Image quality intelligent analysis method and network-attached storage device

Abstract

Disclosed in the present invention are an image quality intelligent analysis method and a network-attached storage device. The method comprises: executing a training operation on a first image analysis model on the basis of a predetermined image data set to obtain a trained second image analysis model; calculating, on the basis of a preset loss calculation function, a model loss parameter corresponding to the second image analysis model, and determining whether the model loss parameter satisfies a preset model training condition; if yes, determining the second image analysis model as a target image analysis model, and inputting an image to be analyzed into the target image analysis model to obtain a model output result; and generating an image analysis result for said image on the basis of the model output result. Therefore, by implementing the present invention, an image can be analyzed in combination with a trained image analysis model, thereby improving the accuracy and reliability of an obtained image analysis result and improving the intelligence and objectivity of the obtained image analysis result.

Description

Intelligent analysis method for image quality and network attached storage device Technical Field

[0001] The present invention relates to the technical field of intelligent image processing, and in particular to an intelligent image quality analysis method and a network attached storage device. Background Art

[0002] With the rapid development of science and technology, artificial intelligence technology has also been applied to various fields of people's daily life and work. With the advancement of image processing technology, the combination of image processing technology and artificial intelligence for image processing has been increasingly widely used.

[0003] Currently, image quality assessment and analysis methods typically evaluate image quality based on peak signal-to-noise ratio (PSNR) or structural similarity. However, in practical applications, these methods differ significantly from the intuitive perception of image quality by humans. Furthermore, these methods employ a relatively single-minded approach to image quality analysis, resulting in low accuracy and difficulty in applying them to complex scenarios to meet increasingly complex requirements. Therefore, it is particularly important to provide a new image quality analysis method to improve the accuracy and intelligence of intelligent image analysis.

[0004] Summary of the Invention

[0005] The present invention provides an intelligent image quality analysis method and a network attached storage device, which can analyze the image to be analyzed in combination with a trained image analysis model, thereby improving the accuracy and reliability of the image analysis results obtained, as well as improving the intelligence and efficiency of the image analysis results obtained.

[0006] In order to solve the above technical problems, the first aspect of the present invention discloses an intelligent analysis method for image quality, the method comprising:

[0007] Performing a training operation on the first image analysis model based on a predetermined image data set to obtain a trained second image analysis model;

[0008] Calculating a model loss parameter corresponding to the second image analysis model according to a preset loss calculation function, and determining whether the model loss parameter meets a preset model training condition;

[0009] When it is determined that the model loss parameter meets the preset model training condition, the second image analysis model is determined as the target image analysis model, and the image to be analyzed is input into the target image analysis model to obtain a model output result;

[0010] Based on the model output result, an image analysis result of the image to be analyzed is generated; wherein the image analysis result includes an image quality analysis result of the image to be analyzed.

[0011] A second aspect of the present invention discloses a network attached storage device, wherein the network attached storage device is configured with at least a display screen and a memory storing executable program code;

[0012] a processor coupled to the memory;

[0013] The processor calls the executable program code stored in the memory to execute the intelligent image quality analysis method disclosed in the first aspect of the present invention.

[0014] A third aspect of the present invention discloses a computer storage medium, wherein the computer storage medium stores computer instructions. When the computer instructions are called, they are used to execute the intelligent image quality analysis method disclosed in the first aspect of the present invention.

[0015] The fourth aspect of the present invention discloses an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the intelligent image quality analysis method disclosed in the first aspect of the present invention.

[0016] The fifth aspect of the present invention discloses a system for intelligent analysis of image quality, the system comprising an electronic device as disclosed in the fourth aspect of the present invention, the electronic device being communicatively connected to a network attached storage device as disclosed in the second aspect of the present invention, and the electronic device being installed with a computer application capable of executing any one of the methods disclosed in the first aspect of the present invention.

[0017] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0018] The implementation of the present invention can obtain image analysis results of the image to be analyzed based on the trained model, and can realize objective analysis and evaluation of the image to be analyzed from multiple dimensions, which is conducive to improving the objectivity of the image analysis results, as well as improving the accuracy and reliability of the image analysis results, and further helping to improve the intelligence of the image analysis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] FIG1 is a schematic flow chart of an intelligent image quality analysis method disclosed in an embodiment of the present invention;

[0021] FIG2 is a flow chart of another intelligent image quality analysis method disclosed in an embodiment of the present invention;

[0022] FIG3 is a schematic diagram of the structure of a network attached storage device disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The present invention discloses an intelligent image quality analysis method, apparatus, and network-attached storage device. These methods can be combined with a trained image analysis model to analyze images to be analyzed, thereby improving the accuracy and reliability of the image analysis results, as well as the intelligence, efficiency, and objectivity of the obtained image analysis results. These methods are described in detail below.

[0024] Example 1

[0025] Please refer to Figure 1, which is a schematic flow chart of an intelligent image quality analysis method disclosed in an embodiment of the present invention. The intelligent image quality analysis method described in Figure 1 can be applied to an intelligent image quality analysis device, which can be integrated into a cloud server or a local server, although this is not limited in the embodiment of the present invention. As shown in Figure 1, the intelligent image quality analysis method may include the following operations:

[0026] 101. Perform a training operation on a first image analysis model based on a predetermined image data set to obtain a trained second image analysis model.

[0027] In the embodiment of the present invention, the AVA dataset is optionally a database for aesthetic quality assessment, including 250,000 photos, each of which has a corresponding score.

[0028] In the embodiment of the present invention, further optionally, the second image analysis model can be used to automatically perform a scoring operation on the image content to obtain an image analysis result for each image.

[0029] 102. Calculate the model loss parameters corresponding to the second image analysis model according to the preset loss calculation function, and determine whether the model loss parameters meet the preset model training conditions.

[0030] In the embodiment of the present invention, optionally, the preset loss calculation function includes one or more of a loss loss function, a cross entropy loss function, and an Earth Mover's Distance (EMD) concept.

[0031] In the embodiment of the present invention, optionally, the above-mentioned determination of whether the model loss parameter meets the preset model training conditions may include:

[0032] Determining a model loss value of the second image analysis model according to the model loss parameter, and determining whether the model loss value is less than or equal to a preset model loss threshold;

[0033] When it is determined that the model loss value is less than or equal to a preset model loss threshold, determining whether the model loss parameter meets the preset model training condition;

[0034] When it is determined that the model loss value is greater than a preset model loss threshold, it is determined that the model loss parameter does not meet the preset model training conditions.

[0035] 103. When it is determined that the model loss parameter meets the preset model training conditions, the second image analysis model is determined as the target image analysis model, and the image to be analyzed is input into the target image analysis model to obtain the model output result.

[0036] In the embodiment of the present invention, further optionally, the above method further includes:

[0037] When it is determined that the model loss parameters do not meet the preset model training conditions, the model update training parameters of the second image analysis model are determined based on the model loss parameters and the preset model training conditions, and based on the model update training parameters, a training operation is performed on the second image analysis model, and the execution is re-triggered to calculate the model loss parameters corresponding to the second image analysis model according to the preset loss calculation function, and to determine whether the model loss parameters meet the preset model training conditions.

[0038] 104. Generate image analysis results of the image to be analyzed based on the model output results.

[0039] In the embodiment of the present invention, the image analysis result includes the image quality analysis result of the image to be analyzed

[0040] In the embodiment of the present invention, the image analysis result may optionally include a quality rating result of the image to be analyzed, wherein the quality analysis result includes an aesthetic rating result of the image to be analyzed.

[0041] It can be seen that the intelligent image quality analysis method described in Figure 1 can perform a training operation on the first image analysis model based on the graphic data set to obtain a second image analysis model, and calculate the model loss parameters of the second image analysis model in combination with the loss calculation function and determine whether the preset model training conditions are met. If so, the second image analysis model is determined as the target image analysis model, and the image to be analyzed is input into the target image analysis model to obtain the model output result and further obtain the image analysis result. The image analysis result of the image to be analyzed can be obtained based on the trained model, and objective analysis and evaluation of the image to be analyzed can be achieved from multiple dimensions, which is conducive to improving the objectivity of the image analysis results, as well as improving the accuracy and reliability of the image analysis results, and further conducive to improving the intelligence of the image analysis results.

[0042] Example 2

[0043] Please refer to Figure 2, which is a flowchart illustrating another intelligent image quality analysis method disclosed in an embodiment of the present invention. The intelligent image quality analysis method described in Figure 2 can be applied to an intelligent image quality analysis device, which can be integrated into a cloud server or a local server, although this is not limited in the present embodiment. As shown in Figure 2, the intelligent image quality analysis method may include the following operations:

[0044] 201. Collect several training images.

[0045] In the embodiment of the present invention, optionally, a method of collecting a plurality of training images may be real-time or according to a preset time period, which is not specifically limited in the embodiment of the present invention.

[0046] 202. For each training image, perform an image scoring operation on the training image to obtain an image scoring result for the training image.

[0047] In the embodiment of the present invention, optionally, performing an image scoring operation on each training image to obtain an image scoring result for the training image may include:

[0048] For each training image, the training image is sent to a plurality of target users, so that each target user performs a scoring operation on the training image, obtains the scoring result of the target user on the training image, and causes the target user to feedback the scoring result of the training image;

[0049] Determine the image score result of the training image based on the rating result fed back by each target user;

[0050] The image scoring result of the training image includes the scoring results of all target users on the training image.

[0051] 203. For each training image, perform a score evaluation calculation operation on the training image according to the image score result of the training image to obtain an image evaluation result of the training image.

[0052] In the embodiment of the present invention, optionally, the image scoring result of each training image includes the scoring results of all target users on the training image.

[0053] In the embodiment of the present invention, optionally, for each training image, performing a score evaluation calculation operation on the training image according to the image score result of the training image to obtain the image evaluation result of the training image may include:

[0054] For each training image, generating a score probability distribution parameter of the training image according to the scoring result of each target user included in the training image;

[0055] For each training image, performing a score evaluation calculation operation on the training image according to the score probability distribution parameter of the training image and all the scoring results of the training image to obtain an image evaluation result of the training image;

[0056] The score evaluation calculation operation includes a variance statistics operation.

[0057] 204. Generate an image dataset based on each training image and the image evaluation result of each training image.

[0058] In an embodiment of the present invention, the image dataset optionally includes each training image and the image evaluation result of each training image. Further optionally, the image dataset includes multiple training images, and the specific number of training images is not specifically limited in this embodiment of the present invention.

[0059] 205. Perform a training operation on the first image analysis model based on the predetermined image data set to obtain a trained second image analysis model.

[0060] 206. Calculate the model loss parameters corresponding to the second image analysis model according to the preset loss calculation function, and determine whether the model loss parameters meet the preset model training conditions.

[0061] 207. When it is determined that the model loss parameter meets the preset model training conditions, the second image analysis model is determined as the target image analysis model, and the image to be analyzed is input into the target image analysis model to obtain the model output result.

[0062] 208. Generate image analysis results of the image to be analyzed based on the model output results.

[0063] In the embodiment of the present invention, for the detailed description of steps 205 to 208 , please refer to the other descriptions of steps 101 to 104 in the first embodiment, and the embodiment of the present invention will not be repeated.

[0064] It can be seen that the intelligent analysis method of image quality described in Figure 2 can obtain an image scoring result based on an image scoring operation performed on each collected training image, and obtain an image evaluation result by performing a score evaluation calculation operation based on the image scoring result. An image dataset is generated based on the image evaluation results of all training images and each training image, and a preprocessing operation can be performed on each training image first to obtain a corresponding scoring result, which is conducive to improving the accuracy and reliability of the generated image dataset, as well as improving the intelligence of the generated image dataset, thereby facilitating improving the accuracy and efficiency of the subsequent training of the first image analysis model to obtain the second image analysis model, and further facilitating improving the accuracy and reliability of the image analysis results of the image to be analyzed obtained by the second image analysis model, as well as improving the intelligence and efficiency of the image analysis results, and further facilitating improving the objectivity of the obtained image analysis results.

[0065] In an optional embodiment, after collecting a plurality of training images, the method further includes:

[0066] For each training image, determining at least one target interest point of the training image, and determining an offset parameter corresponding to each target interest point; and determining target label information of the training image based on each target interest point and the offset parameter corresponding to each target interest point, wherein the target label information includes background area information of the training image and area of ​​interest information of the training image;

[0067] For each training image, the training image is updated according to the training image and the target label information of the training image, and an image scoring operation is triggered for each training image to obtain an image scoring result of the training image.

[0068] In this optional embodiment, optionally, the above-mentioned determination of the offset parameter corresponding to each target point of interest may include:

[0069] For each target interest point, position information of the target interest point in the training image is determined, and an offset parameter corresponding to the target interest point is determined based on the position information.

[0070] In this optional embodiment, further optionally, the offset parameters corresponding to different target points of interest may be different or the same, which is not specifically limited in this embodiment of the present invention.

[0071] In this optional embodiment, the region of interest information for each training image may optionally include one or more of object information, animal information, and person information in the training image; and the background region information for each training image may include background image information in the training image. For example, if the training image depicts a rabbit on grass, the region of interest information includes region information of the rabbit in the corresponding region of the training image, and the grass is the background image information of the training image.

[0072] In this optional embodiment, the offset parameter may optionally include one or more of an offset angle, an offset distance, and an offset displacement.

[0073] In this optional embodiment, optionally, determining the target label information of the training image according to each target interest point and the offset parameter corresponding to each target interest point may include:

[0074] For each target interest point in the training image, generating background distribution information and focus distribution information of the target interest point according to the target interest point and the offset parameter corresponding to the target interest point;

[0075] The target label information of the training image is determined according to the background distribution information and the key distribution information of each target interest point.

[0076] In this alternative embodiment, for example, because the model is pre-trained on ImageNet, the network tends to place most points of interest near salient objects and scatter the remaining points in the background. By randomly assigning an offset to each point, the model also focuses on the background area, avoiding local optima. During training, the weight of this random assignment is gradually increased, thereby improving the intelligence of model training and the accuracy of the resulting trained model.

[0077] It can be seen that the implementation of this optional embodiment can determine the target interest points in each training image and determine the offset parameters corresponding to each target interest point, determine the target label information based on the target interest points and the offset parameters, and update the training images based on the target label information. This enables the model to pay attention to the background area of ​​each training image during model training to avoid falling into a local optimal state in the training image, and can integrate various information in the training image to generate target label information and then update the training image, which is conducive to improving the accuracy and reliability of updating the training images, as well as improving the intelligence of updating the training images, thereby facilitating improving the accuracy and reliability of the image scoring results for each training image, and improving the accuracy and reliability of the generated image data set, thereby facilitating improving the accuracy and reliability of the image analysis results of the image to be analyzed, and improving the intelligence and efficiency of obtaining the image analysis results, and further facilitating improving the objectivity of the obtained image analysis results.

[0078] In another optional embodiment, after generating an image analysis result of the image to be analyzed based on the model output result, the method further includes:

[0079] Perform calculation operations on the image analysis results based on a preset target calculation formula to obtain a comprehensive analysis result of the image to be analyzed;

[0080] Determine whether the comprehensive analysis results match the image analysis results;

[0081] When it is determined that the comprehensive analysis result does not match the image analysis result, an update operation is performed on the image analysis result according to the comprehensive analysis result to update the image analysis result of the image to be analyzed.

[0082] Among them, the preset target calculation formula includes one or more of a saturation score calculation formula, a brightness score calculation formula, a color diversity score calculation formula, and a clarity score calculation formula.

[0083] In this optional embodiment, further optionally, when it is determined that the comprehensive analysis result matches the image analysis result, the process can be terminated.

[0084] In this optional embodiment, optionally, the above-mentioned determination of whether the comprehensive analysis result matches the image analysis result may include:

[0085] Determine the result difference between the comprehensive analysis result and the image analysis result, and judge whether the result difference is greater than or equal to a preset result difference threshold;

[0086] When it is determined that the result difference value is greater than or equal to a preset result difference threshold, it is determined that the comprehensive analysis result does not match the image analysis result;

[0087] When it is determined that the result difference value is less than a preset result difference threshold, it is determined that the comprehensive analysis result matches the image analysis result.

[0088] In this optional embodiment, optionally, the above saturation score calculation formula includes:

[0089] Score = (standard deviation of image saturation - minimum value) / (maximum value - minimum value);

[0090] Among them, Score is the saturation score of the image to be analyzed, the standard deviation of the image saturation is a preset value, the minimum value is the minimum saturation value of the image to be analyzed, and the maximum value is the maximum saturation value of the image to be analyzed.

[0091] In this optional embodiment, optionally, for example, saturation (S) = (CM) / C, where C represents the maximum value of the pure color component in the color (for example, the maximum value of red is 255), and M represents the value of the gray component in the color. Through this formula, a numerical value between 0 and 1 can be obtained to represent the saturation of the color. Furthermore, a saturation of 0 indicates that the color is completely gray, and a saturation of 1 indicates that the color is a pure color.

[0092] In this optional embodiment, the brightness score calculation formula may optionally include:

[0093] Brightness score = 1-|average brightness-base brightness| / base brightness;

[0094] Where average brightness is the average brightness of the image to be analyzed, and reference brightness is a pre-set reference brightness value. Furthermore, the result of this formula ranges from -∞ to 1, where -∞ indicates that the image to be analyzed does not match the reference brightness at all, and 1 indicates that the image to be analyzed matches the reference brightness perfectly. A higher brightness score indicates that the image is closer to the reference brightness, while a lower brightness score indicates that the image differs significantly from the reference brightness.

[0095] In this optional embodiment, the color diversity score calculation formula may include: Score color diversity = (image color diversity variance - minimum value) / (maximum value - minimum value);

[0096] The maximum value is the maximum value of the color diversity of the image to be analyzed, and the minimum value is the minimum value of the color diversity of the image to be analyzed.

[0097] In this optional embodiment, the clarity score calculation formula may include: Score clarity = 1-SSIM(x,y);

[0098] Among them, μ x is the mean value of x, μ y is the mean value of y, σ x 2 is the variance of x, σ y 2 is the variance of y, σ xy is the covariance of x and y, and C1 and C2 are constants.

[0099] Furthermore, the larger the SSIM value, the better the image quality.

[0100] In this optional embodiment, further optionally, the calculation of SSIM can be split into brightness calculation, contrast calculation, and structure calculation, wherein the brightness calculation can be implemented by the following formula:

[0101] The contrast calculation can be achieved through the following formula:

[0102] Among them, the structural calculation can be achieved through the following formula:

[0103] Wherein, C3 is a constant.

[0104] It can be seen that the implementation of this optional embodiment can perform calculations on the image analysis results based on the preset target calculation formula to obtain comprehensive analysis results, and determine whether the comprehensive analysis results match the image analysis results. If they do not match, the image analysis results are updated. It can comprehensively analyze the image to be analyzed in combination with multiple aspects such as saturation, brightness, color diversity and clarity, and can realize comprehensive update operations on the image analysis results of the image to be analyzed, which is conducive to improving the accuracy and reliability of the image analysis results, as well as improving the intelligence and efficiency of the image analysis results, thereby helping to improve the objectivity of the image analysis results.

[0105] In yet another optional embodiment, before performing a training operation on the first image analysis model based on the predetermined image data set to obtain a trained second image analysis model, the method further includes:

[0106] Obtaining model training information of the first image analysis model, wherein the model training information includes one or more of training times, training duration, and training convergence degree information of the first image analysis model;

[0107] Determining dynamic training adjustment parameters of the first image analysis model according to the model training information, wherein the dynamic training adjustment parameters include a dynamic learning rate of the first image analysis model;

[0108] The step of performing a training operation on the first image analysis model based on the predetermined image data set to obtain a trained second image analysis model includes:

[0109] Based on the predetermined image data set and the dynamic training adjustment parameters, a training operation is performed on the first image analysis model to obtain a trained second image analysis model.

[0110] In this optional embodiment, optionally, the training convergence degree information includes training loss information of the first image analysis model; further, the smaller the training loss of the first image analysis model, the more converged the first image analysis model is.

[0111] In this optional embodiment, optionally, determining the dynamic training adjustment parameters of the first image analysis model based on the model training information may include:

[0112] determining a training stage of the first image analysis model according to the model training information;

[0113] According to the training stage of the first image analysis model, determining a learning rate that matches the training stage of the first image analysis model from a preset adjustment parameter database;

[0114] determining a dynamic training adjustment parameter of the first image analysis model according to a learning rate matched to a training phase of the first image analysis model;

[0115] Among them, the learning rates corresponding to different training stages are different.

[0116] In this optional embodiment, for example, since the neural network is very unstable at the beginning of training, the initial learning rate should be set very low, so as to ensure that the network has good convergence, but the low learning rate will make the training process very slow, so here we will adopt a method of gradually increasing the learning rate from a lower learning rate to a higher learning rate to implement the "warm-up" stage of network training, called the warmup stage, but if we minimize the loss of network training, it is not appropriate to always use a higher learning rate, because it will cause the gradient of the weight to oscillate back and forth, and it is difficult to make the loss value of training reach the global lowest point. In this way, determining different learning rates through different training stages of the model can improve the effect of model training, and is conducive to improving the intelligence and accuracy of the training model.

[0117] It can be seen that the implementation of this optional embodiment can determine the dynamic training adjustment parameters based on the acquired model training information of the first image analysis model, and train the first image analysis model based on the image data set and the dynamic training adjustment parameters to obtain the second image analysis model. It can determine the corresponding training adjustment parameters in combination with the training information of the first image analysis model at different stages, which is beneficial to improving the effect of model training for the first image analysis model, and is beneficial to improving the intelligence and efficiency of model training, thereby also helping to improve the accuracy and reliability of training the first image analysis model, and further helping to improve the accuracy of subsequently obtaining the second image analysis model and outputting the image analysis results of each image to be analyzed.

[0118] In yet another optional embodiment, for each training image, determining at least one target interest point of the training image includes:

[0119] For each training image, a thermal statistical operation is performed on the training image based on a preset attention mechanism to obtain the thermal statistical parameters of the training image;

[0120] For each training image, a target region of the training image is determined according to the thermodynamic statistical parameters of the training image, and a target interest point of the training image is determined according to the target region of the training image.

[0121] In this optional embodiment, the attention mechanism is optionally derived from the study of human vision. In cognitive science, due to the bottleneck of information processing, humans selectively focus on a portion of all information while ignoring other visible information. In order to rationally utilize limited visual information processing resources, humans need to select a specific part of the visual area and then focus on it. For example, when people read, usually only a small number of words to be read will be paid attention to and processed. In summary, the attention mechanism has two main aspects: determining which part of the input needs to be paid attention to; allocating limited information processing resources to the important part.

[0122] In this optional embodiment, the thermodynamic statistical parameters of each training image optionally include a thermodynamic statistical value corresponding to a key area of ​​the training image and a thermodynamic statistical value corresponding to a background area. For example, the thermodynamic statistical value corresponding to the key area of ​​the training image is higher, and the thermodynamic statistical value corresponding to the background area of ​​the training image is lower.

[0123] In this optional embodiment, optionally, for each training image, determining the target area of ​​the training image according to the thermodynamic statistical parameters of the training image may include:

[0124] For each training image, a heat map corresponding to the training image is generated according to the heat statistical parameters of the training image, wherein the heat values ​​corresponding to the key areas in the heat map are higher;

[0125] For each training image, according to the heat map corresponding to the training image, a target area having a heat value greater than or equal to a preset heat threshold is determined from the heat map.

[0126] In this optional embodiment, optionally, determining the target interest point of the training image based on the target area of ​​the training image may include:

[0127] According to the target area of ​​the training image, the thermal value of each area point in the target area is determined, and the highest thermal value is determined from all the thermal values, and the area point corresponding to the highest thermal value is determined as the target interest point.

[0128] In this optional embodiment, optionally, for example, because the model is pre-trained on ImageNet, such a network tends to place most of the interest points close to salient objects and scatter the remaining interest points in the background, which are target interest points.

[0129] It can be seen that the implementation of this optional embodiment can perform thermal statistical operations on the training images based on the preset attention mechanism to obtain thermal statistical parameters, and determine the target area and further determine the target interest points based on the thermal statistical parameters. It can combine the attention mechanism and the thermal statistical parameters to comprehensively determine the target area, and can realize the determination of the target area by integrating various information, and can improve the accuracy and reliability of determining the target area, which is conducive to improving the accuracy and reliability of determining the target interest points of each training image, and further conducive to improving the accuracy and reliability of subsequent updates of training images and determination of image score results for each training image, and further conducive to improving the accuracy, reliability and objectivity of subsequent generated image data sets.

[0130] In yet another optional embodiment, calculating the model loss parameter corresponding to the second image analysis model according to a preset loss calculation function includes:

[0131] Inputting at least one preset image to be tested into the second image analysis model to obtain a test output result;

[0132] Based on the preset loss calculation function and all test output results, the probability distribution difference parameters corresponding to the second image analysis model are calculated, and based on the probability distribution difference parameters, the model loss parameters corresponding to the second image analysis model are determined.

[0133] In this optional embodiment, the number of images to be tested may be one or more, which is not specifically limited in this embodiment of the present invention. Optionally, the preset loss calculation function includes one or more of a loss loss function and a cross entropy loss function, which is not specifically limited in this embodiment of the present invention.

[0134] In this optional embodiment, optionally, the test output result includes a quality score result of each image to be tested.

[0135] In this optional embodiment, optionally, the calculation of the probability distribution difference parameter corresponding to the second image analysis model based on the preset loss calculation function and all test output results may include:

[0136] Determine the true label result of each image to be tested, and calculate the image loss parameter between the true label result of the image to be tested and the test output result based on a preset loss calculation function and the test output result of the image to be tested;

[0137] According to the image loss parameters of all images to be tested, probability distribution difference parameters corresponding to the second image analysis model are generated.

[0138] In this optional embodiment, optionally, the above-mentioned determination of the model loss parameter corresponding to the second image analysis model based on the probability distribution difference parameter may include: determining the probability distribution difference parameter as the model loss parameter corresponding to the second image analysis model.

[0139] In this optional embodiment, optionally, the smaller the probability distribution difference parameter is, the smaller the model loss parameter is; and the larger the probability distribution difference parameter is, the larger the model loss parameter is.

[0140] In this optional embodiment, further optionally, determining whether the model loss parameter meets the preset model training conditions may include:

[0141] Determine a probability distribution difference value according to the probability distribution difference parameter, and determine whether the probability distribution difference value is less than or equal to a preset probability distribution difference threshold;

[0142] When it is determined that the probability distribution difference value is less than or equal to a preset probability distribution difference threshold, determining that the model loss parameter meets the preset model training condition;

[0143] When it is determined that the probability distribution difference value is greater than a preset probability distribution difference threshold, it is determined that the model loss parameter does not meet the preset model training conditions.

[0144] It can be seen that the implementation of this optional embodiment can calculate the probability distribution difference parameters corresponding to the second image analysis model and then determine the model loss parameters based on the test output results of all images to be tested after passing through the second image analysis model and the loss calculation function. The probability distribution difference parameters can be calculated in combination with the test output results and the loss calculation function, which is conducive to improving the accuracy and reliability of the calculated probability distribution difference parameters, thereby helping to improve the accuracy and reliability of determining the model loss parameters corresponding to the second image analysis model, and thereby helping to improve the accuracy, reliability and objectivity of the subsequent target image analysis model and the image analysis results corresponding to each image to be analyzed obtained based on the target image analysis model.

[0145] Example 3

[0146] An embodiment of the present invention discloses an electronic device, including a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor performs the steps of the intelligent image quality analysis method described in Embodiment 1 or Embodiment 2 of the present invention.

[0147] In an embodiment of the present invention, optionally, the electronic device includes one or more of a smart phone, a smart tablet, and a smart TV, and a related application is installed in the electronic device, and the application can be used to access the NAS device, thereby realizing an intelligent analysis method for image quality.

[0148] Example 4

[0149] Please refer to Figure 3, which is a schematic diagram of the structure of a network attached storage device disclosed in an embodiment of the present invention. As shown in Figure 3, the network attached storage device is at least equipped with a display screen, and the network attached storage device has a memory 401 storing executable program code;

[0150] a processor 402 coupled to the memory 401;

[0151] The processor 402 calls the executable program code stored in the memory 401 to execute the steps of the intelligent image quality analysis method described in the first embodiment or the second embodiment of the present invention.

[0152] Example 5

[0153] An embodiment of the present invention discloses a computer storage medium storing computer instructions. When the computer instructions are called, they are used to execute the steps of the intelligent image quality analysis method described in the first embodiment or the second embodiment of the present invention.

[0154] Example 6

[0155] An embodiment of the present invention discloses a system for intelligent analysis of image quality, which includes an electronic device as disclosed in Example 3 of the present invention, and the electronic device is communicatively connected to a network attached storage device as disclosed in Example 4 of the present invention, and a computer application capable of executing Example 1 or Example 2 is installed in the electronic device.

[0156] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus the necessary general hardware platform, or of course, by means of hardware. Based on this understanding, the above technical solution, in essence, or the portion that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, including a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disk storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

Claims

1. An intelligent analysis method for image quality, characterized in that, The method includes: Performing a training operation on a first image analysis model based on a pre-determined image dataset to obtain a trained second image analysis model; Calculating a model loss parameter corresponding to the second image analysis model according to a preset loss calculation function, and determining whether the model loss parameter meets a preset model training condition; When it is determined that the model loss parameter meets the preset model training condition, determining the second image analysis model as a target image analysis model, and inputting an image to be analyzed into the target image analysis model to obtain a model output result; Generating an image analysis result of the image to be analyzed based on the model output result; wherein, the image analysis result includes an image quality analysis result of the image to be analyzed.

2. The intelligent analysis method for image quality according to claim 1, wherein Before performing the training operation on the first image analysis model based on the pre-determined image dataset to obtain the trained second image analysis model, the method further includes: Collecting a plurality of training images; For each of the training images, performing an image scoring operation on the training image to obtain an image score result of the training image; For each of the training images, performing a score evaluation calculation operation on the training image according to the image score result of the training image to obtain an image evaluation result of the training image; Generating an image dataset according to each of the training images and the image evaluation result of each of the training images.

3. The intelligent analysis method for image quality according to claim 2, characterized in that, After collecting the plurality of training images, the method further includes: For each of the training images, determining at least one target interest point of the training image, and determining an offset parameter corresponding to each target interest point, and determining target label information of the training image according to each target interest point and the offset parameter corresponding to each target interest point, wherein the target label information includes background region information of the training image and interest region information of the training image; For each of the training images, updating the training image according to the training image and the target label information of the training image, and triggering the operation of performing an image scoring operation on the training image to obtain an image score result of the training image for each of the training images.

4. The intelligent analysis method for image quality according to claim 3, characterized in that, After generating the image analysis result of the image to be analyzed based on the model output result, the method further includes: Performing a calculation operation on the image analysis result according to a preset target calculation formula to obtain a comprehensive analysis result of the image to be analyzed; Determining whether the comprehensive analysis result matches the image analysis result; When it is determined that the comprehensive analysis result does not match the image analysis result, performing an update operation on the image analysis result according to the comprehensive analysis result to update the image analysis result of the image to be analyzed; Wherein, the preset target calculation formula includes one or more of a saturation score calculation formula, a brightness score calculation formula, a color diversity score calculation formula, and a sharpness score calculation formula.

5. The intelligent analysis method for image quality according to claim 3, wherein Before performing a training operation on the first image analysis model based on a pre-determined image dataset to obtain a trained second image analysis model, the method further includes: Obtain the model training information of the first image analysis model, where the model training information includes one or more of the training times information, training duration information, and training convergence degree information of the first image analysis model; Determine the dynamic training adjustment parameters of the first image analysis model according to the model training information, where the dynamic training adjustment parameters include the dynamic learning rate of the first image analysis model; Wherein, performing a training operation on the first image analysis model based on a pre-determined image dataset to obtain a trained second image analysis model includes: Perform a training operation on the first image analysis model based on the pre-determined image dataset and the dynamic training adjustment parameters to obtain a trained second image analysis model.

6. The intelligent analysis method for image quality according to claim 3, characterized in that, For each of the training images, determining at least one target interest point of the training image includes: For each of the training images, perform a heat statistics operation on the training image based on a preset attention mechanism to obtain the heat statistics parameters of the training image; For each of the training images, determine the target region of the training image according to the heat statistics parameters of the training image, and determine the target interest point of the training image according to the target region of the training image.

7. The intelligent analysis method for image quality according to claim 1, characterized in that Calculating the model loss parameters corresponding to the second image analysis model according to a preset loss calculation function includes: Input at least one preset test image into the second image analysis model to obtain a test output result; Calculate the probability distribution difference parameters corresponding to the second image analysis model based on the preset loss calculation function and all the test output results, and determine the model loss parameters corresponding to the second image analysis model based on the probability distribution difference parameters.

8. The intelligent analysis method for image quality according to claim 1, wherein The method further includes: When it is determined that the model loss parameters do not meet the preset model training conditions, determine the model update training parameters of the second image analysis model according to the model loss parameters and the preset model training conditions, perform a training operation on the second image analysis model according to the model update training parameters, and re-trigger the steps of calculating the model loss parameters corresponding to the second image analysis model according to the preset loss calculation function and determining whether the model loss parameters meet the preset model training conditions.

9. The intelligent analysis method for image quality according to claim 2, characterized in that For each of the training images, performing an image scoring operation on the training image to obtain the image score result of the training image includes: For each of the training images, send the training image to a number of target users, so that each target user performs a scoring operation on the training image to obtain the scoring result of the target user on the training image, and enable the target user to feedback the scoring result of the training image; Determine the image score result of the training image according to the rating results feedback by each of the target users; wherein, the image score result of the training image includes the rating results of all the target users for the training image.

10. The intelligent analysis method for image quality according to claim 2, wherein, For each of the training images, perform a score evaluation calculation operation on the training image according to the image score result of the training image to obtain the image evaluation result of the training image, including: For each of the training images, generate the score probability distribution parameter of the training image according to the rating results of each of the target users included in the training image for the training image. For each of the training images, perform a score evaluation calculation operation on the training image according to the score probability distribution parameter of the training image and all the rating results corresponding to the training image to obtain the image evaluation result of the training image; wherein, the score evaluation calculation operation includes a variance statistical operation.

11. The intelligent analysis method for image quality according to claim 3, characterized in that, For each of the training images, the determination of the target label information of the training image according to each of the target interest points and the offset parameter corresponding to each of the target interest points includes: For each of the target interest points in the training image, generate the background distribution information and the key point distribution information of the target interest point according to all the target interest points and the offset parameter corresponding to each of the target interest points. Determine the target label information of the training image according to the background distribution information and the key point distribution information of each of the target interest points.

12. The intelligent analysis method for image quality according to claim 5, wherein The determination of the dynamic training adjustment parameter of the first image analysis model according to the model training information includes: Determine the training stage of the first image analysis model according to the model training information. According to the training stage of the first image analysis model, determine the learning rate matching the training stage of the first image analysis model from a preset adjustment parameter database. Determine the dynamic training adjustment parameter of the first image analysis model according to the learning rate matching the training stage of the first image analysis model; wherein, the learning rates corresponding to different training stages are different.

13. The intelligent analysis method for image quality according to claim 6, wherein For each of the training images, the determination of the target area of the training image according to the heat statistic parameter of the training image includes: For each of the training images, generate a heat map corresponding to the training image according to the heat statistic parameter of the training image. For each of the training images, determine the target area with a heat value greater than or equal to a preset heat threshold from the heat map corresponding to the training image.

14. The intelligent analysis method for image quality according to claim 7, characterized in that, The calculation of the probability distribution difference parameter corresponding to the second image analysis model based on a preset loss calculation function and all the test output results includes: Determine the true label result of each of the test images, and calculate the image loss parameter between the true label result and the test output result of each of the test images based on a preset loss calculation function and the test output result of each of the test images. Generate the probability distribution difference parameter corresponding to the second image analysis model according to the image loss parameters of all the images to be tested.

15. An electronic device, characterized in that, Comprising a memory and a processor, the memory stores a computer program, when the computer program is executed by the processor, the processor is caused to execute the intelligent analysis method for image quality according to any one of claims 1 to 14.

16. A network attached storage device, characterized in that, The network attached storage device is at least configured with a display screen, and the network attached storage device stores a memory with executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory and executes the intelligent analysis method for image quality according to any one of claims 1-14.

17. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, when the computer instructions are called, they are used to execute the intelligent analysis method for image quality according to any one of claims 1-14.

18. An intelligent analysis system for image quality, characterized in that, The system includes the electronic device according to claim 15, and the electronic device is communicatively connected to the network attached storage device according to claim 16, and a computer application program capable of executing the method according to any one of claims 1-14 is installed in the electronic device.

Images