A video object denoising method and system based on contextual spatial information

Abstract

The present invention discloses a method and system for denoising a video target based on contextual spatial information, the method comprising: fusing temporally adjacent video images to obtain a fused image; based on temporally adjacent video images, determining the motion state of the target and constructing an attribute matrix of the fused image; extracting convolution features from the fused image and splicing it with the attribute matrix to obtain an enhanced feature matrix; constructing a cost function of an autoencoder, setting a number of blind spots on the enhanced feature matrix and inputting it into the autoencoder for training until the cost function converges to a preset value; inputting the enhanced feature matrix into the trained autoencoder, reconstructing the enhanced feature matrix, and obtaining a denoised feature matrix; intercepting and transposing the denoised feature matrix to obtain a denoised video image. The present invention adopts video image fusion and autoencoder, and obtains a high-quality denoising model through two-stage denoising processing to improve the denoising performance of video targets.

Description

Technical Field

[0001] The present invention relates to the technical field of video image denoising, and in particular to a video target denoising method and system based on contextual spatial information. Background Art

[0002] When collecting information about a target, background noise is inevitably introduced into the video image. Noise is also introduced into the video image encoding, transmission, and display process. The noise in the video image will have an impact in many aspects. For example, noise will reduce the image clarity, distort the image color, and affect the visual effect of the image. At the same time, the presence of noise will also interfere with video target detection and target motion tracking, thereby increasing the difficulty of image processing. Therefore, video image denoising is a very important link in the field of image processing.

[0003] Traditional video image denoising methods mainly use filtering methods to process pixels, such as using mean filters, median filters and Gaussian filters to smooth pixels to remove noise. Although these methods can achieve the purpose of denoising, they will inevitably cause image blur and lose image fine details and texture information during denoising, which brings new problems. With the further development of technology, the commonly used video image denoising method is to use deep learning methods for denoising. Although the existing methods can denoise video images to a certain extent, the denoising effect is not ideal when the video image quality is poor and the target is blocked. Summary of the invention

[0004] In view of this, the present invention provides a video target denoising method and system based on contextual spatial information, the purpose of which is to: based on temporally adjacent video images, make full use of the correlation of contextual spatial information, enhance image noise characteristics, and achieve high-quality denoising of video images.

[0005] In order to achieve the above object, the present invention discloses a video object denoising method based on contextual spatial information, comprising the following steps:

[0006] S1: Fusing temporally adjacent video images to obtain a fused image;

[0007] S2: Based on the temporally adjacent video images, determine the motion state of the target and construct the attribute matrix of the fused image;

[0008] S3: extract convolution features from the fused image and concatenate them with the attribute matrix to obtain an enhanced feature matrix;

[0009] S4: Using the enhanced feature matrix, train the autoencoder, including:

[0010] S41: Randomly initialize the weights and biases of the encoder and decoder in the autoencoder;

[0011] S42: After setting several blind spots on the enhanced feature matrix, the enhanced feature matrix is ​​input into the autoencoder. Based on the enhanced feature matrix and the output result of the autoencoder, the cost function of the autoencoder is constructed:

[0012] ;

[0013] in, represents the cost function, represents the norm function, represents the enhanced feature matrix, Represents the output result of the autoencoder;

[0014] S43: Calculate the gradient of the cost function with respect to the weights and biases of the encoder and decoder, and update the weights and biases of the encoder and decoder using the gradient descent method;

[0015] S44: repeating step S42 and step S43 to train the autoencoder until the cost function of the autoencoder converges to a preset value, thereby obtaining a trained autoencoder;

[0016] S5: input the enhanced feature matrix into the trained autoencoder, reconstruct the enhanced feature matrix, and obtain the denoised feature matrix;

[0017] S6: Performing a truncation and transposition convolution transformation on the denoised feature matrix to obtain a denoised video image.

[0018] Preferably, the step S1 comprises:

[0019] S11: using a wavelet decomposition method, decomposing the video images of the t-th frame and the t+1-th frame respectively to obtain a high-frequency residual image;

[0020] Using the optical flow method, the video images of the t-th frame and the t+1-th frame are registered according to the high-frequency residual image;

[0021] Estimate the pixel displacement of the video image of the tth frame and the t+1th frame to obtain the displacement map ;

[0022] S12: Reconstruct the Laplacian pyramid of the t-th frame video image:

[0023] ;

[0024] in, represents the reconstructed Laplacian pyramid of the t-th frame video image, Represents the original Laplacian pyramid of the t-th frame video image, represents the Laplace function;

[0025] S13: Reconstruct the Laplacian pyramid of the t+1th frame video image:

[0026] ;

[0027] in, represents the reconstructed Laplacian pyramid of the t+1th frame video image, Represents the original Laplacian pyramid of the t+1th frame video image, represents the weight factor;

[0028] S14: Exploitation , reconstruct the t+1th frame of video image to obtain a fused image.

[0029] Preferably, the wavelet decomposition method in step S11 may be Gaussian decomposition or Fourier decomposition.

[0030] Preferably, the step S13 comprises:

[0031] S131: For all pixels of the video image of the t+1th frame, calculate the standard deviation of the RGB value relative to the brightness value ;

[0032] S132: Utilization , reconstruct the t-th frame of the video image, and calculate the standard deviation of the RGB value relative to the brightness value of all pixels in the reconstructed image ;

[0033] S133: Standard deviation and Normalized, that is, the weight factor .

[0034] Preferably, step S2 comprises:

[0035] S21: determining the motion state of the target according to the video images of the t-th frame and the t+1-th frame;

[0036] S22: Construct the attribute matrix of the fused image:

[0037] ;

[0038] in, represents the attribute matrix, represents the target change rate, Indicates that the target is forward motion, Indicates that the target is moving backward. Indicates that the target is stationary motion.

[0039] Preferably, the step S21 includes:

[0040] If the size of the video object in the tth frame and the t+1th frame is the same, the object is judged to be stationary motion;

[0041] If the size of the target in the video frame t+1 is larger than that in the video frame t, the target is judged to be moving forward;

[0042] If the size of the object in the t-th frame video is larger than that in the t+1-th frame video, the object is determined to be moving backward.

[0043] Preferably, the step S22 includes:

[0044] In the video images of the t-th frame and the t+1-th frame, the target change rate is calculated according to the size change of the target:

[0045] ;

[0046] in, represents the target change rate, Indicates the size of the target in the t+1th frame of the video image, Indicates the size of the target in the t-th frame of the video image.

[0047] Preferably, step S3 comprises:

[0048] S31: Use the Sobel-x convolution kernel to perform two-dimensional convolution on the fused image to obtain a two-dimensional matrix, and add a bias value to each element of the two-dimensional matrix After that, input the activation function in turn , obtain a one-dimensional matrix with nonlinear representation ;

[0049] Use the Sobel-y convolution kernel to perform two-dimensional convolution on the fused image to obtain a two-dimensional matrix, and add a bias value to each element of the two-dimensional matrix After that, input the activation function in turn , obtain a one-dimensional matrix with nonlinear representation ;

[0050] S32: Convert one-dimensional matrix , and the attribute matrix obtained in step S2 Stitched together to obtain an enhanced feature matrix .

[0051] Preferably, the offset value in step S31 is and All are determined through training, the specific process includes:

[0052] Construct single-layer convolutional networks based on Sobel-x convolution kernel and Sobel-y convolution kernel respectively, and randomly initialize all weights and bias functions;

[0053] Construct a dataset of fused images, train two single-layer convolutional networks on the dataset, and obtain the cost function;

[0054] Based on the step size, the bias value is continuously adjusted within the bias value range until the optimal cost function is obtained. At this time, the corresponding bias value is and .

[0055] Preferably, the step S31 further includes:

[0056] Add a bias value to each element of a two-dimensional matrix or After that, distributed in When the interval is , the activation function uses the following function:

[0057] ;

[0058] in, Represents the elements of a two-dimensional matrix to add bias values or The value after

[0059] Add a bias value to each element of a two-dimensional matrix or After that, distributed in When the interval is , the activation function uses the following function:

[0060] ;

[0061] in, Represents the elements of a two-dimensional matrix to add bias values or The value after.

[0062] Preferably, step S5 comprises:

[0063] The matrix Input the trained autoencoder, and the encoder converts the matrix Compressed to a low-dimensional space, the decoder restores the low-dimensional space through the inverse process to obtain the same matrix Output of the same dimension, label the output as the denoised feature matrix .

[0064] Preferably, step S6 comprises:

[0065] S61: Denoised feature matrix Before interception elements, and get the matrix , represents the one-dimensional matrix obtained in step S31 Length;

[0066] S62: Matrix The transposed convolution transform is used to obtain the denoised video image.

[0067] The present invention also discloses a video target denoising system based on contextual spatial information, comprising:

[0068] Image fusion module: Decompose temporally adjacent video images to obtain high-frequency residual images, align temporally adjacent video images according to the high-frequency residual images, reconstruct the Laplacian pyramids of temporally adjacent video images in sequence, and obtain a fused image based on image reconstruction;

[0069] Attribute matrix construction module: Based on the temporally adjacent video images, the target motion state is determined and the attribute matrix of the fused image is constructed;

[0070] Feature matrix construction module: perform convolution feature extraction and nonlinear matrix calculation on the fused image, and concatenate the obtained one-dimensional matrix with the attribute matrix to obtain an enhanced feature matrix;

[0071] Autoencoder training module: construct the cost function of the autoencoder, set blind spots for the feature matrix and input it into the autoencoder for training;

[0072] Feature matrix reconstruction module: Use the trained autoencoder to reconstruct the enhanced feature matrix to obtain the denoised matrix;

[0073] Feature matrix conversion module: The denoised matrix is ​​intercepted and transposed by convolution transformation to obtain the denoised video image.

[0074] The above-mentioned video target denoising method based on contextual spatial information is implemented through this system.

[0075] Beneficial Effects

[0076] The method of the present invention adopts image fusion and autoencoder to realize two-stage denoising and improve the denoising quality of video images; the autoencoder is trained based on blind spot setting, and a high-quality denoising model is obtained through training in the absence of clean target images and noise images, which is beneficial to improving the denoising ability of video targets; the fusion of image feature enhancement and learning is beneficial to learning the characteristics of weak noise; the convolution feature extraction method, the bias value and the activation function determination method are beneficial to improving the network fitting ability; the cost function of the autoencoder is designed according to the autoencoder and the expected output result, which is beneficial to obtaining a better training effect.

[0077] At present, the utilization of contextual spatial information of video images mainly focuses on using the continuous movement of the target for target recognition. Considering that the context of the video image is not only continuous in time but also correlated in noise information, the method of the present invention makes full use of the correlation of contextual spatial information to fuse adjacent video images, which is beneficial to improving the video image denoising capability, specifically including: removing larger noise through image fusion, superimposing smaller noise, improving the noise feature extraction capability, obtaining significant features of smaller noise through feature processing, and facilitating the training of denoising models. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 A flowchart of a method for video object denoising based on contextual spatial information provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0079] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

[0080] Embodiment 1:

[0081] A video object denoising method based on contextual spatial information, such as Figure 1 As shown, the following steps are included:

[0082] S1: Fusing temporally adjacent video images to obtain a fused image;

[0083] S2: Based on the temporally adjacent video images, determine the motion state of the target and construct the attribute matrix of the fused image;

[0084] S3: extract convolution features from the fused image and concatenate them with the attribute matrix to obtain an enhanced feature matrix;

[0085] S4: Train the autoencoder using the enhanced feature matrix;

[0086] S5: input the enhanced feature matrix into the trained autoencoder, reconstruct the enhanced feature matrix, and obtain the denoised feature matrix;

[0087] S6: Performing a truncation and transposition convolution transformation on the denoised feature matrix to obtain a denoised video image.

[0088] The step S1 comprises:

[0089] S11: using a wavelet decomposition method, specifically, Gaussian decomposition or Fourier decomposition, to decompose the video images of the t-th frame and the t+1-th frame respectively to obtain a high-frequency residual image;

[0090] The optical flow method is used to register the video images of the t-th frame and the t+1-th frame according to the high-frequency residual image. Specifically, pixels with significant features in the high-frequency residual image are selected, including pixels containing targets, pixels located at the edge of the image, and pixels with special amplitudes. The position changes of the pixels in the video images of the t-th frame and the t+1-th frame are used to register the video images of the t-th frame and the t+1-th frame.

[0091] Estimate the pixel displacement of the video image of the tth frame and the t+1th frame to obtain the displacement map ;

[0092] S12: Reconstruct the Laplacian pyramid of the t-th frame video image:

[0093] ;

[0094] in, represents the reconstructed Laplacian pyramid of the t-th frame video image, Represents the original Laplacian pyramid of the t-th frame video image, represents the Laplace function;

[0095] S13: Reconstruct the Laplacian pyramid of the t+1th frame video image:

[0096] ;

[0097] in, represents the reconstructed Laplacian pyramid of the t+1th frame video image, Represents the original Laplacian pyramid of the t+1th frame video image, represents the weight factor;

[0098] S14: Exploitation , reconstruct the t+1th frame of video image to obtain a fused image.

[0099] The step S13 further includes:

[0100] S131: For all pixels of the video image of the t+1th frame, calculate the standard deviation of the RGB value relative to the brightness value ;

[0101] S132: Utilization , reconstruct the t-th frame of the video image, and calculate the standard deviation of the RGB value relative to the brightness value of all pixels in the reconstructed image ;

[0102] S133: Standard deviation and Normalized, that is, the weight factor .

[0103] The step S2 comprises:

[0104] S21: Determine the motion state of the target according to the video images of the t-th frame and the t+1-th frame, specifically:

[0105] If the size of the video object in the tth frame and the t+1th frame is the same, the object is judged to be stationary motion;

[0106] If the size of the target in the video frame t+1 is larger than that in the video frame t, the target is judged to be moving forward;

[0107] If the size of the object in the t-th frame video is larger than that in the t+1-th frame video, the object is judged to be moving backward;

[0108] S22: Construct the attribute matrix of the fused image:

[0109] ;

[0110] in, represents the attribute matrix, represents the target change rate, Indicates that the target is forward motion, Indicates that the target is moving backward. Indicates that the target is stationary motion.

[0111] In the video images of the tth frame and the t+1th frame, the target change rate is calculated according to the size change of the target :

[0112] ;

[0113] in, Indicates the size of the target in the t+1th frame of the video image, Indicates the size of the target in the t-th frame of the video image.

[0114] In this embodiment, the target is a target with significant features, including a target with a clear boundary, a target located in the foreground of the video image, and a target with a large contrast with the background; the target size measurement method is based on the longest line segment in the target diagonal.

[0115] The step S3 comprises:

[0116] S31: Use the Sobel-x convolution kernel to perform two-dimensional convolution on the fused image to obtain a two-dimensional matrix, and add a bias value to each element of the two-dimensional matrix After that, input the activation function in turn , obtain a one-dimensional matrix with nonlinear representation ;

[0117] The input activation function The specific process is:

[0118] Add a bias value to each element of a two-dimensional matrix After that, if the distribution is When the interval is , the activation function uses the following function:

[0119] ;

[0120] If distributed in When the interval is , the activation function uses the following function:

[0121] ;

[0122] in, Represents the elements of a two-dimensional matrix to add bias values The value after.

[0123] Use the Sobel-y convolution kernel to perform two-dimensional convolution on the fused image to obtain a two-dimensional matrix, and add a bias value to each element of the two-dimensional matrix After that, input the activation function in turn , obtain a one-dimensional matrix with nonlinear representation ;

[0124] The input activation function The specific process is:

[0125] Add a bias value to each element of a two-dimensional matrix After that, if the distribution is interval, the activation function uses the following function:

[0126] ;

[0127] If distributed in interval, the activation function uses the following function:

[0128] ;

[0129] in, Represents the elements of a two-dimensional matrix to add bias values The value after

[0130] Bias value and All are determined through training, the specific process is:

[0131] Construct single-layer convolutional networks based on Sobel-x convolution kernel and Sobel-y convolution kernel respectively, and randomly initialize all weights and bias functions;

[0132] Construct a dataset of fused images, train two single-layer convolutional networks on the dataset, and obtain the cost function;

[0133] Based on the step size, the bias value is continuously adjusted within the bias value range until the optimal cost function is obtained. At this time, the corresponding bias value is and ;

[0134] S32: Convert one-dimensional matrix , and the attribute matrix obtained in step S2 Stitched together to obtain an enhanced feature matrix .

[0135] The step S4 comprises:

[0136] S41: Randomly initialize the weights and biases of the encoder and decoder in the autoencoder;

[0137] S42: Pair Matrix set up Blind spots, specifically, for the matrix Random Selection elements, and set each element value to 0, and get the matrix , the matrix Input the autoencoder and mark the output of the autoencoder as ;

[0138] Construct the cost function of the autoencoder:

[0139] ;

[0140] in, represents the cost function, represents the norm function;

[0141] S43: Calculate the gradient of the cost function with respect to the weights and biases of the encoder and decoder, and update the weights and biases of the encoder and decoder using the gradient descent method;

[0142] S44: Repeat step S42 and step S43 to train the autoencoder until the cost function of the autoencoder converges to a preset value, thereby obtaining a trained autoencoder.

[0143] In this embodiment, Much smaller than the matrix The number of elements in is randomly set each time Through multiple experiments, we can get multiple matrices with blind spots. , constitute a data set, on which the autoencoder is trained based on the cost function.

[0144] The step S5 comprises:

[0145] The matrix Input the trained autoencoder and solve the output of the autoencoder based on the minimum cost function condition. Specifically, the encoder converts the matrix Compressed to a low-dimensional space, the decoder restores the low-dimensional space through the inverse process to obtain the same matrix Output of the same dimension, label the output as the denoised feature matrix .

[0146] The step S6 comprises:

[0147] S61: Denoised feature matrix Before interception elements, and get the matrix , represents the one-dimensional matrix obtained in step S31 Length;

[0148] S62: Matrix The transposed convolution transform is used to obtain the denoised video image.

[0149] In this embodiment, the matrix A transposed convolution transformation is adopted, and the bias value and activation function adopted refer to the bias value and activation function in step S31, and are fine-tuned.

[0150] Embodiment 2:

[0151] The present invention also provides a video target denoising system based on contextual spatial information, comprising:

[0152] Image fusion module: Decompose temporally adjacent video images to obtain high-frequency residual images, align temporally adjacent video images according to the high-frequency residual images, reconstruct the Laplacian pyramids of temporally adjacent video images in sequence, and obtain a fused image based on image reconstruction;

[0153] Attribute matrix construction module: Based on the temporally adjacent video images, the target motion state is determined and the attribute matrix of the fused image is constructed;

[0154] Feature matrix construction module: perform convolution feature extraction and nonlinear matrix calculation on the fused image, and concatenate the obtained one-dimensional matrix with the attribute matrix to obtain an enhanced feature matrix;

[0155] Autoencoder training module: construct the cost function of the autoencoder, set blind spots for the feature matrix and input it into the autoencoder for training;

[0156] Feature matrix reconstruction module: Use the trained autoencoder to reconstruct the enhanced feature matrix to obtain the denoised matrix;

[0157] Feature matrix conversion module: The denoised matrix is ​​intercepted and transposed by convolution transformation to obtain the denoised video image.

[0158] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0159] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0160] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A video object denoising method based on contextual spatial information, characterized in that: The following steps are involved: S1: Fusing temporally adjacent video images to obtain a fused image; S2: Based on the temporally adjacent video images, determine the motion state of the target and construct the attribute matrix of the fused image; S3: extract convolution features from the fused image and concatenate them with the attribute matrix to obtain an enhanced feature matrix; S4: Using the enhanced feature matrix, train the autoencoder, including: S41: Randomly initialize the weights and biases of the encoder and decoder in the autoencoder; S42: After setting several blind spots on the enhanced feature matrix, the enhanced feature matrix is ​​input into the autoencoder. Based on the enhanced feature matrix and the output result of the autoencoder, the cost function of the autoencoder is constructed: ； in, represents the cost function, represents the norm function, represents the enhanced feature matrix, Represents the output result of the autoencoder; S43: Calculate the gradient of the cost function with respect to the weights and biases of the encoder and decoder, and update the weights and biases of the encoder and decoder using the gradient descent method; S44: repeating step S42 and step S43 to train the autoencoder until the cost function of the autoencoder converges to a preset value, thereby obtaining a trained autoencoder; S5: input the enhanced feature matrix into the trained autoencoder, reconstruct the enhanced feature matrix, and obtain the denoised feature matrix; S6: Performing a truncation and transposition convolution transformation on the denoised feature matrix to obtain a denoised video image.

2. The video object denoising method based on contextual spatial information according to claim 1, characterized in that: The step S1 comprises: S11: using a wavelet decomposition method, decomposing the video images of the t-th frame and the t+1-th frame respectively to obtain a high-frequency residual image; Using the optical flow method, the video images of the t-th frame and the t+1-th frame are registered according to the high-frequency residual image; Estimate the pixel displacement of the video image of the tth frame and the t+1th frame to obtain the displacement map ; S12: Reconstruct the Laplacian pyramid of the t-th frame video image: ； in, represents the reconstructed Laplacian pyramid of the t-th frame video image, Represents the original Laplacian pyramid of the t-th frame video image, represents the Laplace function; S13: Reconstruct the Laplacian pyramid of the t+1th frame video image: ； in, represents the reconstructed Laplacian pyramid of the t+1th frame video image, Represents the original Laplacian pyramid of the t+1th frame video image, represents the weight factor; S14: Exploitation , reconstruct the t+1th frame of video image to obtain a fused image.

3. The video object denoising method based on contextual spatial information according to claim 2, characterized in that: The step S2 comprises: S21: determining the motion state of the target according to the video images of the t-th frame and the t+1-th frame; S22: Construct the attribute matrix of the fused image: ； in, represents the attribute matrix, represents the target change rate, Indicates that the target is forward motion, Indicates that the target is moving backward. Indicates that the target is stationary motion.

4. The video object denoising method based on contextual spatial information according to claim 3, characterized in that: The step S22 comprises: In the video images of the t-th frame and the t+1-th frame, the target change rate is calculated according to the size change of the target: ； in, represents the target change rate, Indicates the size of the target in the t+1th frame of the video image, Indicates the size of the target in the t-th frame of the video image.

5. The video object denoising method based on contextual spatial information according to claim 4, characterized in that: The step S3 comprises: S31: Use the Sobel-x convolution kernel to perform two-dimensional convolution on the fused image to obtain a two-dimensional matrix, and add a bias value to each element of the two-dimensional matrix After that, input the activation function in turn , obtain a one-dimensional matrix with nonlinear representation ; Use the Sobel-y convolution kernel to perform two-dimensional convolution on the fused image to obtain a two-dimensional matrix, and add a bias value to each element of the two-dimensional matrix After that, input the activation function in turn , obtain a one-dimensional matrix with nonlinear representation ; S32: Convert one-dimensional matrix , and the attribute matrix obtained in step S2 Stitched together to obtain an enhanced feature matrix .

6. The video object denoising method based on contextual spatial information according to claim 5, characterized in that: The offset value in step S31 and All are determined through training, the specific process includes: Construct single-layer convolutional networks based on Sobel-x convolution kernel and Sobel-y convolution kernel respectively, and randomly initialize all weights and bias functions; Construct a dataset of fused images, train two single-layer convolutional networks on the dataset, and obtain the cost function; Based on the step size, the bias value is continuously adjusted within the bias value range until the optimal cost function is obtained. At this time, the corresponding bias value is and .

7. The video object denoising method based on contextual spatial information according to claim 5, characterized in that: The step S5 comprises: The matrix Input the trained autoencoder, and the encoder converts the matrix Compressed to a low-dimensional space, the decoder restores the low-dimensional space through the inverse process to obtain the same matrix Output of the same dimension, label the output as the denoised feature matrix .

8. The video object denoising method based on contextual spatial information according to claim 7, characterized in that: The step S6 comprises: S61: Denoising the feature matrix obtained in step S5 Before interception elements, and get the matrix , represents the one-dimensional matrix obtained in step S31 Length; S62: Matrix The transposed convolution transform is used to obtain the denoised video image.

9. A video object denoising system based on contextual spatial information, characterized in that: include: Image fusion module: Decompose temporally adjacent video images to obtain high-frequency residual images, align temporally adjacent video images according to the high-frequency residual images, reconstruct the Laplacian pyramids of temporally adjacent video images in sequence, and obtain a fused image based on image reconstruction; Attribute matrix construction module: Based on the temporally adjacent video images, the target motion state is determined and the attribute matrix of the fused image is constructed; Feature matrix construction module: perform convolution feature extraction and nonlinear matrix calculation on the fused image, and concatenate the obtained one-dimensional matrix with the attribute matrix to obtain an enhanced feature matrix; Autoencoder training module: construct the cost function of the autoencoder, set blind spots for the feature matrix and input it into the autoencoder for training; Feature matrix reconstruction module: Use the trained autoencoder to reconstruct the enhanced feature matrix to obtain the denoised matrix; Feature matrix conversion module: intercepts and transposes the denoised matrix to obtain the denoised video image; To implement the video target denoising method based on contextual spatial information as described in any one of claims 1-8.

Images