A Motion Blur Detection Method and System Based on Deep Learning and Blind Deconvolution

Abstract

The present invention relates to the field of blur detection technology, and discloses a motion blur detection method and system based on deep learning and blind deconvolution, including: collecting an original image and performing sub-region division of the image; performing singular value decomposition on each sub-region respectively; detecting the blur region of the sub-region according to the decomposition result; performing image restoration on the blurred image; using the method of deep learning to segment the target contour information in the blurred restored image, and outputting a clear image with the target contour. By introducing SPD-Conv, the improved YOLOv8 model can more accurately segment low-resolution blurred targets, effectively improving the segmentation accuracy. It can adapt to the low-resolution blurred target segmentation tasks in different scenarios. While improving the segmentation accuracy, the method proposed in the present invention does not significantly increase the computational complexity and inference time of the model, and still maintains the real-time advantage of YOLOv8, meeting the requirements for real-time performance in practical applications.

Description

Technical Field

[0001] The present invention relates to the technical field of fuzzy detection, and specifically to a motion blur detection method and system based on deep learning and blind deconvolution. Background Art

[0002] With the growing global demand for clean energy, wind power generation, as a renewable energy technology, has been widely developed. As the core equipment of wind power generation, the safe operation of wind turbines is of crucial importance. In the wind turbine clearance detection project, accurately detecting the state of the wind turbine blades is one of the key links to ensure the safe operation of the wind turbine.

[0003] To achieve all-weather safety monitoring of the wind turbine clearance, it is necessary to segment the contour of the wind turbine blades in real time in low-light environments. To obtain relatively clear images, it is usually necessary to extend the exposure time of the camera to obtain target features. Due to the high rotational speed of the wind turbine blades, it is inevitable that serious motion blur occurs in the imaging of the wind turbine blades. This kind of blur causes great damage to the image quality, making the contour of the wind turbine blades unclear and greatly reducing the segmentation accuracy.

[0004] Existing related techniques for removing motion blur commonly include inverse filtering, Wiener filtering, the L-R algorithm, and some deep neural networks such as DeblurGAN, etc. Inverse filtering and Wiener filtering are non-blind deconvolution of data, which requires obtaining the blur kernel of the image and has poor blur restoration effect for images with large noise. The L-R algorithm needs to iterate continuously to remove image blur. If the original image has a large guarantee, it will lead to extremely poor quality of the final image restoration. And DeblurGan, as a new star in the field of deblurring in deep learning methods, has a relatively high quality of restoring blurred images, but the method is difficult to reproduce. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is that existing methods for restoring blurred images have problems such as difficult reproduction and inability to accurately segment moving parts.

[0007] To solve the above technical problem, the present invention provides the following technical solution: A motion blur detection method based on deep learning and blind deconvolution, including:

[0008] Collecting the original image and performing sub-region division of the image;

[0009] Performing singular value decomposition on each sub-region respectively;

[0010] Detecting the blurred region of the sub-region according to the decomposition result;

[0011] Restoring the blurred image;

[0012] Use the method of deep learning to segment the target contour information in the blurred restored image and output a clear image with the target contour.

[0013] As a preferred solution of the motion blur detection method based on deep learning and blind deconvolution according to the present invention, wherein: the sub-region division includes dividing the original input image into sub-regions of a fixed size.

[0014] As a preferred solution of the motion blur detection method based on deep learning and blind deconvolution according to the present invention, wherein: the singular value decomposition includes, assuming that the image I is an m×n image, then U is an m×m orthogonal matrix, and its column vectors are left singular vectors; V is an n×n orthogonal matrix, and its column vectors are right singular vectors; S is an m×n diagonal matrix, and the elements on the diagonal are non-negative real numbers and the singular values arranged in descending order;

[0015] For the singular value decomposition of the blurred image I, it is expressed as:

[0016] ;

[0017] wherein, i represents the diagonal index, i∈{1, 2,..., min(m, n)}; T represents the transpose.

[0018] As a preferred solution of the motion blur detection method based on deep learning and blind deconvolution according to the present invention, wherein: the blurred region detection includes, for each sub-region perform singular value decomposition and extract the singular values ; sort the singular value information in descending order, and assume that the singular values corresponding to the main feature information in the image are k ;

[0019] Calculate the ratio sum of each value in the singular value information to the initial value:

[0020] ;

[0021] wherein, is the quality evaluation score of the image region ; j represents the index of the diagonal line in the diagonal matrix of the sub-region; represents the decomposed image after the feature information, the first k eigenvalues are the singular values of the main texture features; represents the singular value corresponding to the key information part of the image;

[0022] Perform Gaussian filtering blur on the image region to generate a reference image and calculate the quality evaluation score ;

[0023] The change amount of the image singular value is:

[0024] ;

[0025] When diff is greater than the threshold , it is determined that the image sub-region belongs to the blurred region.

[0026] As a preferred embodiment of the motion blur detection method based on deep learning and blind deconvolution according to the present invention, wherein: the image restoration includes assuming that the image is formed by convolving a clear image with a blur kernel and adding noise;

[0027]

[0028] wherein, g represents the blurred image, h represents the high-definition image, f(x) represents the blur kernel, and n(x) represents the noise;

[0029] The blind deconvolution method based on MLE constructs a likelihood function to obtain an optimal combination of the clear image and the blur kernel, making the likelihood function reach the maximum value;

[0030] Assume that the noise n(x) follows a Gaussian distribution with a mean of 0 and a variance of , and the probability density function of observing the blurred image :

[0031]

[0032] The goal based on MLE is to maximize the likelihood function to obtain the log-likelihood function:

[0033]

[0034] wherein, C is a constant, and the equation is equivalent to least squares minimization .

[0035] As a preferred embodiment of the motion blur detection method based on deep learning and blind deconvolution according to the present invention, wherein: the specific steps of performing image restoration on the blurred image are as follows:

[0036] Select an initial value to guess the blur kernel f(x) and the clear image g;

[0037] Fix h and update f(x): Assume h is fixed as , then solve ;

[0038] Determine whether the algorithm converges. If the convergence condition is not met, continue the iteration; otherwise, output the estimated clear image h and the blur kernel f(x).

[0039] As a preferred solution of the motion blur detection method based on deep learning and blind deconvolution according to the present invention, wherein: the method of using deep learning to segment the target contour information in the blurred restoration image includes identifying the target contour of the restored image through the improved yolov8 and outputting the feature map;

[0040] The improved yolov8 includes inserting an SPD module on the basis of yolov8 and replacing conv-c2f of the backbone with spd-c2f.

[0041] A motion blur detection system based on deep learning and blind deconvolution adopting the method according to any one of the present invention, characterized in that:

[0042] An acquisition module for acquiring the original image and performing sub-region division of the image;

[0043] A detection module for performing singular value decomposition on each sub-region respectively; and detecting the blurred region of the sub-region according to the decomposition result;

[0044] A processing module for restoring the blurred image;

[0045] An output module for segmenting the target contour information in the blurred restoration image by using the method of deep learning and outputting a clear image with the target contour.

[0046] A computer device, comprising: a memory and a processor; the memory stores a computer program, wherein: when the processor executes the computer program, the steps of the method according to any one of the present invention are implemented.

[0047] A computer-readable storage medium, on which a computer program is stored, wherein: when the computer program is executed by a processor, the steps of the method according to any one of the present invention are implemented.

[0048] Advantages of the present invention: The motion blur detection method based on deep learning and blind deconvolution provided by the present invention can more accurately segment low-resolution blurred targets by introducing SPD-Conv. The improved YOLOv8 model effectively improves the segmentation accuracy, providing a more reliable basis for subsequent target recognition, analysis, and processing. Experimental results on various low-resolution blurred image datasets show that the method proposed in the present invention has good generalization ability and can adapt to low-resolution blurred target segmentation tasks in different scenarios. While improving the segmentation accuracy, the method proposed in the present invention does not significantly increase the computational complexity and inference time of the model, and still maintains the real-time advantage of YOLOv8, meeting the requirements for real-time performance in practical applications. Description of the Drawings

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0050] Figure 1 It is the overall flowchart for detecting the fan blade in the motion blur detection method based on deep learning and blind deconvolution provided by the first embodiment of the present invention;

[0051] Figure 2 It is the comparison diagram of blur detection for the motion blur detection method based on deep learning and blind deconvolution provided by the second embodiment of the present invention; where (a) represents the image before region detection, and (b) represents the image after region detection;

[0052] Figure 3 It is the comparison diagram of blurred image restoration for the motion blur detection method based on deep learning and blind deconvolution provided by the second embodiment of the present invention; where (a) represents the original image before restoration, and (b) represents the restored image;

[0053] Figure 4 It is the comparison diagram of target segmentation for the motion blur detection method based on deep learning and blind deconvolution provided by the second embodiment of the present invention; where (a) represents the image before segmentation, and (b) represents the segmentation result. Detailed Embodiments

[0054] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following provides a detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0055] Example 1. Refer to Figure 1 , which is an embodiment of the present invention, and provides a motion blur detection method based on deep learning and blind deconvolution, including:

[0056] S1: Collect the original image and perform sub-region division of the image.

[0057] First, the original input image is divided into several small sub-regions (for example, 9×9). This is because directly performing singular value decomposition on the original image involves a large amount of calculation, takes a long time, and the blur degree and blur type of each region of the image are different. Dividing into several small sub-regions can not only accelerate the singular value decomposition process of the image data but also process data with different blur types and different blur degrees, avoiding the loss of blur features.

[0058] S2: Perform singular value decomposition on each sub-region separately.

[0059] It should be noted that a blurred image is essentially composed of an original high-definition image and a blur kernel. The blur kernel will smooth the high-frequency features in the image, such as the contour features of objects, making them disappear. Image data is actually a set of pixel matrices, and the contour features in the image can be manifested from the numerical distribution features of the pixel matrices. The singular value distribution of the pixel matrix of a clear image has relatively obvious features. Generally speaking, the decay of its singular values is relatively slow. For example, for a natural landscape image with rich details and clear edges, the first few singular values will be relatively large because these singular values mainly correspond to the main structures in the image (such as the horizon, mountain contours, etc.) and the high-energy parts. As the serial number of the singular values increases, the singular values will gradually decrease but will not quickly approach zero. This is because a clear image contains various frequency components, from the large-area color regions of low frequency to the fine textures and edges of high frequency, and all these information is reflected in the singular value decomposition.

[0060] Blur will cause changes in the singular value distribution of the image pixel matrix. First of all, the blurring operation is similar to a low-pass filter, which will cause the image to lose high-frequency information. In terms of the singular value distribution, it is manifested as a rapid decay of the singular values.

[0061] The singular value decomposition of the blurred image I can be expressed as:

[0062]

[0063] Assume that the image I is an m×n image, then U is an m×m orthogonal matrix whose column vectors are left singular vectors; V is an n×n orthogonal matrix whose column vectors are right singular vectors;

[0064]

[0065] S is an m×n diagonal matrix, and the elements on its diagonal are non - negative real numbers and singular values arranged in descending order, .

[0066] S3: According to the decomposition result, perform fuzzy region detection on the sub - regions.

[0067] For each sub - region perform SVD (Singular Value Decomposition) and extract its singular value information , where the first singular value information corresponds to the main feature information in the image, such as contour features; the rest corresponds to the detail information of the image. Gaussian blur will reduce the detail features of the image, that is, increase the singular values. Compared with a clear image, the proportion of the first main feature information in the blurred image is relatively large. Therefore, calculating the ratio of each value in the singular value information to the initial value can be used as an evaluation of the image blur score:

[0068] ;

[0069] where, is the quality evaluation score of the image region ; j represents the index of the diagonal in the diagonal matrix of the sub - region; represents the singular values of the main texture features among the first feature values sorted in descending order after decomposing the feature information of the decomposed image , and the rest are detail features;

[0070] According to the singular value decomposition theory, the image I can be understood as first being rotated by and then scaled in the new coordinate system by , and then rotated back by . is the maximum scaling degree in a certain direction during this optimal (in a certain mathematical sense, such as the transformation that minimizes the energy loss) transformation process, carrying the most important, overall key information or the strongest energy component of the image.

[0071] Perform Gaussian filtering blur (Gaussian kernel is 3, filled with normal distribution) on the image region to generate a reference image and calculate the quality evaluation score .

[0072] The variation of the singular value of the image is as follows:

[0073]

[0074] When diff is greater than the threshold , then the current image sub-region belongs to the blurred region.

[0075] S4: Image restoration is performed on the blurred image.

[0076] Assume that the image is formed by convolving a clear image with a blur kernel and then adding noise:

[0077]

[0078] Among them, g represents the blurred image, h represents the high-definition image, f(x) represents the blur kernel, and n(x) represents the noise.

[0079] The blind deconvolution method based on MLE (Maximum Likelihood Estimation) constructs a likelihood function to find the most likely combination of the clear image and the blur kernel, so that the likelihood function reaches the maximum value given the observed blurred image . The likelihood function is usually constructed based on the probability distribution assumption of the noise. For example, assuming that the noise is Gaussian distributed, then the likelihood function can be constructed according to the probability density function of the Gaussian distribution.

[0080] Assume that the noise n(x) follows a Gaussian distribution with a mean of 0 and a variance of , then the probability density function of observing the blurred image is:

[0081]

[0082] The goal based on MLE is to maximize the likelihood function to obtain the log-likelihood function:

[0083]

[0084] Among them, C is a constant, and the equation is equivalent to least squares minimization .

[0085] The steps of blurred image restoration include:

[0086] Select appropriate initial values to guess the blur kernel f(x) and the clear image g. The initialization parameter values will affect the convergence speed and the final result of the algorithm.

[0087] Fix h and update f(x): Assume that h is fixed as , then solve .

[0088] Determine whether the algorithm converges. The convergence condition can be that the change in the log-likelihood function is less than a certain threshold, or the relative change in f(x) is less than a certain threshold, etc. If the convergence condition is not met, continue the iteration; otherwise, output the estimated clear image h and the blur kernel f(x).

[0089] S5: Use deep learning methods to segment the target contour information in the blurred restored image and output a clear image with the target contour.

[0090] Use deep learning methods to segment the target contour information in the blurred restored image. Insert an SPD module on the basis of yolov8 (replace conv-c2f in the backbone with spd-c2f), introduce global context and local feature information, improve the feature extraction ability of the model for low-resolution and small contours, and improve the robustness in complex scenarios.

[0091] Meanings of each layer of the SPD module:

[0092] (a) shows the standard feature map.

[0093] (b) shows the space-to-depth operation, where the spatial information is rearranged into the depth channels.

[0094] (c) shows the increase in the depth of the resulting feature map.

[0095] (d) represents the non-strided convolutional layer applied after the SPD operation.

[0096] (e) shows the output feature map after convolution with a stride of 1, which maintains the spatial resolution but changes the depth dimension.

[0097] On the other hand, this embodiment also provides a motion blur detection system based on deep learning and blind deconvolution, which includes:

[0098] An acquisition module that acquires the original image and performs sub-region division of the image.

[0099] A detection module that performs singular value decomposition on each sub-region respectively; according to the decomposition result, detects the blurred region of the sub-region.

[0100] A processing module that restores the blurred image.

[0101] An output module that uses deep learning methods to segment the target contour information in the blurred restored image and outputs a clear image with the target contour.

[0102] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0103] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

[0104] More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection parts with one or more wirings (electronic devices), portable computer disk cartridges (magnetic devices), random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memories), fiber optic devices, and portable compact disc read-only memories (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or otherwise processing it as necessary, and then storing it in a computer memory.

[0105] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits with logic gate circuits for implementing logic functions on data signals, application specific integrated circuits with suitable combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0106] Embodiment 2, such as Figures 2 - 4 is an embodiment of the present invention, which provides an implementation process of a motion blur detection method based on deep learning and blind deconvolution.

[0107] (1) Data preprocessing: Preprocess the input low-resolution blurred image, including operations such as image normalization, cropping, and scaling, to make it meet the requirements of network input.

[0108] (2) Model training: Use a labeled low-resolution blurred target image dataset to train the improved YOLOv8 model. During the training process, adjust the parameters of the model, such as learning rate, weight decay, etc., to optimize the performance of the model. At the same time, data augmentation techniques, such as rotation, flipping, adding noise, etc., can be adopted to increase the diversity of the dataset and improve the generalization ability of the model.

[0109] (3) Model inference: Input the low-resolution blurred image to be segmented into the trained model, and obtain the segmentation result of the target through forward propagation calculation. During the inference process, post-processing can be performed on the output result according to actual needs, such as removing noise, filling holes, etc., to obtain a more accurate and smoother segmentation mask.

[0110] (4) Performance evaluation: Use metrics such as accuracy, recall rate, and mean average precision (mAP) to evaluate the segmentation performance of the model. Through comparative experiments with the original YOLOv8 model and other existing segmentation methods, verify the effectiveness of the method proposed in the present invention in improving the segmentation accuracy of low-resolution blurred targets.

[0111] After performing singular value decomposition on the blurred image I, perform blurred region detection to obtain the blurred region, such as Figure 2 shown. The left side represents the image before region detection; the right side is the image after region detection, where the red part represents the detected background part, and the other parts are the detected blurred regions.

[0112] After completing the detection, restore the blurred part, such as Figure 3As shown, the left side represents the original image before restoration, and the right side represents the restored image. It can be seen that the present invention can effectively clarify blurred images.

[0113] Perform object segmentation on the clarified image, such as Figure 4 As shown, the left side represents before segmentation, and the right side represents the segmentation result. The red color represents the target object, realizing the blurred detection of moving objects.

[0114] The core of the present invention lies in introducing SPD-Conv into the YOLOv8 network structure. By processing the input feature map through SPD-Conv, the feature extraction ability of the model for low-resolution blurred targets is effectively improved, thereby enhancing the segmentation accuracy.

[0115] Principle and advantages of SPD-Conv: SPD-Conv consists of a space-to-depth (SPD) layer and a non-strided convolution (Conv) layer. The SPD layer reduces each spatial dimension of the input feature map to the channel dimension while retaining the information within the channels, reducing the spatial dimension size; the Conv layer performs standard convolution operations after the SPD layer, avoiding the over-sampling problem that may be caused by strided convolution and retaining more fine-grained information. This combination method can improve the detection performance of the model for low-resolution images and small objects and reduce the dependence on high-quality inputs.

[0116] Improved YOLOv8 network structure: In the backbone network and / or neck network of YOLOv8, replace some traditional convolution layers with SPD-Conv modules. Specifically, according to the characteristics and requirements of the object segmentation task, select a suitable position to insert the SPD-Conv module to give full play to its advantages and enhance the network's ability to capture features of low-resolution blurred targets.

[0117] Principle of improved segmentation accuracy: The feature information of low-resolution blurred targets is relatively weak and difficult to extract. Traditional convolution layers may lose important information when processing such targets. However, SPD-Conv can better retain and utilize this feature information. By converting spatial information into depth information, the network can more comprehensively learn the feature representation of the target, thereby more accurately segmenting low-resolution blurred targets.

[0118] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not restrictive. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A motion blur detection method based on deep learning and blind deconvolution, characterized in that, Including: Collect the original image and perform sub-region division of the image; Perform singular value decomposition on each sub-region separately; Detect the blurred region of the sub-region according to the decomposition result; Restore the blurred image; Use deep learning methods to segment the target contour information in the blurred restored image and output a clear image with the target contour; The image restoration includes assuming that the image is formed by convolving a clear image with a blur kernel and adding noise, and forming a clear image through iteration; ， Among them, g represents the blurred image, h represents the high-definition image, f(x) represents the blur kernel, and n(x) represents the noise; The MLE-based blind deconvolution method constructs a likelihood function to obtain the optimal combination of a clear image and a blur kernel, maximizing the likelihood function to reach the maximum value; Let the noise \(n(x)\) follow a Gaussian distribution with a mean of 0 and a variance of , and the probability density function of observing the blurred image is: ， The goal based on MLE is to maximize the likelihood function to obtain the log-likelihood function: ， where C is a constant and the equation is equivalent to least squares minimization .

2. The motion blur detection method based on deep learning and blind deconvolution according to claim 1, characterized in that: The sub-region division includes dividing the original input image into sub-regions of a fixed size.

3. The motion blur detection method based on deep learning and blind deconvolution according to claim 2, wherein: The singular value decomposition includes: Let the image I be an m×n image. Then U is an m×m orthogonal matrix, and its column vectors are left singular vectors; V is an n×n orthogonal matrix, and its column vectors are right singular vectors; S is an m×n diagonal matrix, and the elements on the diagonal are non-negative real numbers and singular values arranged in descending order; The singular value decomposition of the blurred image I is expressed as: ； Among them, i represents the diagonal index, i ∈ {1, 2,..., min(m, n)}; T represents the transpose.

4. The motion blur detection method based on deep learning and blind deconvolution according to claim 3, wherein: The fuzzy region detection includes, for each sub-region performing singular value decomposition and extracting singular values ; sorting the singular value information in descending order, and setting the singular values corresponding to the main feature information in the image to be k ; Calculate the sum of the ratios of each value in the singular value information to the initial value; ； Among them, is the quality evaluation score of the image region ; j represents the index of the diagonal in the diagonal matrix of the sub-region; represents that after decomposing the feature information of the image , the first k eigenvalues are the singular values of the main texture features; represents the singular value corresponding to the key information part of the image. For the image region perform Gaussian filtering blur to generate a reference image and calculate the quality evaluation score ; The change amount of the image singular value is: ； When diff is greater than the threshold , it is determined that the image sub-region belongs to the blurred region.

5. The motion blur detection method based on deep learning and blind deconvolution according to claim 4, characterized in that: The specific steps for restoring the blurred image are as follows: Select an initial value to guess the blur kernel f(x) and the clear image g; Fix h and update f(x): Let h be fixed as , then solve ; Judge whether the algorithm converges. If the convergence condition is not satisfied, continue the iteration; otherwise, output the estimated clear image h and the blur kernel f(x).

6. The motion blur detection method based on deep learning and blind deconvolution according to claim 5, characterized in that: The use of deep learning methods to segment the target contour information in the blurred restored image includes identifying the target contour of the restored image through the improved yolov8 and outputting a feature map; The improved yolov8 includes inserting an SPD module on the basis of yolov8 and replacing conv-c2f of the backbone with spd-c2f.

7. A motion blur detection system based on deep learning and blind deconvolution using the method according to any one of claims 1-6, characterized in that: A collection module that collects the original image and performs sub-region division of the image; A detection module that performs singular value decomposition on each sub-region separately; Detect the blurred region of the sub-region according to the decomposition result; A processing module that restores the blurred image; An output module that uses deep learning methods to segment the target contour information in the blurred restored image and outputs a clear image with the target contour.

8. A computer device, comprising: A memory and a processor; The memory stores a computer program, characterized in that: when the processor executes the computer program, the steps of the method according to any one of claims 1-6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the steps of the method according to any one of claims 1-6 are implemented.

Images