Noise image denoising and dynamic target enhancement method based on joint filtering in spatiotemporal domain

Abstract

The present invention discloses a method for denoising noisy images and enhancing dynamic targets based on joint spatial-temporal filtering, comprising: performing downsampling differential pyramid filtering on each frame of an image to decompose the image information in different frequency bands; analyzing high-frequency noise characteristics to obtain a spatial domain filtering convolution template that suppresses maximum values; performing spatial domain convolution on each layer of the image using the spatial domain filtering convolution template that suppresses maximum values ​​to calculate motion normalization parameters; and performing motion compensation denoising in the time domain by combining the DOP image of the current frame with the result of motion compensation of the previous frame. The present invention can effectively remove various types of noise from noisy images in noisy scenes while balancing computational speed and algorithm stability, and retain dynamic targets as much as possible, making the scene clear and enhancing the details of the targets.

Description

Technical Field

[0001] The present invention belongs to the technical field of image denoising and enhancement, and specifically relates to a method for denoising a noisy image and enhancing a dynamic target based on joint filtering in time and space domains. Background Art

[0002] Noise is a significant cause of image interference. Especially in high-noise environments, cameras can generate numerous types of noise, disrupting image quality. This noise can occur during transmission or during processing such as quantization. It can severely impact visual perception and the observation of image detail. This noise can range from additive noise to pervasive noise. However, current denoising methods are unable to simultaneously remove noise and preserve target detail. This can be specifically described in the following aspects:

[0003] (1) Denoising algorithms based on the spatial domain: Commonly used classic methods include median filtering, mean filtering, Gaussian filtering, etc. They use the pixel information of the pixel point and its fixed neighborhood to modify its own grayscale, color and other intensity values, which will lead to smoothing. The spatial domain denoising algorithm is a method of denoising by directly calculating the grayscale values ​​of image pixels in two-dimensional space. The spatial domain filtering algorithm is simple and easy to implement. When used for denoising, it does not consider the time domain information, and the noise at the same position between frames is random, which can easily lead to obvious flickering between adjacent frames after filtering.

[0004] (2) Denoising algorithm based on the time domain: The main application of time domain denoising is the multi-frame noise reduction algorithm. Time domain filtering uses the information redundancy between video frames to restore noise points by performing motion detection and motion estimation on adjacent frames, but it is prone to noise residual or "ghosting" phenomenon. The VBM3D algorithm proposed for video denoising can incorporate redundant information in the time domain into the video denoising algorithm, but this type of time domain denoising algorithm often requires multiple block matching, which has high computational complexity and long processing time.

[0005] (3) Learning-based methods: Learning-based methods focus on learning the latent mapping from noisy images to clean images and can be divided into traditional learning-based methods and deep network-based learning methods. In recent years, deep network-based methods have become the mainstream method because they have achieved more promising denoising results than filtering-based, model-based, and traditional learning-based methods.

[0006] Other image denoising methods, such as the block set that can be represented as many similar blocks in the same image, can use the BM3D algorithm based on block matching. However, such algorithms are computationally too complex and cannot adapt to images with stronger noise or may cause distortion due to improper parameter settings.

[0007] In summary, the problems existing in the existing technology can be summarized as: poor denoising effect; blurred edges; complex calculations; too many parameters, etc. Summary of the Invention

[0008] The present invention aims to provide a method for denoising noisy images and enhancing dynamic objects based on joint spatial and temporal filtering. This method can suppress image noise while preserving image details and object outlines as much as possible. Furthermore, in scenes with motion or significant scene changes, it can effectively suppress the "smearing" phenomenon caused by moving objects in the image.

[0009] The technical solution for achieving the purpose of the present invention is: a method for denoising a noisy image and enhancing a dynamic target based on joint filtering in the spatiotemporal domain, comprising the following steps:

[0010] (10) Downsampling differential pyramid filtering: Downsampling differential pyramid filtering is performed on each frame of the image to decompose the image into different frequency bands;

[0011] (20) High-frequency noise suppression: Analyze the characteristics of high-frequency noise and obtain the spatial domain filtering convolution template that suppresses the maximum value;

[0012] (30) Four-directional spatial edge filtering: Use the spatial filtering convolution template that suppresses the maximum value to perform spatial convolution on each layer of the image and calculate the motion normalization parameters;

[0013] (40) Temporal domain motion compensation: Combine the DOP image of the current frame with the result of motion compensation of the previous frame to perform motion compensation denoising in the temporal domain.

[0014] Preferably, the specific steps of downsampling differential pyramid filtering in step (10) are:

[0015] (11) 2x downsampling:

[0016] The original image is downsampled 6 times by a factor of 2 to obtain 6 downsampled images with different resolutions, i.e. 6 scale spaces;

[0017] (12) 2x upsampling:

[0018] Double the current downsampled image to make it have the same resolution as the image in the previous scale space;

[0019] (13) Subtract the differences of each scale:

[0020] Subtracting two images with different blur levels in the same scale space can obtain difference images DOP1, DOP2, DOP3, DOP4, DOP5, and DOP6 with different resolutions, which can be expressed as follows:

[0021] {DoP} k={prydown} k -{resize} k

[0022] Where, {DoP} k represents the k-th level differential pyramid downsampled image, {prydown} k represents the k-th pyramid downsampled image, {resize} k Represents the restored image at the kth level.

[0023] Preferably, in step (20) high frequency noise suppression, the specific method for obtaining the filter template is:

[0024] By analyzing the noise characteristics of visible light images under dark light conditions, it can be concluded that high-frequency noise follows a Gaussian distribution, that is, the grayscale value has a high value in the middle and a low value in the surrounding neighborhood. Therefore, the high-frequency noise in each layer of DOP is mostly the maximum value point in its neighborhood. When finding the extreme value of a curve, the local extreme value can often be found by taking the derivative:

[0025]

[0026] Where f(x) represents the original function, f'(x) represents the derivative of the original function, and δx is an infinitesimal quantity.

[0027] However, the pixel values ​​of the image are discrete points, so the step size must be an integer, that is, 1 pixel, so the maximum point can be expressed as:

[0028]

[0029]

[0030] High-frequency noise often only has the characteristics of local maximum values ​​but no directional information. Using an operator based on gradient filtering can eliminate isolated noise points and extract edge contours with directional information.

[0031] For images, digital filtering can often be converted into a convolution operation process between a template and an image window. The filtering process can be achieved by replacing the center pixel of the window with the result of the convolution operation. Therefore, the above local noise suppression can be converted into a convolution template in four directions, as shown in the following formula:

[0032]

[0033] In the picture They represent the spatial convolution templates of the image in the directions of 0°, 45°, 90°, and 135° respectively. The high-frequency noise of each layer DOP can be suppressed by filtering the templates.

[0034] Preferably, the specific steps of the four-directional spatial edge filtering in step (30) are:

[0035] The high-frequency noise suppression filter templates in four directions obtained in step 20 are used to operate on a layer of DOP to obtain high-frequency noise suppression images in four directions. The process can be expressed as follows:

[0036] {I f} C =I*{Conv} C

[0037] where {Conv} C is a high-frequency noise suppression filter template in four directions, I is a layer of DOP, {I f} C are high-frequency noise suppression images in four directions.

[0038] (32) Repeat step 31 and perform spatial high-frequency noise filtering on each layer of DOP. The high-frequency noise suppression image in the four directions of each layer is expressed as:

[0039] {I f C} k =I k *{Conv C} k

[0040] Where {Conv C} k is the high-frequency noise suppression filter template in four directions of the k-layer, I k is the k-layer DOP, {I f C} k is the high-frequency noise suppression image in four directions of the k-layer.

[0041] (33) The six layers of DOP in each direction are fused to obtain the fused images in four directions. The process can be expressed as:

[0042]

[0043] Fuse is the fusion process, is the fused image of k layers in each direction.

[0044] (34) Calculate the motion signal normalization parameters:

[0045] The four-directional spatial fusion image calculated according to step 33 After averaging and normalization, the a_sig parameter can be obtained. This parameter is the update coefficient of the edge feature information. The larger the value of a_sig, the greater the impact of the current frame DOP value on the denoising result.

[0046]

[0047] Where a_sig n k represents the normalization parameter of motion signal, I f k0 , I f k45 , I f k90 , I f k135 represents the fused images at the four directions of 0°, 45°, 90°, and 135° obtained in step 33, and |·| represents the absolute value operation.

[0048] (35) Calculate the normalization parameters of the differential signal

[0049] The data in the corresponding direction obtained after spatial filtering of the current frame and the previous frame are differentially calculated. The four directions are added, averaged, and normalized to obtain the a_dif parameter, which is the motion area update coefficient. The larger the a_dif value, the greater the motion amplitude and the faster the update.

[0050]

[0051] In the formula, a_dif n k Differential signal normalization parameter, I f k0 -I f (k-1)0 Indicates the calculation of the differential signal between the 0°k layer and the k-1 layer, I f k45 -I f (k-1)45 Indicates the calculation of the differential signal between the 45° k layer and the k-1 layer, I f k90 -I f (k-1)90 Indicates the calculation of the differential signal between the 90° k layer and the k-1 layer, I f k135 -I f (k-1)135 represents the calculation of the differential signal between the kth layer and the k-1th layer at 135°, and |·| represents the absolute value operation.

[0052] Preferably, the specific steps of time domain motion compensation in step (40) include:

[0053] (41) Image motion compensation:

[0054] Combine the DOP image of the current frame with the result of motion compensation of the previous frame, and perform motion compensation denoising in the time domain. The denoising formula is as follows:

[0055]

[0056] in, represents the result of DOP motion compensation denoising at the kth layer of the nth frame image, {Dop} k represents the k-th level differential pyramid downsampled image;

[0057] (42) Image fusion of each layer:

[0058] The six-layer DOP image is restored to the resolution of the original image and fused. The fusion formula is as follows:

[0059]

[0060] Among them, I n It is the image after the fusion of the nth frame, DOP_mean_re n k It is an image with the resolution size of the original image restored by each layer of DOP after denoising;

[0061] (43) Second spatiotemporal filtering:

[0062] to I n Perform four-directional spatial edge filtering and fusion.

[0063] Compared with the prior art, the present invention has the following significant advantages:

[0064] 1. Obvious denoising effect: The present invention can deal with multiple types of noise at the same time, and has obvious denoising effect.

[0065] 2. For motion scenes: Existing algorithms can only process single-frame images in most cases. If it is a continuous frame image, the freshness is poor. The present invention calculates the displacement of multiple frames through phase correlation and estimates the camera motion parameters. While ensuring denoising, it will not be affected by the camera motion state.

[0066] 3. Small information loss: Traditional denoising algorithms will more or less lose some edge information while denoising, and the stronger the noise, the more serious the loss. The present invention has a certain protection effect on the edge, and the loss of details is small. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a flow chart of the noisy image denoising and dynamic target enhancement method based on joint filtering in the spatiotemporal domain.

[0068] Figure 2 This is the downsampling differential pyramid filtering flow chart.

[0069] Figure 3 are differential images at different scales.

[0070] Figure 4 The difference diagrams of different frequency bands and their three-dimensional response diagrams.

[0071] Figure 5 It is a four-directional spatial edge filtering map.

[0072] Figure 6 This is the effect diagram of time domain processing.

[0073] Figure 7 The figure is the processing effect (1).

[0074] Figure 8 This is the processing effect diagram (2). DETAILED DESCRIPTION

[0075] like Figure 1 As shown in FIG, a method for denoising a noisy image and enhancing a dynamic target based on joint filtering in the spatiotemporal domain includes the following steps:

[0076] (10) Downsampling differential pyramid filtering: Downsampling differential pyramid filtering is performed on each frame of the image to decompose the image into different frequency bands;

[0077] (20) High-frequency noise suppression: Analyze the characteristics of high-frequency noise and obtain the spatial domain filtering convolution template that suppresses the maximum value;

[0078] (30) Four-directional spatial edge filtering: Use the spatial filtering convolution template that suppresses the maximum value to perform spatial convolution on each layer of the image and calculate the motion normalization parameters;

[0079] (40) Temporal domain motion compensation: Combine the DOP (pyramid downsampling difference map) image of the current frame with the result of motion compensation of the previous frame to perform motion compensation denoising in the temporal domain.

[0080] In a further embodiment, the (10) downsampling differential pyramid filtering step is specifically as follows:

[0081] (11) 2x downsampling:

[0082] The original image is downsampled 6 times by 2 times to obtain 6 downsampled images with different resolutions, that is, 6 scale spaces

[0083] (12) 2x upsampling:

[0084] Double the current downsampled image to make it have the same resolution as the image in the previous scale space

[0085] (13) Subtract the differences of each scale:

[0086] The difference image can be obtained by subtracting two images with different blur levels in the same scale space. DOP1, DOP2, DOP3, DOP4, DOP5, and DOP6 are downsampled difference images with different resolutions. The difference image can be expressed as:

[0087] {Dop} k ={prydown} k -{resize} k

[0088] Where, {Dop} k represents the k-th level differential pyramid downsampled image, {prydown} k represents the k-th pyramid downsampled image, {resize} k Represents the restored image at the kth level.

[0089] The downsampling differential pyramid filtering flow chart is as follows Figure 2 shown.

[0090] Different scale difference images such as Figure 3 shown

[0091] The difference graphs of different frequency bands and their three-dimensional response graphs are as follows: Figure 4 shown

[0092] exist Figure 2 In the figure, the green line represents a 2x downsampling operation, which is performed 6 times by 2x, resulting in 6 downsampled images of different resolutions, i.e., 6 scale spaces. The blue line represents a 2x upsampling operation, which doubles the resolution of the current downsampled image to make it the same as the image in the previous scale space. The orange line represents a subtraction operation, which subtracts two images of the same scale space but with different blur levels to obtain a difference image. DOP1, DOP2, DOP3, DOP4, DOP5, and DOP6 are downsampled difference images of different resolutions, respectively.

[0093] exist Figure 3 In the figure, (a), (b), (c), (d), (e), and (f) represent DOP1, DOP2, DOP3, DOP4, DOP5, and DOP6, respectively.

[0094] exist Figure 4 From the three-dimensional response diagram of DOP in each frequency band in the above figure, we can see that DOP1 contains the most high-frequency noise because it has only been downsampled once. The more times the downsampling is performed, the less high-frequency noise the corresponding DOP contains. It can be seen that DOP5 and DOP6 basically do not contain high-frequency noise.

[0095] In a further embodiment, the step (20) of suppressing high frequency noise and obtaining a filter template includes:

[0096] By analyzing the noise characteristics of visible light images under dark light conditions, it can be concluded that high-frequency noise follows a Gaussian distribution, that is, the grayscale value has a high value in the middle and a low value in the surrounding neighborhood. Therefore, the high-frequency noise in each layer of DOP is mostly the maximum value point in its neighborhood. When finding the extreme value of a curve, the local extreme value can often be found by taking the derivative:

[0097]

[0098] Where f(x) represents the original function, f'(x) represents the derivative of the original function, and δx represents an infinitesimal quantity.

[0099] However, the pixel values ​​of the image are discrete points, so the step size must be an integer, that is, 1 pixel, so the maximum point can be expressed as:

[0100]

[0101]

[0102] High-frequency noise often only has the characteristics of local maximum values ​​but no directional information. Using an operator based on gradient filtering can eliminate isolated noise points and extract edge contours with directional information.

[0103] For images, digital filtering can often be converted into a convolution operation process between a template and an image window. The filtering process can be achieved by replacing the center pixel of the window with the result of the convolution operation. Therefore, the above local noise suppression can be converted into a convolution template in four directions, as shown in the following formula:

[0104]

[0105] In the picture They represent the spatial convolution templates of the image in the directions of 0°, 45°, 90°, and 135° respectively. The high-frequency noise of each layer DOP can be suppressed by filtering the templates.

[0106] In a further embodiment, the (30) four-directional spatial edge filtering step includes:

[0107] (31) Suppress high-frequency noise:

[0108] The high-frequency noise suppression filter templates in four directions obtained in step 20 are used to operate on a layer of DOP to obtain high-frequency noise suppression images in four directions. The process can be expressed as follows:

[0109] {I f} C =I*{Conv} C

[0110] where {Conv}C is a high-frequency noise suppression filter template in four directions, I is a layer of DOP, {I f} C are high-frequency noise suppression images in four directions.

[0111] (32) Repeat step 31 and perform spatial high-frequency noise filtering on each layer of DOP (pyramid downsampling difference map). The high-frequency noise suppression image in the four directions of each layer is expressed as:

[0112] {I f C} k =I k *{Conv C} k

[0113] Where {Conv C} k is the high-frequency noise suppression filter template in four directions of the k-layer, I k is the k-layer DOP, {I f C ] k is the high-frequency noise suppression image in four directions of the k-layer.

[0114] (33) The six layers of DOP in each direction are fused to obtain the fused images in four directions. The process can be expressed as:

[0115]

[0116] Fuse is the fusion process, is the fused image of k layers in each direction.

[0117] (34) Calculate the motion signal normalization parameters:

[0118] The four-directional spatial fusion image calculated according to step 33 After averaging and normalization, the a_sig parameter can be obtained. This parameter is the update coefficient of the edge feature information. The larger the value of a_sig, the greater the impact of the current frame DOP value on the denoising result.

[0119]

[0120] Where a_sig n k represents the normalization parameter of motion signal, I f k0 , I f k45 , I f k90 , I f k135represents the fused images at the four directions of 0°, 45°, 90°, and 135° obtained in step 33, and |·| represents the absolute value operation.

[0121] (35) Calculate the normalization parameters of the differential signal

[0122] The data in the corresponding direction obtained after spatial filtering of the current frame and the previous frame are differentially calculated. The four directions are added, averaged, and normalized to obtain the a_dif parameter, which is the motion area update coefficient. The larger the a_dif value, the greater the motion amplitude and the faster the update.

[0123]

[0124] In the formula, a_dif n k Differential signal normalization parameter, I f k0 -I f (k-1)0 Indicates the calculation of the differential signal between the 0°k layer and the k-1 layer, I f k45 -I f (k-1)45 Indicates the calculation of the differential signal between the 45° k layer and the k-1 layer, I f k90 -I f (k-1)90 Indicates the calculation of the differential signal between the 90° k layer and the k-1 layer, I f k135 -I f (k-1)135 represents the calculation of the differential signal between the kth layer and the k-1th layer at 135°, and |·| represents the absolute value operation.

[0125] The spatial edge filtering diagram of the four directions of the spatial domain is as follows Figure 5 shown.

[0126] Depend on Figure 5 It can be seen that after the original image containing noise is downsampled and filtered by differential pyramid filtering, four images can be obtained by using spatial filtering templates in the four directions of 0°, 45°, 90°, and 135°, which emphasize edge extraction in different directions and suppress local maximum values. It can be seen that the noise in the image obtained after the fusion of the four images is significantly reduced, and the feature information is more prominent.

[0127] In a further embodiment, the (40) temporal motion compensation step comprises:

[0128] (41) Image motion compensation:

[0129] Combine the DOP image of the current frame with the result of motion compensation of the previous frame, and perform motion compensation denoising in the time domain. The denoising formula is as follows:

[0130]

[0131] in, Indicates the result of the k-th layer DOP motion compensation denoising of the n-th frame. The denoising result of the current frame is determined by the denoising result of the previous frame and the DOP image of the current frame.

[0132] (42) Image fusion of each layer:

[0133] Since the resolutions of the six-layer DOP images obtained after noise suppression in the above three steps are different, they need to be restored to the resolution of the original images and then fused. The fusion formula is as follows:

[0134]

[0135] Among them, I n It is the image after the fusion of the nth frame, DOP_mean_re n k It is an image with the resolution of the original image restored by each layer of DOP after denoising.

[0136] (43) Second spatiotemporal filtering:

[0137] Since the fused image I n There will inevitably be some noise in the I n Perform a four-way spatial edge filtering and fusion to further enhance the image feature information and suppress image noise. The final spatial filtered image I n It is the final result of the denoising algorithm based on spatiotemporal filtering.

[0138] The effect of temporal motion compensation is shown in the figure Figure 6 shown.

[0139] The effect diagram of the spatiotemporal algorithm in scene 1 is as follows Figure 7 shown.

[0140] The effect diagram of the spatiotemporal algorithm in scenario 2 is as follows Figure 8 shown.

Claims

1. A method for denoising noisy images and enhancing dynamic targets based on joint filtering in spatiotemporal domains, characterized in that: The steps include: (10) Downsampling differential pyramid filtering: Downsampling differential pyramid filtering is performed on each frame of the image to decompose the image into different frequency bands. The specific steps are as follows: (11) 2x downsampling: The original image is downsampled 6 times by a factor of 2 to obtain 6 downsampled images with different resolutions, i.e. 6 scale spaces; (12) 2x upsampling: Double the current downsampled image to make it have the same resolution as the image in the previous scale space; (13) Subtract the differences of each scale: Subtracting two images with different blur levels in the same scale space can obtain difference images DOP1, DOP2, DOP3, DOP4, DOP5, and DOP6 with different resolutions, which can be expressed as follows: {DoP} k =(prydown} k -{resize} k Where, {DoP} k represents the k-th level differential pyramid downsampled image, {prydown} k represents the k-th pyramid downsampled image, {resize} k Represents the k-th level scale restored image; (20) High-frequency noise suppression: Analyze the characteristics of high-frequency noise and obtain the spatial domain filtering convolution template that suppresses the maximum value. The specific method is: By analyzing the noise characteristics of visible light images under dark light conditions, it can be concluded that high-frequency noise obeys a Gaussian distribution, that is, the grayscale value has the characteristics of being high in the middle and low in the surrounding neighborhood. Therefore, the high-frequency noise in each layer of DOP is mostly the maximum point in its neighborhood. When finding the extreme value in a curve, the local extreme value can often be achieved by taking the derivative: Where f(x) represents the original function, f′(x) represents the derivative of the original function, and δx is an infinitesimal quantity; However, the pixel values ​​of the image are discrete points, so the step size must be an integer, that is, 1 pixel, so the maximum point can be expressed as: High-frequency noise often only has the characteristics of local maxima, but no directional information. Using an operator based on gradient filtering can eliminate isolated noise and extract edge contours with directional information; For images, digital filtering can often be converted into a convolution operation process between a template and an image window. The filtering process can be achieved by replacing the center pixel of the window with the result of the convolution operation. Therefore, the above local noise suppression can be converted into a convolution template in four directions, as shown in the following formula: In the picture Respectively represent the spatial convolution templates of the image in the directions of 0°, 45°, 90°, and 135°. The high-frequency noise of each layer DOP can be suppressed by filtering the templates. (30) Four-directional spatial edge filtering: Use the spatial filtering convolution template that suppresses the maximum value to perform spatial convolution on each layer of the image and calculate the motion normalization parameters. The specific steps are as follows: The high-frequency noise suppression filter templates in four directions obtained in step 20 are used to operate on a layer of DOP to obtain high-frequency noise suppression images in four directions. The process can be expressed as follows: {I f } C =I*{Conv} C where {Conv} C is a high-frequency noise suppression filter template in four directions, I is a layer of DOP, {I f } C are high-frequency noise suppression images in four directions; (32) Repeat step 31 and perform spatial high-frequency noise filtering on each layer of DOP. The high-frequency noise suppression image in the four directions of each layer is expressed as: {I f C } k =I k *{Conv C } k Where {Conv C } k is the high-frequency noise suppression filter template in four directions of the k-layer, I k is the k-layer DOP, {I f C } k is the high-frequency noise suppression image in four directions of the k-layer; (33) The six layers of DOP in each direction are fused to obtain the fused images in four directions. The process can be expressed as: Fuse is the fusion process, is the fused image of k layers in each direction; (34) Calculate the motion signal normalization parameters: The four-directional spatial fusion image calculated according to step 33 After averaging and normalization, we can get the a_sig parameter, which is the update coefficient of edge feature information. The larger the a_sig value is, the greater the impact of the current frame DOP value on the denoising result; Where a_sig n k represents the normalization parameter of motion signal, I f k0 , I f k45 , I f k90 , I f k135 represents the fused images at the four directions of 0°, 45°, 90°, and 135° obtained in step 33, and |·| represents the absolute value operation; (35) Calculate the normalization parameters of the differential signal The data of the corresponding direction obtained after spatial filtering of the current frame and the previous frame are differentially calculated. The four directions are added and averaged and then normalized to obtain the a_dif parameter, which is the motion area update coefficient. The larger the a_dif value, the greater the motion amplitude and the faster the update. In the formula, a_dif n k Differential signal normalization parameter, I f k0 -I f (k-1)0 Indicates the calculation of the differential signal between the 0°k layer and the k-1 layer, I f k45 -I f (k-1)45 Indicates the calculation of the differential signal between the 45° k layer and the k-1 layer, I f k90 -I f (k-1)90 Indicates the calculation of the differential signal between the 90° k layer and the k-1 layer, I f k135 -I f (k-1)135 represents the calculation of the differential signal between the kth layer and the k-1th layer at 135°, and |·| represents the absolute value operation; (40) Temporal domain motion compensation: Combine the DOP image of the current frame with the result of motion compensation of the previous frame to perform motion compensation denoising in the temporal domain.

2. The method for denoising noisy images and enhancing dynamic targets based on spatiotemporal joint filtering according to claim 1, characterized in that: The specific steps of step (40) temporal domain motion compensation include: (41) Image motion compensation: Combine the DOP image of the current frame with the result of motion compensation of the previous frame, and perform motion compensation denoising in the time domain. The denoising formula is as follows: in, represents the result of the k-th layer DOP motion compensation denoising of the n-th frame image, {DoP} k represents the k-th level differential pyramid downsampled image; (42) Image fusion of each layer: The six-layer DOP image is restored to the resolution of the original image and fused. The fusion formula is as follows: Among them, I n It is the image after the fusion of the nth frame, DOP_mean_re n k It is an image with the resolution size of the original image restored by each layer of DOP after denoising; (43) Second spatiotemporal filtering: to I n Perform four-directional spatial edge filtering and fusion.

Images