dy2,Dy3,...,Dy N ) is the variable to be optimized; (A1,A2,A3,...,A N ) is used to adjust the overall amplitude of each frame image; (δx1,δx2,δx3,...,δx N )、(δy1,δy2,δy3,...,δy N ) are used to adjust the pixel movement distance of each frame image in the x and y directions; (Dx1, Dx2, Dx3, ..., Dx N )、(Dy1,Dy2,Dy3,...,Dy N ) are used to adjust the slope of Gaussian convolution in the x and y directions of each frame image respectively.

3. The method for extracting feature points from dark-light images based on machine learning according to claim 1, wherein: The S3 includes: S31, assuming that N frames of adjacent or similar images I1, I2, I3, ..., I N Take any image I k , image I k Use dense optical flow method with the remaining N-1 frames, k∈{1,2,...,N}, to get the optical flow matrix O1,O2,...,O N-1 ; S32, calculate the optical flow matrix O1, O2, ..., O N-1 The variance matrix var(O1),var(O2),...,var(O N-1 ), and calculate the mean of each variance matrix mean(var(O1)),mean(var(O2)),...,mean(var(O N-1 )); S33, extract the variance matrix var(O1),var(O2),...,var(O N-1 ) is greater than mean(var(O1)),mean(var(O2)),...,mean(var(O N-1 ))'s element coordinates, and the image grayscale value at the corresponding element coordinates in the N-1 frame image is taken as 0.

4. The method for extracting feature points from dark-light images based on machine learning according to claim 1, wherein: In S1, a pulse coupled neural network is used to remove background noise from N adjacent or similar images.

5. A computer-readable storage device storing a computer program, characterized in that: When the computer program is executed, the method for extracting feature points from dark-light images based on machine learning as claimed in any one of claims 1 to 4 is implemented.

6. A device for extracting feature points from dark-light images based on machine learning, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor, wherein: The processor executes the computer program to implement the method for extracting feature points from dark-light images based on machine learning as described in any one of claims 1 to 4.

7. A method for correcting the posture of an optical communication terminal, characterized in that: The method comprises: Establish a regression equation for the difference between the azimuth and elevation angles at any two adjacent or similar moments; Using the method for extracting feature points of dark light images based on machine learning according to any one of claims 1 to 4, a total of M feature images extracted under adjacent or similar dark light conditions of the satellite optical communication terminal are obtained; Comparing two feature images from the M adjacent or similar feature images extracted under dark light conditions, calculating pixel deviations in the x and y directions of the two images, substituting the pixel deviations into a regression equation for the difference between the azimuth and elevation angles, and obtaining constraint equations corresponding to the two feature images; M feature images extracted under adjacent or similar dark light conditions are compared with each other to obtain constraint equations; right The nonlinear least square method is applied to the constraint equation to obtain the optimal value of the azimuth error angle and complete the azimuth error correction.

8. The method for correcting the posture of an optical communication terminal according to claim 7, wherein: The method further comprises: Establish a regression equation for the deviation between the standard position and the actual position; Sending the M feature images extracted under adjacent or similar dark light conditions to a ground station; The ground station obtains multiple adjacent or similar frames of cloud images at corresponding moments, and simultaneously processes the cloud images using the feature point extraction method for dark light images based on machine learning as described in any one of claims 1 to 4 to obtain M feature images at corresponding moments as standard feature images; The received M feature images extracted under adjacent or similar dark light conditions are matched and compared with the standard feature images at the corresponding moment, and the pixel deviation of the two compared feature images in the x and y directions is calculated. The pixel deviation is substituted into the regression equation of the deviation between the standard position and the actual position to obtain the corresponding constraint equation. A total of M constraint equations are obtained, and then the least squares method is used to obtain the zero-bit error vector to complete the zero-bit error correction.

Images