An event camera based joint de-motion-blur and video interpolation method

Abstract

The application discloses a kind of joint motion deblurring and video interpolation method based on event camera, comprising:1, the acquisition and preprocessing of low frame rate blurred video data based on event camera,2, construct joint motion deblurring and video interpolation network under blind exposure limit, including: feature extraction module, event-guided exposure prior estimation module, time-exposure joint control module and clear video frame reconstruction module,3, adopt the training strategy of stage optimization network.The application can use event camera to record the motion of scene, effectively overcome exposure ambiguity, obtain accurate exposure estimation, so as to solve blind exposure limit, and effectively fuse video frame data and event data using developed deep learning algorithm, so as to effectively remove motion blur and realize video interpolation, achieve optimal quantization and visual effect.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and in particular to an event camera-based joint de-motion blurring and video frame interpolation method. Background Art

[0002] In recent years, computer vision researchers have conducted extensive research in the field of interpolation of blurred video frames. The development of interpolation algorithms for blurred video frames is still in its early stages. A cascaded multi-task approach is the simplest approach. It combines two tasks, video deblurring and interpolation, in tandem, achieving simultaneous motion deblurring and improved frame rate. Compared to this cascaded multi-task approach, an integrated restoration approach is a more efficient option because it avoids the cumulative errors inherent in cascading multiple tasks. While existing interpolation approaches for blurred video frames have achieved some success, they rely on certain exposure assumptions, such as a fixed exposure time equal to the inverse of the frame rate. This assumption of non-blind exposure contradicts actual video capture conditions. In real-world video capture, complex motion and lighting conditions often result in variable and unknown exposure times. This variable and unknown exposure is also known as blind exposure constraints. This blind exposure assumption poses significant challenges to existing non-blind exposure-based interpolation algorithms for blurred video frames, resulting in significant performance degradation and even the introduction of additional noise.

[0003] The key to solving the problem of interpolating blurred videos under blind exposure constraints is obtaining accurate exposure estimates. However, accurate exposure estimation is difficult to obtain from blurred videos alone. This is because the blurriness of a blurred video frame is determined by two factors: motion speed and exposure time. This is known as exposure ambiguity. This means that the blur produced by high-speed motion under short exposure is likely to be the same as the blur produced by slow motion under long exposure. This exposure ambiguity causes existing methods to produce severe visual blur and loss of detail. Summary of the Invention

[0004] In order to address the shortcomings of the above-mentioned existing technologies, the present invention proposes a joint de-motion blurring and video interpolation method based on an event camera, in order to effectively overcome the exposure ambiguity of the scene motion recorded by the event camera, obtain accurate exposure estimation, and solve the blind exposure limitation. By fusing video frame data and event data, the motion blur is effectively removed and video interpolation is realized, thereby achieving optimal quantization and visual effects.

[0005] In order to achieve the above-mentioned object, the present invention adopts the following technical solutions:

[0006] The present invention provides a method for joint motion blur removal and video frame interpolation based on an event camera, comprising the following steps:

[0007] Step 1: Collection and preprocessing of low-frame-rate blurry video data from event cameras:

[0008] After shooting the dynamic scene with the event camera, we get continuous clear video frames and their asynchronous event stream signals, and let one of the clear video frames be denoted as I gt , let the asynchronous event stream signal corresponding to the clear video frame I be recorded as Among them, e k represents the kth event, and e k =p k δ(tt k , xx k ,yy k ), p k Represents the kth event e k Polarity, p k ∈{-1, 1}; t k Represents the kth event e k The time of occurrence; x k and y k Represents the kth event e k The spatial coordinates of occurrence; N represents the total number of events; (t, x, y) represents the time-space projection coordinates; δ is the indicator function;

[0009] The continuous clear video frames are averaged as the clear video frame I gt The corresponding frame is blurred video frame B; and the blurred video frame B is converted into a blur degree BL using Laplace transform function Laplace();

[0010] Step 2: Use formula (1) to convert the asynchronous event stream signal Convert to event frame V∈R 2×H×W :

[0011] V=∑ k＝1 δ(xx k ,yy k ,pp k ) (1)

[0012] In formula (1), H and W represent the height and width of the event frame V respectively;

[0013] Step 3: Construct a joint de-motion blurring and video interpolation network under blind exposure constraints, including: a feature extraction module, an event-guided exposure prior estimation module, a joint time-exposure control module, and a clear video frame reconstruction module;

[0014] Step 3.1: The feature extraction module processes the event frame V and the blurred video frame B using a convolution layer to obtain the event features. and blurred video frame features

[0015] Step 3.2: The event-guided exposure prior estimation module processes the event frame V to obtain a predicted exposure prior EP;

[0016] Step 3.3: The time-exposure joint control module processes the exposure prior EP to obtain the joint control feature Feat control ;

[0017] Step 3.4: The clear video frame reconstruction module performs joint control feature Feat control Processing is performed to obtain the initial clear frame I init And the final clear frame I final ;

[0018] Step 4: Optimize the network using a phased training strategy:

[0019] Step 4.1: Use formula (10) to construct the loss function of the first-stage event-guided exposure prior estimation module

[0020]

[0021] In formula (10), EP gt is the true exposure prior; represents the mean square error loss function;

[0022] Step 4.2: Use formula (11) to construct the loss function of the remaining module in the second stage

[0023]

[0024] In formula (11), Represents the network loss function, and has:

[0025]

[0026] In formula (12), represents the composite loss function, and has:

[0027]

[0028] In formula (13), represents the Laplace loss function, Denotes the Hamming loss function, and has:

[0029]

[0030]

[0031] In formula (14) and formula (15), I represents I init or I final , S represents the number of layers, L s () represents the Laplace pyramid of level s, Dis() represents the Hamming distance, and Cen() represents the census transform;

[0032] Step 4.3: Use the gradient descent method to train the joint demotion blurring and video interpolation network in stages. In the first stage, use formula (10) to optimize the event-guided exposure prior estimation module. In the second stage, use formula (11) to optimize the remaining modules and fine-tune the network parameters to obtain the trained joint demotion blurring and video interpolation network, which is used to realize the deblurring and interpolation processing of any low-frame-rate blurred video frame.

[0033] The joint de-motion blurring and video interpolation method based on event camera of the present invention is also characterized in that the event-guided exposure prior estimation module includes two feature extraction networks with the same structure but no shared parameters. and a feature aggregation network Φ aggre ;

[0034] The feature extraction network consists of a convolutional layer and a LeakyReLU activation function;

[0035] The feature aggregation network Φ aggre It consists of two convolutional layers. The activation function of the first convolution layer is LeakyReLU, and the second convolution layer has no activation function.

[0036] Step 3.2.1: The event frame V is input into the feature extraction network Process it and output the event feature Feat e ;

[0037] The frame blurred video frame B and the blur level BL are stacked in the channel dimension and then input into the feature extraction network Process it and output the fuzzy video feature Feat bf ;

[0038] Step 3.2.2: The feature aggregation network Φ aggre Using formula (5) to get the aggregated feature Feat aggre :

[0039] Feat aggre =Φ aggre ([Feat e ×Sigmoid(GN(Feat e )×GN(Feat bf)), Feat bf ]) (5)

[0040] In formula (5), Sigmoid() represents the sigmoid activation function, and GN() represents the group norm normalization process;

[0041] The feature aggregation network Φ aggre The predicted exposure prior EP is obtained using formula (6):

[0042] EP=AP(Feat aggre ) (6)

[0043] In formula (6), EP ranges from 0 to 1, and AP represents average pooling.

[0044] The time-exposure joint control module includes n cascaded submodules, each of which consists of two feature transfer convolutional layers Two control convolutional layers And a fused convolutional layer Φ fuse composition;

[0045] Among them, the i-th submodule uses formula (7) to obtain the i-th intermediate control feature Feat i , so that the nth cascaded submodule outputs the joint control feature Feat control ;

[0046]

[0047] In formula (7), Feat i-1 represents the intermediate control feature output by the i-1th submodule; when i=1, let i∈[1,N].

[0048] The clear video frame reconstruction module consists of an initial clear frame prediction submodule and a final clear frame recovery submodule Φ detail ;

[0049] Step 3.4.1, the initial clear frame prediction submodule consists of a convolution module, a kernel prediction module KPN, a connection module and an image reconstruction module Φ recons constitute;

[0050] The convolution module is used to jointly control the feature Feat control Perform convolution processing to obtain the convolution feature Feat′ control ;

[0051] The kernel prediction module KPN uses formula (9) to obtain the initial convolution feature

[0052]

[0053] In formula (9), Conv() represents the processing of convolution layer and LeakyReLU activation function, * represents convolution;

[0054] The connection module uses formula (10) to obtain the initial clear frame feature

[0055]

[0056] The image reconstruction module Φ recons It consists of a convolution layer, a PixelShuffle layer and a LeakyReLU activation function, and After processing, the initial clear frame I is obtained init ;

[0057] Step 3.4.2, the final clear frame recovery submodule Φ detail Will I init After splicing with B, the final clear frame I is obtained final .

[0058] An electronic device of the present invention includes a memory and a processor, and is characterized in that the memory is used to store a program that supports the processor to execute any of the joint de-motion blur and video interpolation methods, and the processor is configured to execute the program stored in the memory.

[0059] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program executes any step of the joint de-motion blurring and video interpolation method when executed by a processor.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. This invention proposes an event-guided exposure prior estimation module that utilizes a high-temporal-resolution event camera to derive motion priors. Taking an asynchronous event stream, blurred video frames, and blur levels as input, the module is constructed using a convolutional neural network. Group normalization techniques are employed to eliminate the discrepancy between the video and event stream modalities, thereby achieving an accurate exposure prior estimate. This module leverages the high temporal resolution of the event camera to provide precise motion capture capabilities, making the blind exposure problem solvable. This overcomes the limitations of existing motion deblurring and video interpolation techniques, resulting in optimal motion deblurring and video interpolation results.

[0062] 2. This invention proposes a combined time-exposure control module that takes interpolation time, exposure priors, and event frame features as inputs and uses a convolutional neural network to build the main module. This module uses interpolation time and exposure priors as two control factors and designs a dual control path to regulate event frame features. This allows the two control factors to work together efficiently, overcoming the data bias associated with single-path control. Furthermore, this module expands the entire control process, using an iterative optimization strategy to progressively control event frame features, making the entire control process more accurate and easier to optimize through neural networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A network structure diagram of the event camera-based joint de-motion blurring and video interpolation implemented in the present invention;

[0064] Figure 2 A flowchart for implementing the event-guided exposure prior estimation module of the present invention;

[0065] Figure 3 Flowchart of the time-exposure joint control module implemented in the present invention. DETAILED DESCRIPTION

[0066] In this embodiment, a joint de-motion blurring and video interpolation method based on an event camera is proposed. First, an event-guided exposure estimation module is used to obtain an exposure prior, thereby fully utilizing the high temporal resolution of the event camera to transform the blind exposure problem into a non-blind exposure problem, eliminating the interference caused by the blind exposure problem. Then, a time-exposure joint control module is used to inject the time and exposure factors into the network processing process, and a two-way iterative strategy is introduced to achieve stable control. For the detailed process, see [1]. Figure 1 The method is carried out as follows:

[0067] Step 1: Collection and preprocessing of low-frame-rate blurry video data from event cameras:

[0068] After shooting the dynamic scene with the event camera, we get continuous clear video frames and their asynchronous event stream signals, and let one of the clear video frames be denoted as I gt , let the asynchronous event stream signal corresponding to the clear video frame I be recorded as Among them, e k represents the kth event, and e k =p k δ(tt k , xx k ,yy k ), p k Represents the kth event e k Polarity, p k ∈{-1, 1}; t kRepresents the kth event e k The time of occurrence; x k and y k Represents the kth event e k The spatial coordinates of occurrence; N represents the total number of events; (t, x, y) represents the time-space projection coordinates; δ is the indicator function;

[0069] The continuous clear video frames are averaged as the clear video frame I gt The corresponding frame is blurred video frame B; and the blurred video frame B is converted into a blur degree BL using Laplace transform function Laplace();

[0070] In this example, the RealSharD-DAVIS dataset was used for training and evaluation. To simulate blind exposure settings, the shutter period is denoted by T and the exposure mode is denoted by mn, where m represents the exposure time, n represents the data readout time, and m + n = T. To simulate blurred video, the blurred video frames are generated by averaging the clear frames within the exposure time. In the experiments, T was set to 16 and 12 to simulate different exposure conditions. When T = 16, the exposure time m was set to 9 to 15. When T = 12, the exposure time m was set to 7 to 11.

[0071] Step 2: Use formula (1) to convert the asynchronous event stream signal Convert to event frame V∈R 2×H×W :

[0072] V=∑ k＝1 δ(xx k ,yy k ,pp k ) (1)

[0073] In formula (1), H and W represent the height and width of the event frame V respectively;

[0074] Step 3: Construct a joint de-motion blurring and video interpolation network under blind exposure constraints, including: a feature extraction module, an event-guided exposure prior estimation module, a joint time-exposure control module, and a clear video frame reconstruction module;

[0075] Step 3.1: The feature extraction module uses a convolutional layer to process the event frame V and the blurred video frame B respectively, and obtains the event features accordingly. and blurred video frame features

[0076] Step 3.2, such as Figure 2 As shown, the event-guided exposure prior estimation module contains two feature extraction networks with the same structure but no shared parameters. and a feature aggregation network Φ aggre;

[0077] The feature extraction network consists of a convolutional layer and a LeakyReLU activation function;

[0078] Feature aggregation network Φ aggre It consists of two convolutional layers. The activation function of the first convolution layer is LeakyReLU, and the second convolution layer has no activation function.

[0079] Step 3.2.1: Event frame V is input into the feature extraction network And process, and output event feature Feat e ;

[0080] The frame blur video frame B and the blur degree BL are stacked in the channel dimension and input into the feature extraction network Process it and output the fuzzy video feature Feat bf ;

[0081] Step 3.2.2, Feature Aggregation Network Φ aggre Using formula (5) to get the aggregated feature Feat aggre :

[0082] Feat aggre =Φ aggre ([Feat e ×Sigmoid(GN(Feat e )×GN(Feat bf )), Feat bf ]) (5)

[0083] In formula (5), Sigmoid() represents the sigmoid activation function, and GN() represents the group norm normalization process;

[0084] Feature aggregation network Φ aggre The predicted exposure prior EP is obtained using formula (6):

[0085] EP=AP(Feat aggre ) (6)

[0086] In formula (6), EP ranges from 0 to 1, and AP represents average pooling;

[0087] Step 3.3, such as Figure 3 As shown, the time-exposure joint control module consists of n cascaded submodules, each of which consists of two feature transfer convolutional layers Two control convolutional layers And a fused convolutional layer Φ fuse composition;

[0088] Among them, the i-th submodule uses formula (7) to obtain the i-th intermediate control feature Feat i , so that the nth cascaded submodule outputs the joint control feature Feat control ;

[0089]

[0090] In formula (7), Feat i-1 represents the intermediate control feature output by the i-1th submodule; when i=1, let i∈[1,N];

[0091] Step 3.4: The clear video frame reconstruction module consists of the initial clear frame prediction submodule and the final clear frame recovery submodule Φ detail ;

[0092] Step 3.4.1, the initial clear frame prediction submodule consists of the kernel prediction module KPN, the convolution module, the connection module and the image reconstruction module Φ recons constitute;

[0093] Convolutional module for joint control feature Feat control Perform convolution processing to obtain the convolution feature Feat′ control ;

[0094] The kernel prediction module KPN uses formula (9) to obtain the initial convolution feature

[0095]

[0096] In formula (9), Conv() represents the processing of convolution layer and LeakyReLU activation function, * represents convolution;

[0097] In this embodiment, the size of the convolution kernel predicted by the kernel prediction module KPN is 5×5.

[0098] The connection module uses formula (10) to obtain the initial clear frame features

[0099]

[0100] Image reconstruction module Φ recons It consists of a convolution layer, a PixelShuffle layer and a LeakyReLU activation function, and After processing, the initial clear frame I is obtained init ;

[0101] Step 3.4.2, final clear frame recovery submodule Φ detailWill I init After splicing with B, the final clear frame I is obtained final ;

[0102] Step 4: Optimize the network using a phased training strategy:

[0103] Step 4.1: Use formula (10) to construct the loss function of the first-stage event-guided exposure prior estimation module

[0104]

[0105] In formula (10), EP gt is the true exposure prior; represents the mean square error loss function;

[0106] Step 4.2: Use formula (11) to construct the loss function of the remaining module in the second stage

[0107]

[0108] In formula (11), Represents the network loss function, and has:

[0109]

[0110] In this embodiment, the weights λ=1, μ=0.1 before 10K iterations, and are then set to λ=0.1, μ=1.

[0111] In formula (12), represents the composite loss function, I gt Indicates a true clear frame, with:

[0112]

[0113] In formula (13), represents the Laplace loss function, Denotes the Hamming loss function, and has:

[0114]

[0115]

[0116] In formula (14) and formula (15), I represents I init or I final , S represents the number of layers, L s () represents the Laplace pyramid of level s, Dis() represents the Hamming distance, and Cen() represents the census transform;

[0117] In this embodiment, S is set to 5.

[0118] Step 4.3: Use a phased training strategy and gradient descent to train the joint de-motion blurring and video interpolation network, and calculate the loss function:

[0119] In the first stage, the event-guided exposure prior estimation module is optimized using formula (10);

[0120] The second stage uses formula (11) to optimize the remaining modules;

[0121] Finally, use formula (11) to fine-tune the entire network.

[0122] In this embodiment, a learning rate of 0.0001 is used in the first and second stages; and a learning rate of 0.00001 is used in the final fine-tuning stage.

[0123] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0124] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are executed.

[0125] Experimental example:

[0126] To verify the effectiveness of the method of the present invention, this embodiment selected the commonly used RealSharp-DAVIS dataset for training and testing. The method is trained based on the RealSharp-DAVIS dataset. The RealSharp-DAVIS dataset is divided into a training set and a test set. The entire dataset contains 59 different scenes, captured using a DAVIS-color-346 event camera with a resolution of 346×260. During training, the exposure time was set to 9 to 15 and the shutter period was 16. Then, a generalization test was performed. During the test, the exposure time was set to 7 to 11 and the shutter period was 12.

[0127] In this example, four common methods, EVDI, LEDVDI, UTI, and BIN, are selected for comparison, and PSNR is selected as the evaluation indicator, as shown in Table 1:

[0128]

[0129]

[0130] Experimental results demonstrate that our method outperforms the best previous restoration model for joint motion deblurring and video interpolation based on event cameras, validating its ability to effectively address video restoration tasks under blind exposure constraints. Furthermore, our method significantly outperforms other methods in inference time, demonstrating its high efficiency.

Claims

1. A joint de-motion blurring and video interpolation method based on event camera, characterized in that: The steps include: Step 1: Collection and preprocessing of low-frame-rate blurry video data from event cameras: After shooting the dynamic scene with the event camera, we get continuous clear video frames and their asynchronous event stream signals, and let one of the clear video frames be denoted as I gt , let the asynchronous event stream signal corresponding to the clear video frame I be recorded as Among them, e k represents the kth event, and e k =p k δ(tt k ,xx k ,yy k ), p k Represents the kth event e k Polarity, p k ∈{-1,1}; t k Represents the kth event e k The time of occurrence; x k and y k Represents the kth event e k The spatial coordinates of occurrence; N represents the total number of events; (t, x, y) represents the time-space projection coordinates; δ is the indicator function; The continuous clear video frames are averaged as the clear video frame I gt The corresponding frame is blurred video frame B; and the blurred video frame B is converted into a blur degree BL using Laplace transform function Laplace(); Step 2: Use formula (1) to convert the asynchronous event stream signal Convert to event frame V∈R 2×H×W : V=∑ k＝1 δ(xx k ,yy k ,pp k ) (1) In formula (1), H and W represent the height and width of the event frame V respectively; Step 3: Construct a joint de-motion blurring and video interpolation network under blind exposure constraints, including: a feature extraction module, an event-guided exposure prior estimation module, a joint time-exposure control module, and a clear video frame reconstruction module; Step 3.1: The feature extraction module processes the event frame V and the blurred video frame B using a convolution layer to obtain the event features. and blurred video frame features Step 3.2: The event-guided exposure prior estimation module processes the event frame V to obtain a predicted exposure prior EP; Step 3.3: The time-exposure joint control module processes the exposure prior EP to obtain the joint control feature Feat control ; Step 3.4: The clear video frame reconstruction module performs joint control feature Feat control Processing is performed to obtain the initial clear frame I init And the final clear frame I final ; Step 4: Optimize the network using a phased training strategy: Step 4.1: Use formula (10) to construct the loss function of the first-stage event-guided exposure prior estimation module In formula (10), EP gt is the true exposure prior; represents the mean square error loss function; Step 4.2: Use formula (11) to construct the loss function of the remaining module in the second stage In formula (11), Represents the network loss function, and has: In formula (12), represents the composite loss function, and has: In formula (13), represents the Laplace loss function, Denotes the Hamming loss function, and has: In formula (14) and formula (15), I represents I init or I final , S represents the number of layers, L s () represents the Laplace pyramid of level s, Dis() represents the Hamming distance, and Cen() represents the census transform; Step 4.3: Use the gradient descent method to train the joint demotion blurring and video interpolation network in stages. In the first stage, use formula (10) to optimize the event-guided exposure prior estimation module. In the second stage, use formula (11) to optimize the remaining modules and fine-tune the network parameters to obtain the trained joint demotion blurring and video interpolation network, which is used to realize the deblurring and interpolation processing of any low-frame-rate blurred video frame.

2. The method for joint de-motion blurring and video interpolation based on event camera according to claim 1, characterized in that: The event-guided exposure prior estimation module contains two feature extraction networks with the same structure but no shared parameters. and a feature aggregation network Φ aggre ; The feature extraction network consists of a convolutional layer and a LeakyReLU activation function; The feature aggregation network Φ aggre It consists of two convolutional layers. The activation function of the first convolution layer is LeakyReLU, and the second convolution layer has no activation function. Step 3.2.1: The event frame V is input into the feature extraction network Process it and output the event feature Feat e ; The frame blurred video frame B and the blur level BL are stacked in the channel dimension and then input into the feature extraction network Process it and output the fuzzy video feature Feat bf ; Step 3.2.2: The feature aggregation network Φ aggre Using formula (5) to get the aggregated feature Feat aggre : Feat aggre =Φ aggre ([Feat e ×Siamoid(GN(Feat e )×GN(Feat bf )),Feat bf ]) (5) In formula (5), Sigmoid() represents the sigmoid activation function, and GN() represents the groupnorm normalization process; The feature aggregation network Φ aggre The predicted exposure prior EP is obtained using formula (6): EP=AP(Feat aggre ) (6) In formula (6), EP ranges from 0 to 1, and AP represents average pooling.

3. The method for joint de-motion blurring and video interpolation based on event camera according to claim 2, characterized in that: The time-exposure joint control module includes n cascaded submodules, each of which consists of two feature transfer convolutional layers Two control convolutional layers And a fused convolutional layer Φ fuse composition; Among them, the i-th submodule uses formula (7) to obtain the i-th intermediate control feature Feat i , so that the nth cascaded submodule outputs the joint control feature Feat control ; In formula (7), Feat i-1 represents the intermediate control feature output by the i-1th submodule; when i=1, let 4. The method for joint de-motion blurring and video interpolation based on event camera according to claim 3, characterized in that: The clear video frame reconstruction module consists of an initial clear frame prediction submodule and a final clear frame recovery submodule Φ detail ; Step 3.4.1, the initial clear frame prediction submodule consists of a convolution module, a kernel prediction module KPN, a connection module and an image reconstruction module Φ recons constitute; The convolution module is used to jointly control the feature Feat control Perform convolution processing to obtain the convolution feature Feat' control ; The kernel prediction module KPN uses formula (9) to obtain the initial convolution feature In formula (9), Conv() represents the processing of convolution layer and LeakyReLU activation function, * represents convolution; The connection module uses formula (10) to obtain the initial clear frame feature The image reconstruction module Φ recons It consists of a convolution layer, a PixelShuffle layer and a LeakyReLU activation function, and After processing, the initial clear frame I is obtained init ; Step 3.4.2, the final clear frame recovery submodule Φ detail Will I init After splicing with B, the final clear frame I is obtained final .

5. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the joint de-motion blur and video interpolation method according to any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the joint de-motion blurring and video interpolation method according to any one of claims 1 to 4 are executed.

Images