Image blurriness evaluation method and device, computer device and medium

Abstract

This invention relates to the field of image recognition technology, and more particularly to an image ambiguity assessment method, apparatus, computer device, and medium. The method involves acquiring an image dataset including at least one image to be assessed; extracting features from all images to be assessed using a backbone network in a ambiguity determination model to obtain assessment image features; evaluating all assessment image features using the boundary distance method in the ambiguity determination model to obtain image assessment values; and filtering the images to be assessed based on all image assessment values ​​to obtain target images. The backbone network in this invention is constructed based on an inverse residual network structure, enabling the extraction of more features from the images to be assessed, thus enriching the assessment image features. Furthermore, the boundary distance method is used to evaluate the assessment image features, enabling the determination of image assessment values, thereby achieving accurate judgment of the ambiguity degree of the images to be assessed and improving the accuracy of the ambiguity determination model.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to an image blur assessment method, apparatus, computer device, and medium. Background Technology

[0002] With the rapid development of image recognition technology, various algorithms based on data, models, and loss functions have emerged. However, in certain specific application scenarios, such as face recognition algorithms, the input data is one of the most important factors affecting algorithm performance; the quality of the input image has a significant impact on model performance. Similarly, in some archiving tasks, some low-quality image data may appear during the capture process, requiring the exclusion of blurry image data to improve the performance of the archived data.

[0003] Traditional gradient-function-based fuzzy image recognition algorithms are not very accurate in identifying face images with varying degrees of blur. While this method can effectively identify most images with high blur levels, it is prone to errors when judging images with relatively low blur levels. For example, it may mistake slightly blurry images for clear ones, resulting in low accuracy in assessing image blur levels and thus degrading the performance of archived data during image filtering.

[0004] Therefore, improving the accuracy of image blur assessment has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide an image blur assessment method, apparatus, computer equipment, and medium to solve the problem of low accuracy in image blur assessment.

[0006] In a first aspect, embodiments of the present invention provide an image blur assessment method, the image blur assessment method comprising:

[0007] Obtain an image dataset, which includes at least one image to be evaluated;

[0008] A fuzzy decision model is obtained, and features are extracted from all the images to be evaluated through the backbone network in the fuzzy decision model to obtain the evaluation image features corresponding to the images to be evaluated; the backbone network is constructed based on the inverted residual structure.

[0009] Image evaluation is performed on all the features of the evaluation image using the boundary distance method in the fuzzy decision model to obtain the image evaluation value corresponding to the image to be evaluated.

[0010] The images to be evaluated are filtered based on all the image evaluation values ​​to obtain the target image.

[0011] Secondly, embodiments of the present invention provide an image blur assessment device, the image blur assessment device comprising:

[0012] The data acquisition module is used to acquire an image dataset, which includes at least one image to be evaluated;

[0013] The feature extraction module is used to obtain a fuzzy decision model, and to extract features from all the images to be evaluated through the backbone network in the fuzzy decision model to obtain the evaluation image features corresponding to the images to be evaluated; the backbone network is constructed based on the inverted residual structure;

[0014] The image evaluation module is used to evaluate all the features of the evaluation image using the boundary distance method in the fuzzy decision model, and obtain the image evaluation value corresponding to the image to be evaluated.

[0015] The image filtering module is used to filter the images to be evaluated based on all the image evaluation values ​​to obtain the target image.

[0016] Thirdly, embodiments of the present invention provide a computer device, the computer device including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the image blur assessment method as described in the first aspect.

[0017] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the image blur assessment method as described in the first aspect.

[0018] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows:

[0019] An image dataset is acquired, including at least one image to be evaluated. A fuzzy decision model is obtained, and features are extracted from all images to be evaluated using the backbone network of the fuzzy decision model to obtain evaluation image features corresponding to the images to be evaluated. The backbone network is constructed based on an inverted residual structure. Image evaluation is performed on all evaluation image features using the boundary distance method in the fuzzy decision model to obtain image evaluation values ​​corresponding to the images to be evaluated. The images to be evaluated are then filtered based on all image evaluation values ​​to obtain target images. The backbone network in the fuzzy decision model of this invention is constructed based on an inverted residual network structure, thereby enabling the extraction of more image features from the images to be evaluated, enriching the evaluation image features. Furthermore, the image evaluation of the evaluation image features using the boundary distance method enables the determination of image evaluation values, improves the accuracy of the fuzzy decision model, and thus achieves accurate judgment of the fuzziness level of the images to be evaluated, reduces the influence of non-target images, and improves the accuracy of target image selection. Furthermore, in the training of the fuzzy judgment model, the initial image is degraded by a preset degradation method in the degradation model, thereby simulating various fuzzy images in reality, and thus ensuring that the fuzzy judgment model can accurately evaluate the degree of fuzziness of various fuzzy images. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating an image blur assessment method provided in Embodiment 1 of the present invention;

[0022] Figure 2 This is a flowchart illustrating an image blur assessment method provided in Embodiment 2 of the present invention;

[0023] Figure 3 This is a schematic diagram of the structure of an image blur assessment device provided in Embodiment 3 of the present invention;

[0024] Figure 4 This is a schematic diagram of the structure of a computer device provided in Embodiment 4 of the present invention. Detailed Implementation

[0025] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0026] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0027] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0028] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0029] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0030] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0031] To illustrate the technical solution of the present invention, specific embodiments are described below.

[0032] See Figure 1 This is a flowchart illustrating an image blur assessment method provided in Embodiment 1 of the present invention, as shown below. Figure 1 As shown, the image blur assessment method may include the following steps:

[0033] Step S10: Obtain an image dataset, which includes at least one image to be evaluated.

[0034] Understandably, the image to be evaluated is the image whose blur level needs to be assessed. The image to be evaluated can be collected from different databases, or it can be a pre-prepared image dataset sent from the client to the database. The image dataset includes at least one image to be evaluated. For example, in archiving tasks, some non-face data and low-quality face data may appear during the capture process. To improve the performance of the archived data, it is necessary to evaluate the blur level of the images, and then filter out the images that meet the requirements.

[0035] In this invention, the image to be evaluated can be a face image, or it can be set to an image other than a face image according to the image recognition requirements.

[0036] Step S20: Obtain a fuzzy decision model, and extract features from all the images to be evaluated through the backbone network in the fuzzy decision model to obtain the evaluation image features corresponding to the images to be evaluated; the backbone network is constructed based on the inverted residual structure.

[0037] The fuzzy decision model was trained on a large amount of data. The evaluation image features are used to characterize the image to be evaluated. The backbone network is constructed based on an inverse residual structure.

[0038] Specifically, the trained fuzzy decision model is retrieved, and all images to be evaluated are input into it. The backbone network of the fuzzy decision model, built based on an inverse residual structure, extracts features from the images. This involves using the extension layers in the backbone network to upscale all images according to a set expansion coefficient, mapping the images from low-dimensional space to high-dimensional space to obtain high-dimensional image features. Then, the convolutional layers in the backbone network perform depthwise convolution on these high-dimensional image features, extracting features from the high-dimensional images to obtain convolutional image features. Finally, the projection layers in the backbone network perform dimensionality reduction on the convolutional image features, restoring the extracted convolutional image features to the dimensions of the images to be evaluated, thus obtaining the evaluation image features.

[0039] Step S30: Image evaluation is performed on all the evaluation image features using the boundary distance method in the fuzzy determination model to obtain the image evaluation value corresponding to the image to be evaluated.

[0040] Step S40: Filter the images to be evaluated based on all the image evaluation values ​​to obtain the target image.

[0041] Understandably, image evaluation values ​​are scores given to the image being evaluated. The target image is the image to be evaluated that meets the requirements, such as face data that meets the archiving requirements in an archiving task.

[0042] Specifically, after obtaining the evaluation image features, the boundary distance method in the fuzzy judgment model is used to score all evaluation image features. That is, each evaluation image feature is evaluated sequentially according to the fuzziness levels defined during training, thus obtaining the image fuzziness level of each feature. For example, image fuzziness levels are set as A, B, C, and D. The evaluation image features of each image are classified according to the principle that level A images have the highest fuzziness and level D images have the lowest fuzziness (highest clarity), obtaining the fuzziness level corresponding to each image. Then, the image to be evaluated is scored based on the boundary distance method and the image fuzziness level of the image to be evaluated. This determines the boundary difference between the image fuzziness level and the highest clarity level of the image to be evaluated. The image is scored using the boundary distance method, with lower evaluation values ​​indicating a more blurred image and higher evaluation values ​​indicating a clearer image, thus obtaining the image evaluation value corresponding to each image to be evaluated.

[0043] Furthermore, a preset threshold is obtained, and the image evaluation value of the image to be evaluated is compared with the preset threshold. When the image evaluation value is less than or equal to the preset threshold, the image to be evaluated corresponding to the image evaluation value that is less than or equal to the preset threshold is deleted, and the client is prompted that the image to be evaluated does not meet the requirements and is requested to re-upload a new image. When the image evaluation value is greater than the preset threshold, the image to be evaluated corresponding to the image evaluation value that is greater than the preset threshold is retained and determined as the target image.

[0044] In this embodiment, the backbone network of the fuzzy decision model is constructed based on an inverted residual network structure, thereby enabling the extraction of more image features from the image to be evaluated and enriching the features of the evaluation image. Image evaluation of the evaluation image features is performed using the boundary distance method, thus determining the image evaluation value, improving the accuracy of the fuzzy decision model, and ultimately achieving accurate judgment of the degree of fuzziness of the image to be evaluated. This reduces the influence of non-target images and improves the accuracy of target image selection.

[0045] In one embodiment, in step S20, feature extraction is performed on all the images to be evaluated using the backbone network in the fuzzy decision model to obtain the evaluation image features corresponding to the images to be evaluated; the backbone network is constructed based on an inverted residual structure and includes:

[0046] Step S201: The image to be evaluated is subjected to dimensionality upscaling through the extended layer in the backbone network to obtain high-dimensional image features.

[0047] Understandably, high-dimensional image features are obtained by increasing the dimensionality of the image to be evaluated through the extended layer.

[0048] Specifically, after obtaining the fuzzy decision model, all images to be evaluated are input into the fuzzy decision model. The backbone network of the fuzzy decision model extracts features from all images to be evaluated. This involves first performing dimensionality upscaling on all images to be evaluated through the extension layer in the backbone network, specifically by using a pre-set extension coefficient. A 1x1 convolutional matrix is ​​used to increase the dimensionality of the number of channels in all images to be evaluated using the extension coefficient, resulting in dimensionality-upgraded image features. Then, the ReLU6 activation function is used to activate these dimensionality-upgraded image features, thus obtaining high-dimensional image features corresponding to each image to be evaluated. For example, for input features h*w*k (height, width, channel), a 1x1 convolutional kernel is used to perform dimensionality upscaling on the input features. After dimensionality upscaling, the channel of the output feature matrix becomes tk, where t represents the extension coefficient. The ReLU6 activation function defaults to 0 when the input value is less than 0. It does not change the input value within the range [0, 6]. When the input value is greater than 6, the output value is set to 6.

[0049] Step S202: Perform deep convolution processing on the high-dimensional image features through the convolutional layers in the backbone network to obtain convolutional image features.

[0050] Understandably, convolutional image features are obtained by depthwise convolution of high-dimensional image features by convolutional layers.

[0051] Specifically, after obtaining the high-dimensional image features, deep convolution processing is performed on all high-dimensional image features through convolutional layers in the backbone network. This involves extracting image features from the high-dimensional image features using a 3x3 deep convolution kernel with a stride of 's' (the stride can be set according to actual conditions, such as 's=2'). The extracted image features are then activated using the ReLU6 activation function, resulting in convolutional image features corresponding to each image to be evaluated. For example, corresponding to the output h*w*tk of the extended layer, the convolutional layer uses DW convolution with a 3x3 kernel, a stride of 's', and the ReLU6 activation function. After DW convolution, the channel size remains unchanged; the output channels are the same as the input. Because the stride is 's', the output feature matrices h and w become 1 / s times their original size.

[0052] Step S203: The convolutional image features are dimensionality reduced by the projection layer in the backbone network to obtain the evaluation image features.

[0053] Understandably, the image features being evaluated are the features extracted from the image by the backbone network.

[0054] Specifically, after obtaining the convolutional image features, the projection layer in the backbone network performs dimensionality reduction on all convolutional image features. That is, the projection layer restores the dimensions of all convolutional image features to their original dimensions (i.e., the dimensions of the image to be evaluated). In other words, a 1*1 convolutional kernel is used to reduce the dimensionality of all convolutional image features, resulting in dimensionality-reduced image features. Then, a linear activation function is used to activate the dimensionality-reduced image features, thus obtaining the evaluation image features corresponding to each image to be evaluated. For example, corresponding to the output h / s*w / s*tk of the convolutional layer, the projection layer uses a 1*1 convolutional kernel and a linear activation function. After convolution with the 1*1 kernel, the channel size is changed, and the output channel is the same as the input of the expansion layer. Since the stride is 1 / s, the output feature matrix h and w become s times their original size, resulting in output features of h*w*k.

[0055] In this embodiment, by first increasing the dimensionality of the image to be evaluated and then performing deep convolution processing on the high-dimensional image features, more image information is extracted from the image to be evaluated, thereby enriching the convolutional image features. Furthermore, by using a linear function in the projection layer to reduce the dimensionality of the convolutional image features, information collapse is avoided, thus reducing information loss in the image to be evaluated during the dimensionality reduction process and reducing the computational load of the dimensionality reduction process.

[0056] In one embodiment, step S30, namely, performing image evaluation on all the evaluation image features using the boundary distance method in the fuzzy decision model to obtain the image evaluation value corresponding to the image to be evaluated, includes:

[0057] Step S301: Perform a blur assessment on the image to be evaluated to obtain the image blur level of the image to be evaluated.

[0058] Step S302: Image evaluation is performed on all the features of the evaluation image using the boundary distance method in the fuzzy determination model and the image fuzziness level of the image to be evaluated, to obtain the image evaluation value corresponding to the image to be evaluated.

[0059] Understandably, the image blur level is the degree of blurriness of the image being evaluated. The image evaluation value is a score given to the image being evaluated.

[0060] Specifically, after obtaining the features of the evaluation image, the image to be evaluated is fuzzy by using the hierarchical classification learned during training. That is, based on the degree of fuzziness of the image to be evaluated, it is classified into a corresponding image fuzziness level. Thus, by using the learned hierarchical classification to evaluate the fuzziness of all images to be evaluated, the image fuzziness level of each image is obtained. Further, image evaluation is performed on all evaluation image features using the boundary distance method in the fuzziness judgment model and the image fuzziness level of the image to be evaluated. That is, the boundary difference between the image fuzziness level of the image to be evaluated and the level of the clear image is calculated. Then, the image fuzziness level of the image to be evaluated is scored using the boundary distance method and the boundary difference, thus obtaining the image evaluation value corresponding to each image to be evaluated.

[0061] In this embodiment, a fuzzy judgment model is used to determine the fuzziness level of the image to be evaluated, thereby determining the fuzziness level of the image. Image evaluation is performed by assessing image features using the image fuzziness level and the boundary distance method, thus determining the image evaluation value and achieving accurate judgment of the image to be evaluated, improving the accuracy of the fuzzy judgment model.

[0062] See Figure 2 This is a flowchart illustrating an image blur assessment method provided in Embodiment 2 of the present invention. In this image blur assessment method, before obtaining the blur determination model, a preset training model is trained to obtain a trained blur determination model.

[0063] The training process for a pre-defined training model includes the following steps:

[0064] Step S501: Obtain a sample training dataset, which includes at least one sample image and a sample level label corresponding to the sample image, wherein the sample level label is a characterization of the blur level of the sample image;

[0065] Step S502: Obtain a preset training model, and extract features from the sample images through the backbone network of the preset training model to obtain sample image features corresponding to each sample image.

[0066] Understandably, the sample images can be obtained from different clients or pre-prepared degraded images collected from different databases. Each sample image corresponds to a sample level label, which represents the blur level of the sample image. A sample training dataset is then constructed based on all the obtained sample images and sample level labels. The preset training model is a pre-built training model. Sample image features are local images representing the sample images.

[0067] Specifically, before obtaining the fuzzy judgment model, a pre-prepared sample training dataset is acquired. A preset training model is retrieved from the database, and all sample images are input into the preset training model. The preset training model is trained using a large number of sample images, that is, feature extraction is performed on all sample images through the backbone network of the preset training model. This involves first performing dimensionality upscaling on all sample images through an expansion layer, increasing the channel dimension according to the expansion coefficient (which can be set according to actual conditions, such as 4 times). Then, a convolutional layer performs depthwise convolution on the dimensionality-upgraded sample images, that is, using lightweight depthwise convolution to filter the dimensionality-upgraded sample images. Finally, a projection layer performs dimensionality reduction on the convolutional sample images, thereby obtaining the sample image features corresponding to each sample image.

[0068] Step S503: The sample image features are classified into levels using the classification module of the preset training model to obtain the predicted fuzziness level;

[0069] Step S504: The sample image features are evaluated and ranked by the ranking module of the preset training model to obtain the predicted ranking value;

[0070] Understandably, the predicted blur level is obtained by predicting the degree of blur of the sample image features. The predicted ranking value is obtained by scoring the degree of blur of the sample image features.

[0071] Specifically, after obtaining the sample image features, the classification module of a pre-trained model classifies all sample image features into levels based on their blurriness. This means dividing all sample image features into a predetermined number of blur levels (e.g., 4 or 5) according to their blurriness. Training with a large amount of data enables the classification module to classify the sample image features according to their blurriness levels, thus obtaining the predicted blur level corresponding to each sample image feature. Further, the ranking module of the pre-trained model evaluates and ranks all sample image features. This involves scoring each sample image feature using its predicted blur level and the boundary distance method in the ranking module, according to pre-defined scoring requirements: higher blurriness results in a lower score, and clearer images result in a higher score. This yields the predicted ranking value corresponding to each sample image feature. For example, the evaluation value for the adjusted image corresponding to a sample image is 0.8, and the predicted ranking value for the most blurred level of the sample image, calculated using the boundary distance method, is 0.4.

[0072] Step S505: Determine the prediction loss value of the preset training model based on the sample level label corresponding to the same sample image, the prediction fuzziness level, and the prediction ranking value.

[0073] Step S506: When the predicted loss value reaches the convergence condition, the preset training model after convergence is determined as a fuzzy decision model.

[0074] Understandably, the prediction loss value is generated during the training of the pre-defined training model. The blur determination model is used to judge the degree of blur in an image. The convergence condition can be either the prediction loss value being less than a set threshold, or the prediction loss value being very small after 500 calculations and no longer decreasing.

[0075] Specifically, after obtaining the predicted fuzziness level and predicted ranking value, the predicted fuzziness levels are arranged according to the training of sample images in the sample training dataset. Then, the sample ranking label corresponding to the sample image is compared with the predicted fuzziness level of the same sequence. That is, based on the sample ranking label and predicted fuzziness level corresponding to the same sample image, the classification loss value of that sample image is determined. Next, the ranking loss value of the sample image is calculated based on the boundary distance method and the predicted fuzziness level. Finally, the overall loss value of the preset training model is calculated based on the classification loss value and the ranking loss value, that is, the overall loss value is calculated according to the weighted values ​​of the classification loss value and the ranking loss value, to obtain the predicted loss value of the preset training model.

[0076] Furthermore, it is determined whether the predicted loss value has reached the convergence condition. If the predicted loss value has not reached the preset convergence condition, the initial parameters of the preset training model are adjusted based on the predicted loss value, and all sample images are re-inputted into the preset training model with adjusted initial parameters. The preset training model with adjusted initial parameters is retrained using the sample images to obtain the predicted loss value corresponding to the preset training model with adjusted initial parameters. If the predicted loss value has not reached the preset convergence condition, the initial parameters of the preset recognition model are adjusted again based on the predicted loss value, so that the predicted loss value of the preset training model with adjusted initial parameters reaches the preset convergence condition. In this way, the output of the preset training model can continuously approach the accurate result, making the prediction accuracy of the model higher and higher, until the predicted loss value of the preset training model reaches the preset convergence condition. At this point, the converged preset training model is recorded as the fuzzy decision model.

[0077] In this embodiment, a pre-defined training model is trained using a large number of sample images. Classification and ranking loss values ​​are calculated using a loss function, thus realizing the overall loss value of the pre-defined training model. The parameters of the pre-defined training model are then adjusted multiple times based on the predicted loss value until the model converges, thereby determining the fuzzy decision model and ensuring its high accuracy.

[0078] In one embodiment, step S505, determining the prediction loss value of the preset training model based on the sample level label corresponding to the same sample image, the predicted fuzziness level, and the predicted ranking value, includes:

[0079] Step S5051: Determine the classification loss value corresponding to the preset training model based on the sample level label and the predicted fuzziness level corresponding to the same sample image;

[0080] Understandably, the classification loss value is generated during the process of classifying sample images into ranks. The cross-entropy loss is expressed as follows: Where N is the number of samples, p i This represents the probability that a sample image is classified into the correct fuzz level.

[0081] Specifically, the predicted fuzziness level corresponding to each sample image is arranged according to the order of the sample images in the sample training dataset. Then, the sample level label corresponding to the sample image is compared with the predicted fuzziness level of the sample images in the same sequence. That is, according to the sample images, the sample level label corresponding to the first sample image is compared with the predicted fuzziness level corresponding to the first sample image. The loss value between the sample level label and the predicted fuzziness level is calculated using the classification loss function, namely cross-entropy loss. Then, the sample level label corresponding to the second sample image is compared with the predicted fuzziness level corresponding to the second sample image. This process continues until the predicted fuzziness level and the sample level label are compared, and then the classification loss value of the preset training model can be determined.

[0082] Step S5052: Determine the ranking loss value corresponding to the preset training model based on the predicted blur level and the predicted ranking value corresponding to the same sample image;

[0083] Understandably, the ranking loss is generated during the image scoring process of the sample images. The expression for the boundary distance method is: Where N represents the number of samples, R i Represents the ranking loss for each sample. Among them, pred i Let m be the ranking score for sample i. ij Represents the boundary distance between sample i and sample j, 1 ij The indicator function is 1 when the ambiguity of sample i is higher than that of sample j, and 0 otherwise.

[0084] Specifically, the ranking loss of the preset training model is calculated by using the predicted blur level and predicted ranking value corresponding to the same sample image. That is, by adjusting the blur level of the corresponding image (e.g., dividing the blur level into four levels, represented by A, B, C, and D, where the blur level decreases sequentially, with A representing the most blurred image and D representing the clearest image), the evaluation value of each sample image is determined using the boundary distance method based on the blur level and predicted blur level of the corresponding adjusted image for the same sample image. Then, based on the evaluation value and predicted ranking value corresponding to the same sample image, i.e., according to the sample images, the evaluation value corresponding to the first sample image is compared with the predicted ranking value corresponding to the first sample image, and the loss value between the evaluation value and the predicted ranking value is calculated using a loss function. This process is repeated for the second sample image, and so on, until both the predicted ranking value and the evaluation value are compared, thus determining the ranking loss value of the preset training model.

[0085] Step S5053: Obtain the preset relational expression and preset weighting factor, and determine the prediction loss value of the preset training model based on the preset relational expression, the preset weighting factor, the classification loss value and the ranking loss value.

[0086] Understandably, the preset relation is a formula for calculating the overall loss value of a preset training model. The preset weighting factor is the weight of the ranking loss value in the overall loss function. The expression for the overall loss function is L = L cls +λL rank During training, λ is set to 0.1, where L cls L is the classification loss value. rank This represents the sorting loss value.

[0087] Specifically, after obtaining the ranking loss value, the overall loss value of the preset training model is calculated based on the classification loss value and the ranking loss value. That is, the classification loss value and the ranking loss value are substituted into the corresponding positions in the expression of the overall loss function, and the overall loss value of the preset training model is calculated based on the expression of the overall loss function, thereby obtaining the predicted loss value of the preset training model.

[0088] In this embodiment, the loss value for classifying sample images is calculated using the cross-entropy loss function, thus determining the classification loss value. This helps avoid gradient vanishing during gradient descent calculation, which could lead to a decrease in the learning rate. The ranking loss is calculated using the margin, further determining the ranking loss value. This ranking loss value is then incorporated into the overall loss calculation, improving the accuracy of the fuzzy judgment model.

[0089] In one embodiment, the steps of constructing the sample training dataset before obtaining the sample training dataset include:

[0090] Step S601: Obtain at least one initial image;

[0091] Step S602: Adjust the initial image to obtain an adjusted image;

[0092] Step S603: Obtain a degradation model, and perform degradation processing on all the adjusted images using the preset degradation method in the degradation model to obtain at least one degradation image corresponding to each of the sample images;

[0093] Understandably, the initial image can be collected from different websites, different databases, or even a clear image from different clients. The adjusted image is obtained by changing the border shape and size of the initial image. The degradation model is a model built based on all the degradation modules, constructing corresponding degradation modules for each preset degradation method. The degraded image is a blurred image. The preset degradation methods are pre-set image degradation methods, such as downsampling, Gaussian white noise, JPEG noise, and Gaussian blur.

[0094] Specifically, before acquiring the sample training dataset, at least one clear initial image is obtained. All initial images are then adjusted, specifically by changing the border shape to a rectangle (or a circle, depending on the specific requirements) and resizing them to 128*128 pixels (adjusted according to business needs). This determines the adjusted image corresponding to each initial image. Further, a degradation model is retrieved from the database, and all adjusted images are input into it. Each degradation module uses a preset degradation method to blur all adjusted images. This means each image can be degraded through each degradation module, or all adjusted images can be divided into multiple batches, with each batch blurred through one or more degradation modules, resulting in at least one degraded image corresponding to each adjusted image. In this embodiment, downsampling, Gaussian white noise, JPEG noise, Gaussian blur, and motion blur are used for degradation processing.

[0095] In one embodiment, step S603 includes:

[0096] The first embodiment employs a downsampling method to degrade all the adjusted images, a process that includes the following steps:

[0097] Obtain a preset downsampling ratio, and determine the downsampling size based on the preset downsampling ratio;

[0098] All the adjusted images are downsampled using bilinear interpolation and the downsampling size to obtain downsampled images;

[0099] The downsampled images are magnified by the bilinear interpolation to obtain a first degraded image corresponding to each of the adjusted images.

[0100] Specifically, the preset sampling ratio of the image is determined according to the actual situation (e.g., 0.4 or 0.6, etc.). This sampling ratio can be adjusted according to different requirements; in this embodiment, the preset sampling ratio ranges from 0.1 to 1.0. Then, the downsampling size is determined based on the preset sampling ratio, such as S = 128*A. The 128*128 image is downsampled using bilinear interpolation. Specifically, four points Q11, Q12, Q21, and Q22 are selected. Interpolation in the X direction yields R1 and R2, and the coordinates of R1 and R2 are determined based on these four points. Interpolation in the Y direction yields P(x, y), and the coordinates of P are determined based on the coordinates of R1 and R2. This yields the downsampled image corresponding to each adjusted image. Finally, all downsampled images are enlarged using bilinear interpolation, restoring the image size to 128*128, thus obtaining the first degraded image corresponding to each adjusted image.

[0101] In this embodiment, the method of downsampling and then upsampling is used to lose information in the image, increase image noise, degrade the image, and thus reduce the image clarity.

[0102] The second embodiment uses Gaussian white noise or JPEG noise to degrade all the adjusted images, as follows:

[0103] Specifically, Gaussian white noise has an amplitude distribution that follows a Gaussian distribution and a power spectral density that follows a uniform distribution. Based on a preset increase in noise concentration (i.e., standard deviation), a noise matrix of the same size as the original image is first generated. Then, the noise matrix and the adjusted image are directly weighted and summed. Finally, the assigned values ​​are cropped within a certain pixel value range to obtain the second degraded image corresponding to each adjusted image. For example, adding Gaussian noise with a noise concentration of 30 means that the amplitude distribution of the added noise conforms to a Gaussian distribution with a standard deviation of 30. Given that the mean of the Gaussian distribution is 0, the assigned noise points have approximately a 68% probability of falling within the range [-30, 30]. Therefore, increasing the standard deviation of the Gaussian distribution also increases the noise concentration of each pixel with a certain probability.

[0104] Furthermore, JPEG generates noise during compression, and the degree of compression is affected by the compression factor [0, 100]. The smaller the compression factor, the greater the noise generated, and the larger the compression factor, the less noise generated. In this embodiment, uniform sampling is used within the JPEG compression factor range [20, 100]. That is, all adjusted images are uniformly sampled according to the JPEG compression factor range to obtain the third degraded image corresponding to each adjusted image.

[0105] In this embodiment, the noise concentration of pixels in the image is increased by adding Gaussian white noise or JPEG noise, which degrades the image and reduces its clarity.

[0106] The third embodiment uses Gaussian blur or motion blur to degrade all the adjusted images, as follows:

[0107] Specifically, Gaussian blurring includes isotropic Gaussian blurring that smooths all regions equally, and anisotropic Gaussian blurring that produces different smoothing effects for different regions. In this embodiment, the Gaussian blur kernel size is uniformly selected between {7x7, 9x9, ..., 21x21}, wherein the kernel width of the isotropic Gaussian kernel is uniformly sampled between [0.1, 2.4] (x2) and [0.1, 2.8] (x4). For the anisotropic Gaussian kernel, the rotation angle is uniformly sampled between [0, π], and the length of each cycle is uniformly sampled between [0.5, 6] (x2) and [0.5, 8] (x4). By inputting all or part of the adjusted images into this degradation module, and performing degradation processing on the adjusted images using the above Gaussian sampling range, a fourth degradation image corresponding to each adjusted image can be obtained.

[0108] Furthermore, motion blur refers to the blur caused by the relative motion between the camera and the object. In this embodiment, the kernel size of the motion blur is uniformly sampled within the range of [5, 25], the degree is uniformly sampled within the range of [5, 25], and the angle is sampled within the range of [0, π]. Thus, by performing motion blur degradation on all adjusted images using the above-mentioned motion blur range, a fifth degraded image corresponding to each adjusted image can be obtained.

[0109] In this embodiment, Gaussian blur or motion blur is used to blur the information in the image, thereby degrading the image and reducing its sharpness. It should be understood that the five degradation methods in the above three embodiments are illustrative examples of this invention. This invention includes, but is not limited to, the above five degradation methods for degrading the adjusted image; it may involve two, three, or four methods simultaneously, or other degradation methods for image degradation.

[0110] Step S604: Compare all the degraded images and the adjusted images to obtain the sample blur level corresponding to each degraded image;

[0111] Step S605: Construct a sample training dataset based on the adjusted image, the degraded image, and the sample blur level corresponding to the degraded image.

[0112] Understandably, the sample blur level is the blur level of each degraded image, which corresponds to the level classification in the pre-trained model.

[0113] Specifically, after obtaining at least one degraded image corresponding to each adjusted image, all degraded and adjusted images corresponding to each adjusted image are compared. That is, the similarity between the degraded and adjusted images is calculated using a preset comparison method, specifically by calculating the similarity between each degraded and adjusted image using methods such as Euclidean distance or cosine similarity. The blur level of the degraded image is then divided based on the similarity between each degraded and adjusted image; each blur level corresponds to a similarity range, thus obtaining the sample blur level corresponding to each degraded image. For example, the degradation method in the degradation model uses different sequences of degradation methods to degrade clear images. Different combinations of degradation methods generate images with different degrees of blur. In this embodiment, the blur levels are divided into four grades, represented by A, B, C, and D, with the blur level decreasing sequentially. A represents the image with the most severe blur, and D represents the clear image.

[0114] In this embodiment, the initial image is degraded using multiple preset degradation methods in the degradation model, thereby simulating various blurred images in reality. This ensures that the fuzzy judgment model can accurately assess the degree of blur in various blurred images, thus improving the accuracy of the fuzzy judgment model in judging images of various degrees of blur.

[0115] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0116] See Figure 3 An image blur assessment device is provided, which corresponds one-to-one with the image blur assessment method described in the above embodiments. The image blur assessment device includes a data acquisition module 11, a feature extraction module 12, an image assessment module 13, and an image filtering module 14. Detailed descriptions of each functional module are as follows:

[0117] Data acquisition module 11 is used to acquire an image dataset, the image dataset including at least one image to be evaluated;

[0118] Feature extraction module 12 is used to obtain a fuzzy decision model, and to extract features from all the images to be evaluated through the backbone network in the fuzzy decision model to obtain the evaluation image features corresponding to the images to be evaluated; the backbone network is constructed based on the inverted residual structure;

[0119] Image evaluation module 13 is used to evaluate all the features of the evaluation image by means of the boundary distance method in the fuzzy determination model, and obtain the image evaluation value corresponding to the image to be evaluated;

[0120] The image filtering module 14 is used to filter the image to be evaluated based on all the image evaluation values ​​to obtain the target image.

[0121] Optionally, the feature extraction module 12 mentioned above includes:

[0122] The feature upscaling unit is used to upscale the image to be evaluated through the extended layer in the backbone network to obtain high-dimensional image features.

[0123] A deep convolutional unit is used to perform deep convolution processing on the high-dimensional image features through the convolutional layers in the backbone network to obtain convolutional image features.

[0124] The feature reduction unit is used to perform dimensionality reduction processing on the convolutional image features through the projection layer in the backbone network to obtain the evaluation image features.

[0125] Optionally, the image evaluation module 13 mentioned above includes:

[0126] A fuzziness judgment unit is used to perform fuzziness judgment on the image to be evaluated and obtain the image fuzziness level of the image to be evaluated.

[0127] The image evaluation unit is used to evaluate all the features of the evaluation image by using the boundary distance method in the fuzzy determination model and the image fuzziness level of the image to be evaluated, so as to obtain the image evaluation value corresponding to the image to be evaluated.

[0128] Optionally, the above-mentioned image blur assessment device further includes:

[0129] The sample data acquisition module is used to acquire a sample training dataset, which includes at least one sample image and a sample level label corresponding to the sample image, wherein the sample level label is a characterization of the blur level of the sample image.

[0130] The image feature module is used to obtain a preset training model, and to extract features from the sample images through the backbone network of the preset training model to obtain sample image features corresponding to each sample image.

[0131] The classification module is used to classify the features of the sample image by means of the classification module of the preset training model to obtain the predicted fuzzy level;

[0132] The evaluation and ranking module is used to evaluate and rank the features of the sample image through the ranking module of the preset training model to obtain the predicted ranking value.

[0133] The loss prediction module is used to determine the prediction loss value of the preset training model based on the sample level label corresponding to the same sample image, the prediction fuzziness level, and the prediction ranking value.

[0134] The model convergence module is used to determine the converged preset training model as a fuzzy decision model when the predicted loss value reaches the convergence condition.

[0135] Optionally, the above-mentioned image blur assessment device further includes:

[0136] The initial image acquisition module is used to acquire at least one initial image.

[0137] An image adjustment module is used to adjust the initial image to obtain an adjusted image;

[0138] An image degradation module is used to obtain a degradation model and perform degradation processing on all the adjusted images using a preset degradation method in the degradation model to obtain at least one degraded image corresponding to each of the sample images;

[0139] The blur level determination module is used to compare all the degraded images and the adjusted images to obtain the sample blur level corresponding to each of the degraded images;

[0140] The dataset construction module is used to construct a sample training dataset based on the adjusted image, the degraded image, and the sample blur level corresponding to the degraded image.

[0141] The aforementioned data image degradation module includes:

[0142] A sampling range unit is used to obtain a preset downsampling ratio and determine the downsampling size based on the preset downsampling ratio;

[0143] The downsampling unit is used to downsample all the adjusted images using bilinear interpolation and the downsampling size to obtain downsampled images;

[0144] The image degradation unit is used to enlarge all the downsampled images through the bilinear interpolation to obtain a first degradation image corresponding to each of the adjusted images.

[0145] Optionally, the above loss prediction module includes:

[0146] The classification loss unit is used to determine the classification loss value corresponding to the preset training model based on the sample level label and the predicted fuzziness level corresponding to the same sample image.

[0147] The ranking loss unit is used to determine the ranking loss value corresponding to the preset training model based on the predicted blur level and the predicted ranking value corresponding to the same sample image.

[0148] The overall loss unit is used to obtain a preset relation and a preset weighting factor, and to determine the prediction loss value of the preset training model based on the preset relation, the preset weighting factor, the classification loss value and the ranking loss value.

[0149] It should be noted that the information interaction and execution process between the above modules and units are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0150] Figure 4 This is a schematic diagram of the structure of a computer device provided in Embodiment 4 of the present invention. Figure 4 As shown, the computer device of this embodiment includes: at least one processor ( Figure 4 Only one is shown in the diagram), a memory, and a computer program stored in the memory and capable of running on at least one processor, which, when executed by the processor, implements the steps in any of the above-described embodiments of the image blur assessment methods.

[0151] This computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that... Figure 4 The examples of computer devices are merely examples and do not constitute a limitation on computer devices. Computer devices may include more or fewer components than shown in the illustration, or combinations of certain components, or different components, such as network interfaces, displays, and input devices.

[0152] The processor referred to can be a CPU, but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0153] Memory includes readable storage media, internal memory, etc., wherein internal memory can be the RAM of a computer device, providing an environment for the operation of the operating system and computer-readable instructions stored in the readable storage media. The readable storage media can be the hard drive of a computer device, or in other embodiments, it can be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, memory can include both internal storage units and external storage devices of a computer device. Memory is used to store the operating system, applications, bootloader, data, and other programs, such as program code for computer programs. Memory can also be used to temporarily store data that has been output or will be output.

[0154] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the functions described above can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this invention. The specific working process of the units and modules in the above device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention can implement all or part of the processes in the methods of the above embodiments by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium can include at least: any entity or device capable of carrying computer program code, a recording medium, a computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0155] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be accomplished by a computer program product. When the computer program product is run on a computer device, the computer device executes the steps in the above method embodiments.

[0156] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0157] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0158] In the embodiments provided by this invention, it should be understood that the disclosed apparatus / computer devices and methods can be implemented in other ways. For example, the apparatus / computer device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0159] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0160] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for evaluating image blurriness, characterized in that, The image blur assessment method includes: Obtain an image dataset, which includes at least one image to be evaluated; A fuzzy decision model is obtained, and features are extracted from all the images to be evaluated through the backbone network in the fuzzy decision model to obtain the evaluation image features corresponding to the images to be evaluated; the backbone network is constructed based on the inverted residual structure. Image evaluation is performed on all the features of the evaluation image using the boundary distance method in the fuzzy decision model to obtain the image evaluation value corresponding to the image to be evaluated. The images to be evaluated are filtered based on all the image evaluation values ​​to obtain the target images; Before obtaining the fuzzy determination model, the following steps are included: Obtain a sample training dataset, which includes at least one sample image and a sample level label corresponding to the sample image, wherein the sample level label is a characterization of the blur level of the sample image; A preset training model is obtained, and the backbone network of the preset training model is used to extract features from the sample images to obtain sample image features corresponding to each sample image. The sample image features are classified into levels using the classification module of the preset training model to obtain the predicted fuzziness level; The sample image features are evaluated and ranked by the ranking module of the preset training model to obtain the predicted ranking value; The prediction loss value of the preset training model is determined based on the sample level label corresponding to the same sample image, the predicted fuzziness level, and the predicted ranking value. When the predicted loss value reaches the convergence condition, the preset training model after convergence is determined as a fuzzy decision model.

2. The image blur assessment method according to claim 1, characterized in that, The step of extracting features from all the images to be evaluated through the backbone network in the fuzzy decision model to obtain the evaluation image features corresponding to the images to be evaluated includes: The image to be evaluated is up-dimensionally processed by the extended layer in the backbone network to obtain high-dimensional image features; The high-dimensional image features are processed by deep convolution through the convolutional layers in the backbone network to obtain convolutional image features; The convolutional image features are reduced in dimensionality by the projection layer in the backbone network to obtain the evaluation image features.

3. The image blur assessment method according to claim 1, characterized in that, The step of evaluating all the image features using the boundary distance method in the fuzzy decision model to obtain the image evaluation value corresponding to the image to be evaluated includes: The image to be evaluated is subjected to a blur determination to obtain the image blur level of the image to be evaluated; Image evaluation is performed on all the features of the evaluation image using the boundary distance method in the fuzzy determination model and the image fuzziness level of the image to be evaluated, so as to obtain the image evaluation value corresponding to the image to be evaluated.

4. The image blur assessment method according to claim 1, characterized in that, Before obtaining the sample training dataset, the following steps are included: Obtain at least one initial image; The initial image is adjusted to obtain an adjusted image; Obtain a degradation model, and perform degradation processing on all the adjusted images using a preset degradation method in the degradation model to obtain at least one degradation image corresponding to each of the sample images; All the degraded images and the adjusted images are compared to obtain the sample blur level corresponding to each degraded image; A sample training dataset is constructed based on the adjusted image, the degraded image, and the sample blur level corresponding to the degraded image.

5. The image blur assessment method according to claim 4, characterized in that, The step of performing degradation processing on all the adjusted images using a preset degradation method in the degradation model to obtain at least one degraded image corresponding to each of the adjusted images includes: Obtain a preset downsampling ratio, and determine the downsampling size based on the preset downsampling ratio; All the adjusted images are downsampled using bilinear interpolation and the downsampling size to obtain downsampled images; The downsampled images are magnified by the bilinear interpolation to obtain a first degraded image corresponding to each of the adjusted images.

6. The image blur assessment method according to claim 1, characterized in that, The step of determining the prediction loss value of the preset training model based on the sample level label corresponding to the same sample image, the predicted fuzziness level, and the predicted ranking value includes: Based on the sample level label and the predicted fuzziness level corresponding to the same sample image, determine the classification loss value corresponding to the preset training model; Based on the predicted blur level and the predicted ranking value corresponding to the same sample image, determine the ranking loss value corresponding to the preset training model; Obtain a preset relation and a preset weighting factor, and determine the prediction loss value of the preset training model based on the preset relation, the preset weighting factor, the classification loss value, and the ranking loss value.

7. An image blur assessment device, characterized in that, include: The data acquisition module is used to acquire an image dataset, which includes at least one image to be evaluated; The feature extraction module is used to obtain a fuzzy decision model, and to extract features from all the images to be evaluated through the backbone network in the fuzzy decision model to obtain the evaluation image features corresponding to the images to be evaluated; the backbone network is constructed based on the inverted residual structure; The image evaluation module is used to evaluate all the features of the evaluation image using the boundary distance method in the fuzzy decision model, and obtain the image evaluation value corresponding to the image to be evaluated. An image filtering module is used to filter the images to be evaluated based on all the image evaluation values ​​to obtain target images; The image blur assessment device further includes: The sample data acquisition module is used to acquire a sample training dataset, which includes at least one sample image and a sample level label corresponding to the sample image, wherein the sample level label is a characterization of the blur level of the sample image. The image feature module is used to obtain a preset training model, and to extract features from the sample images through the backbone network of the preset training model to obtain sample image features corresponding to each sample image. The classification module is used to classify the features of the sample image by means of the classification module of the preset training model to obtain the predicted fuzzy level; The evaluation and ranking module is used to evaluate and rank the features of the sample image through the ranking module of the preset training model to obtain the predicted ranking value. The loss prediction module is used to determine the prediction loss value of the preset training model based on the sample level label corresponding to the same sample image, the prediction fuzziness level, and the prediction ranking value. The model convergence module is used to determine the converged preset training model as a fuzzy decision model when the predicted loss value reaches the convergence condition.

8. A computer device, characterized in that, The computer device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the image blur assessment method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the image blur assessment method as described in any one of claims 1 to 6.

Images