Image quality model generation, evaluation method and system, and device and medium thereof

Abstract

The present invention provides a method and system for generating and evaluating an image quality model, and its equipment and medium. The method for generating the image quality model includes: constructing a training image sample set with N types of image features, where N is a positive integer, and inputting the training image samples in the training image sample set into a convolutional neural network model to be trained, wherein the convolutional neural network model to be trained includes at least two deep neural network basic models, extracting features from the training image samples using the convolutional neural network model to be trained to obtain training image feature information, calculating the loss value between the training image feature information and the standard image feature information using a loss function, adjusting the parameters in the convolutional neural network model to be trained according to the loss value, and generating an image quality model including the parameters. The image quality model provided by the present invention is used to assist test engineers in evaluating the imaging quality of terminal cameras, thereby improving test efficiency and evaluation accuracy.

Description

Technical Field

[0001] The present invention relates to the field of image testing technology, and in particular to a method and system for generating and evaluating an image quality model, and a device and medium thereof. Background Art

[0002] As competition in the smartphone market becomes increasingly fierce, the camera mode of smartphones largely determines the key to the success or failure of the competition among manufacturers.

[0003] Nowadays, people take photos in their daily lives and when traveling. The number of photos stored in personal albums is growing explosively. This requires image test engineers to measure visual perception when adjusting the camera modes of smartphones. However, even professional photographers find it difficult to explain which features have a greater impact on the aesthetic quality of images. Although aesthetics are difficult to describe, image test engineers still evaluate images based on basic factors such as auto-exposure, white balance, and auto-focus.

[0004] The current image quality assessment method mainly relies on image test engineers to evaluate the quality of camera imaging based on their personal subjective preferences. This method is inefficient, highly subjective, and the reliability of the assessment results is insufficient. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for generating and evaluating an image quality model, as well as a device and medium thereof, so as to improve the efficiency of image quality evaluation and the accuracy of evaluation results.

[0006] To achieve the above-mentioned purpose, in a first aspect, the present invention provides a method for generating an image quality model, the method comprising: constructing a training image sample set with N types of image features, N being a positive integer, and inputting the training image samples in the training image sample set into a convolutional neural network model to be trained, the convolutional neural network model to be trained comprising at least two deep neural network basic models, and then using the convolutional neural network model to be trained to perform feature extraction on the training image samples to obtain training image feature information, using a loss function to calculate the loss value between the image feature information of the training image samples and the image standard feature information of the verification image samples, adjusting the parameters in the convolutional neural network model to be trained according to the loss value, and generating an image quality model including the parameters.

[0007] The beneficial effect of the method for generating an image quality model provided by the present invention is that: a set of training image samples of N types of image features are input into a convolutional neural network model to be trained for training, and features of the training image samples are extracted using the convolutional neural network model to be trained. Since the convolutional neural network model to be trained includes at least two deep neural network basic models, the training image feature information is relatively comprehensive. The loss value between the training image feature information and the standard image feature information is calculated by a loss function, and the parameters in the convolutional neural network model to be trained are adjusted according to the loss value, thereby generating an image quality model of the parameters. Since the parameters in the image quality model are adjusted according to the loss value, when the image quality model is used to evaluate the image to be tested, the accuracy of the evaluated image quality can be improved.

[0008] Optionally, the convolutional neural network model to be trained includes at least two deep neural network basic models, including: the convolutional neural network model to be trained includes GoogleNet, ResNet, MobileNet and ResMobile, wherein the ResMobile includes a residual network module and a depth-separable convolution module. Its beneficial effect is that: the convolutional neural network model to be trained includes GoogleNet, ResNet, MobileNet and ResMobile, and ResMobile includes a residual network module and a depth-separable convolution module, that is, the residual idea in ResNet is applied to the MobileNet network. Specifically, first, the ResNet approach is used for reference, and parallel 1×1 convolution branches and identity mapping branches are added to each convolution block during training to form a ResMobileBlock, and then the ResMobileBlock is converted into a convolution, making the architecture more flexible and easy to change the width of each layer. Multiple single-channel architectures have high parallelism and can save memory consumption.

[0009] Optionally, the training image sample set includes training image samples with different exposure compensations, training image samples with different focusing modes, and training image samples with different white balances. The beneficial effect is that by inputting the training image samples with different exposure compensations, training image samples with different focusing modes, and training image samples with different white balances into the convolutional neural network model to be trained, various image features are used for training, thereby improving the reliability of the convolutional neural network model to be trained.

[0010] Optionally, the generation of the image quality model including the parameters includes: when the number of iterations of the convolutional neural network model to be trained reaches a set value or the loss value of the loss function reaches a target value, the image quality model including the parameters is generated. The beneficial effect is that by presetting the number of iterations, when the preset number of iterations is reached, the training is stopped, thereby avoiding cyclic training during long iterative operations. Alternatively, the parameters in the convolutional neural network model to be trained are adjusted according to the loss value until the loss value calculated using the loss function reaches the preset target value, thereby obtaining the image quality model including the parameters, thereby achieving fast and reliable generation of the image quality model.

[0011] In a second aspect, an embodiment of the present invention provides an image quality model evaluation method, which is applied to the above-mentioned image quality model, and the method includes:

[0012] A test image is input into the image quality model to obtain an output result of the image quality model, wherein the output result includes reference scores corresponding to N types of image features of the test image; and a weighted sum is performed on the reference scores corresponding to the N types of image features to obtain an image quality score of the test image.

[0013] The beneficial effect of the image quality model evaluation method provided by the present invention is that: by inputting a test image into the image quality model, an output result in the image quality model is obtained, and the output result includes reference scores corresponding to N types of image features in the test image, and the reference scores corresponding to the N types of image features are weighted summed to obtain an image quality score for the test image, thereby improving the efficiency of image quality evaluation and the accuracy of the evaluation results. The high or low scores make it easier for test engineers to quickly understand the quality of the corresponding images.

[0014] In a third aspect, the present invention provides a system for generating an image quality model, comprising:

[0015] A construction unit is used to construct a training image sample set with N types of image features, where N is a positive integer; a processing unit is electrically connected to the construction unit, and the processing unit is used to input training image samples in the training image sample set into a convolutional neural network model to be trained, wherein the convolutional neural network model to be trained includes at least two deep neural network basic models, and the convolutional neural network model to be trained is used to perform feature extraction on the training image samples to obtain training image feature information. An acquisition unit is electrically connected to the processing unit, and the acquisition unit is used to obtain standard image feature information of verification image samples. A calculation unit is electrically connected to the acquisition unit, and is used to calculate a loss value between the training image feature information and the standard image feature information using a loss function. The processing unit is also used to adjust parameters in the convolutional neural network model to be trained according to the loss value to generate an image quality model including the parameters.

[0016] The beneficial effect of the image quality model generation system provided by the present invention is that: a set of training image samples of N types of image features are input into the convolutional neural network model to be trained through a processing unit, and features of the training image samples are extracted using the convolutional neural network model to be trained. Since the convolutional neural network model to be trained includes at least two deep neural network basic models, the training image feature information is relatively comprehensive. Then, the calculation unit calculates the loss value between the image feature information of the training image samples and the image standard feature information of the verification image samples through a loss function. Finally, the processing unit adjusts the parameters in the convolutional neural network model to be trained according to the loss value, thereby generating an image quality model of the parameters.

[0017] Optionally, the convolutional neural network model to be trained includes at least two deep neural network basic models, including: the convolutional neural network model to be trained includes GoogleNet, ResNet, MobileNet and ResMobile, wherein ResMobile includes a residual network module and a deep separable convolution module. Its beneficial effect is that the convolutional neural network model to be trained in the processing unit includes GoogleNet, ResNet, MobileNet and ResMobile, and ResMobile includes a residual network module and a deep separable convolution module, that is, the residual idea in ResNet is applied to the MobileNet network. Specifically, first, the ResNet approach is used for reference, and parallel 1×1 convolution branches and identity mapping branches are added to each convolution block during MobileNet network training to form a ResMobileBlock, and then the ResMobileBlock is converted into a convolution, making the architecture more flexible and easy to change the width of each layer. Multiple single-channel architectures have high parallelism and can save memory consumption.

[0018] Optionally, the training image sample set constructed by the construction unit includes exposure compensation training image samples, focus mode training image samples, and white balance training image samples. The beneficial effect is that by inputting the exposure compensation training image samples, focus mode training image samples, and white balance training image samples into the convolutional neural network model to be trained, various image features are used for training, thereby improving the reliability of the convolutional neural network model to be trained.

[0019] Optionally, the processing unit is used to generate an image quality model including the parameters when the number of iterations of the convolutional neural network model to be trained reaches a set value or the loss value of the loss function reaches a target value. Its beneficial effect is that by presetting the number of iterations, when the preset number of iterations is reached, the training is stopped, thereby avoiding cyclic training during long iterative operations. Alternatively, the parameters in the convolutional neural network model to be trained are adjusted according to the loss value until the loss value calculated using the loss function reaches the preset target value, thereby obtaining an image quality model including the parameters, thereby achieving fast and reliable generation of image quality models.

[0020] In a fourth aspect, an embodiment of the present invention provides an evaluation system for an image quality model, which is applied to the generation system of the above-mentioned image quality model, and the evaluation system includes:

[0021] The input unit is used to input the test image into the image quality model; the acquisition unit is electrically connected to the input unit, and is used to acquire the output result of the image quality model, wherein the output result includes the reference scores corresponding to the N types of image features of the test image. The calculation unit is electrically connected to the acquisition unit, and is used to perform weighted summation of the reference scores corresponding to the N types of image features to obtain the image quality score of the test image.

[0022] The beneficial effects of the image quality model evaluation system provided by the present invention are: a test image is input into the above-mentioned image quality model through an input unit to test the test image, an acquisition unit acquires the output result of the image quality model, and the output result includes reference scores corresponding to N types of image features of the test image, and finally a calculation unit performs weighted summation on the reference scores corresponding to the N types of image features to obtain the image quality score of the test image, which can assist test engineers in testing and evaluating image quality and improve the efficiency of image quality evaluation and the accuracy of evaluation results.

[0023] In a fifth aspect, the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method steps in the first aspect when executing the computer program.

[0024] The beneficial effect of the electronic device of the present invention is that the above method is implemented by executing the computer program through the processor.

[0025] In a sixth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the method steps in the above-mentioned first aspect when executed by a processor.

[0026] The beneficial effect of the computer-readable storage medium of the present invention is that the operation of the above method is realized by executing the computer program. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flow chart of a method for generating an image quality model provided for implementation of the present invention;

[0028] Figure 2 A schematic diagram of the first three layers of the MobileNet network structure provided for the implementation of the present invention;

[0029] Figure 3 A schematic diagram of applying the residual idea to one layer of MobileNet provided for the implementation of the present invention;

[0030] Figure 4 A schematic diagram of a data set of training image samples with different exposure compensations corresponding to verification image samples with different exposure compensations provided by an embodiment of the present invention;

[0031] Figure 5 A flowchart of a method for evaluating an image quality model provided by an embodiment of the present invention;

[0032] Figure 6 A schematic diagram of an image quality model evaluation method provided by an embodiment of the present invention;

[0033] Figure 7 A schematic diagram of an operation interface and an execution display interface of an evaluation system for an image quality model provided by an embodiment of the present invention;

[0034] Figure 8 A schematic diagram of the structure of a system for generating an image quality model provided by an embodiment of the present invention;

[0035] Fig. 9 A schematic diagram of the structure of an image quality model evaluation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein should be understood by people with general skills in the field to which the present invention belongs. "Including" and similar words used in this article mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0037] As the competition in the smartphone market becomes increasingly fierce, the camera mode of smartphones largely determines the key to the success or failure of each manufacturer, so the quality of the pictures taken by the camera determines the size of the market. However, the current image quality evaluation method mainly relies on image test engineers to evaluate the quality of camera imaging based on their personal subjective preferences, which is inefficient, highly subjective, and the reliability of the evaluation results is insufficient.

[0038] In view of the existing problems, an embodiment of the present invention provides a method for generating an image quality model. Figure 1 As shown, the method includes:

[0039] S101: Construct a training image sample set with N types of image features, where N is a positive integer.

[0040] In this step, the constructed training set of N types of image features mainly includes training image samples with different exposure compensations, training image samples with different focusing modes, and training image samples with different white balances.

[0041] S102: Inputting training image samples in the training image sample set into a convolutional neural network model to be trained, wherein the convolutional neural network model to be trained includes at least two deep neural network basic models.

[0042] In this step, training image samples including exposure compensation, training image samples of focus mode and training image samples of white balance are input into the convolutional neural network model to be trained, and various picture features are used for training, thereby improving the reliability of the training of the convolutional neural network model to be trained.

[0043] In addition, the convolutional neural network model to be trained includes GoogleNet, ResNet, MobileNet and ResMobile. In this embodiment, the convolutional neural network model to be trained consists of GoogleNet, ResNet, MobileNet and ResMobile, wherein the ResMobile includes a residual network module and a depthwise separable convolution module. Figure 2 and Figure 3 As shown, Figure 2 It is a schematic diagram of the network structure of the first three layers of MobileNet. Figure 3 The schematic diagram of applying the residual idea to one of the layers of MobileNet, that is, the ResMobile applies the residual idea in ResNet to the MobileNet network. Specifically, first, drawing on the ResNet approach, parallel 1×1 convolution branches and identity mapping branches are added to each convolution block during training to form a ResMobileBlock, and then the ResMobileBlock is converted into a convolution, making the architecture more flexible and easy to change the width of each layer. The multiple single-channel architectures have high parallelism and can save memory consumption.

[0044] S103: Using the convolutional neural network model to be trained to perform feature extraction on the training image samples to obtain training image feature information.

[0045] In this step, the training image samples are feature extracted by the convolutional neural network model to be trained to obtain training image feature information. Because in this embodiment, the convolutional neural network model to be trained is composed of the GoogleNet, the ResNet, the MobileNet and the ResMobile, the image feature information will be more comprehensive and detailed.

[0046] S104: Obtain standard image feature information of the verification image sample.

[0047] By obtaining the standard image feature information of the verification image sample, a benchmark is provided for the subsequent adjustment of the parameters in the convolutional neural network model to be trained.

[0048] S105: Calculate the loss value between the training image feature information and the standard image feature information using a loss function.

[0049] In this step, the reliability of parameter adjustment can be improved by calculating the loss value between the training image feature information and the standard image feature information, and using the obtained actual loss value for subsequent parameter adjustment.

[0050] Further, combined with Figure 4 As shown, Figure 4 The diagram is a diagram of a dataset of validation image samples with different exposure compensations corresponding to training image samples with different exposure compensations. The datasets in both the training image samples and the validation image samples include nine categories of images: "overall too dark, overall darker, some areas too dark, some areas darker, overall too bright, overall brighter, some areas too bright, some areas brighter, and normal". For example, the image feature information of "overall too dark" in the training image samples and the validation image samples is first calculated using a loss function to obtain the mutual loss values, and the datasets in the training image samples and the validation image samples are calculated in turn according to the scheme.

[0051] It should be noted that the loss values ​​between the training image samples of different focus modes and the verification image samples of different focus modes, and the loss values ​​between the training image samples of different white balances and the verification image samples of different white balances are also calculated using the above method and are not explained separately here.

[0052] S106: Adjust parameters in the convolutional neural network model to be trained according to the loss value.

[0053] In this step, the convolutional neural network model to be trained is optimized by continuously adjusting parameters.

[0054] S107: Generate an image quality model including the parameters.

[0055] In this embodiment, a set of training image samples of N types of image features are input into a convolutional neural network model to be trained for training, and features of the training image samples are extracted using the convolutional neural network model to be trained. Since the convolutional neural network model to be trained includes at least two deep neural network basic models, the training image feature information is relatively comprehensive. The loss value between the training image feature information and the standard image feature information is calculated by a loss function, and the parameters in the convolutional neural network model to be trained are adjusted according to the loss value, thereby generating an image quality model with adjusted parameters.

[0056] In another embodiment disclosed in the present invention, a method for evaluating an image quality model is provided, referring to Figure 5 As shown, the method is applied to the image quality model implemented above, and the method includes:

[0057] S501: Input a test image into the image quality model.

[0058] In this step, combine Figure 6 As shown, Figure 6It is a schematic diagram of the evaluation method of the image quality model. A test picture is input into the image quality model. The image quality model includes trained GoogleNet, ResNet, MobileNet and ResMobile. Each test picture needs to be input into GoogleNet, ResNet, MobileNet and ResMobile.

[0059] S502: Obtain an output result of the image quality model, wherein the output result includes reference scores corresponding to N types of image features of the test image.

[0060] In this step, the output result of the image quality model is obtained, the features are aggregated using a statistical aggregation structure, and the probability obtained by a logistic regression model (softmax) represents the reference score.

[0061] S503: Perform weighted summation on the reference scores corresponding to the N types of image features to obtain an image quality score of the test image.

[0062] Exemplarily, the reference scores corresponding to N types of image features include a reference score corresponding to the exposure compensation image feature in the test image, a reference score corresponding to the focus mode, and a reference score corresponding to the white balance. By taking a weighted sum of the reference scores corresponding to these types of image features, the image quality score of the test image can be obtained.

[0063] In this step, in order to make the reference score more directly convenient for the test engineer to observe, the image quality score of the test image is obtained by weighted summing the reference score. The test engineer can directly evaluate the quality of the test image according to the size of the image quality score of the test image.

[0064] In this embodiment, the test image is input into the image quality model to obtain the output result of the image quality model, and the output result includes a reference score corresponding to the exposure compensation image feature in the test image, a reference score corresponding to the focus mode, and a reference score corresponding to the white balance. Then, the reference scores corresponding to these types of image features are weightedly summed to obtain the image quality score of the test image, thereby improving the efficiency and accuracy of image quality assessment. The high and low scores enable test engineers to quickly understand the quality of the corresponding images.

[0065] Specific, combined Figure 7 As shown, Figure 7It is the operation interface and execution display interface of the image quality model evaluation system. The operation interface includes a training image sample set saved to path 1, a verification image sample set saved to path 2, an image quality model saved to path 3, a selection button, a training button, a training display area, an Epoch input box, a Batch-Size input box, a convolutional neural network model to be trained, and a graphics processing unit (GPU).

[0066] When training the image quality model, select the training image sample set, the verification image sample set, and the convolutional neural network model to be trained by selecting the button, then enter the number of calculations and the number of images in the iteration number input box and the size of each batch of samples input box respectively, and finally click the button to implement the image quality model training task in the image processor. The training display area will always display the specific situation of the training, which is convenient for the test engineer to observe. When the training is completed, the generated image quality model will be saved to path 3.

[0067] Furthermore, the execution display interface includes a folder display area, a folder or file display area, a picture display area, a path 4 corresponding to the generated image quality model, a path 5 corresponding to the test image, an identification button, and a batch identification button. Among them, the folder display area is used to assist the test engineer in finding the location of the test image file. When preparing to perform image quality assessment, select the corresponding test image in path 5. The folder or file display area will display the folder and image of the corresponding test image in path 5, and load the image quality model generated in the operation interface from path 4. By clicking the identification button or the batch identification button, the quality identification and evaluation of one image or batches of images will be performed. The image display area will display the test image, and the identification results of the evaluation will be displayed below the corresponding test image. The pictures in the picture display area can be freely enlarged and reduced, which is convenient for image testing engineering to view specific image details.

[0068] In another embodiment disclosed in the present invention, a system for generating an image quality model is provided, referring to Figure 8As shown, the system includes a construction unit 801, a processing unit 802, an acquisition unit 803 and a calculation unit 804, wherein the construction unit 801 is used to construct a training image sample set with N types of image features, where N is a positive integer, and the processing unit 802 is electrically connected to the construction unit 801, and is used to input the training image samples in the training image sample set to the convolutional neural network model to be trained, wherein the convolutional neural network model to be trained includes at least two deep neural network basic models, and the convolutional neural network model to be trained is used to extract features from the training image samples to obtain training image feature information, and the acquisition unit 803 is electrically connected to the processing unit 802, and the acquisition unit 803 is used to obtain standard image feature information of the verification image sample. The calculation unit 804 is used to calculate the loss value between the training image feature information and the standard image feature information using a loss function, and the processing unit 802 is also used to adjust the parameters in the convolutional neural network model to be trained according to the loss value to generate an image quality model including the parameters.

[0069] In this embodiment, a set of training image samples of N types of image features are input into the convolutional neural network model to be trained for training by the processing unit 802, and features of the training image samples are extracted using the convolutional neural network model to be trained. Since the convolutional neural network model to be trained includes at least two deep neural network basic models, the training image feature information is relatively comprehensive. Then, the calculation unit 804 calculates the loss value between the training image feature information and the standard image feature information through the loss function. Finally, the processing unit 802 adjusts the parameters in the convolutional neural network model to be trained according to the loss value, thereby generating an image quality model of the parameters. When the image quality model is used to test the test picture, the efficiency and accuracy of image quality assessment can be improved.

[0070] Optionally, the training image sample set constructed by the construction unit 801 includes training image samples for exposure compensation, training image samples for focus mode, and training image samples for white balance. By adopting various picture features for training, the reliability of the training of the convolutional neural network model to be trained is improved. In addition, the processing unit 802 is also used to generate an image quality model including the parameters when the number of iterations of the convolutional neural network model to be trained reaches a set value or the loss value of the loss function reaches a target value. By presetting the number of iterations, when the preset number of iterations is reached, the training is stopped, thereby avoiding cyclic training during long iterative operations, or adjusting the parameters in the convolutional neural network model to be trained according to the loss value until the loss value calculated by the loss function reaches the preset target value, the image quality model including the parameters is obtained, thereby achieving fast and reliable generation of image quality models.

[0071] In another embodiment disclosed in the present invention, an evaluation system for an image quality model is provided, referring to Fig. 9 As shown, the system is based on the generation system of the above-mentioned image quality model, and the evaluation system of the image quality model includes: an input unit 901 for inputting a test image into the image quality model, an acquisition unit 902 is electrically connected to the input unit 901, and is used to obtain an output result of the image quality model, the output result includes reference scores corresponding to N types of image features of the test image, and a calculation unit 903 is electrically connected to the acquisition unit 902, and is used to perform weighted summation of the reference scores corresponding to the N types of image features to obtain an image quality score of the test image.

[0072] In this embodiment, the test image is input into the above-mentioned image quality model through the input unit 901 to test the test image, and the acquisition unit 902 acquires the output result of the image quality model, which includes the reference scores corresponding to the N types of image features of the test image. Finally, the calculation unit 903 performs weighted summation on the reference scores corresponding to the N types of image features to obtain the image quality score of the test image, which can assist test engineers in testing and evaluating image quality and improve the efficiency of image quality evaluation and the accuracy of evaluation results.

[0073] In another embodiment disclosed in the present invention, based on the above embodiment, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for generating an image quality model as in the above embodiment is implemented.

[0074] In other embodiments of the present application, an embodiment of the present application discloses an electronic device, which may include: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can be used to execute the various steps in the corresponding embodiment.

[0075] The above is only a specific implementation of the embodiment of the present application, but the protection scope of the embodiment of the present application is not limited thereto, and any changes or replacements within the technical scope disclosed in the embodiment of the present application should be included in the protection scope of the embodiment of the present application. Therefore, the protection scope of the embodiment of the present application should be based on the protection scope of the claims.

Claims

1. A method for generating an image quality model, characterized in that: The method comprises: Construct a training image sample set with N types of image features, where N is a positive integer; Inputting the training image samples in the training image sample set into the convolutional neural network model to be trained; Using the convolutional neural network model to be trained to perform feature extraction on the training image samples to obtain training image feature information; Obtain standard image feature information of verification image samples; Calculating the loss value between the training image feature information and the standard image feature information using a loss function; Adjusting parameters in the convolutional neural network model to be trained according to the loss value; generating an image quality model including the parameters; Among them, the convolutional neural network model to be trained is composed of GoogleNet, ResNet, MobileNet and ResMobile; the ResMobile applies the residual idea in ResNet to the MobileNet network: first, drawing on the ResNet approach, parallel 1×1 convolution branches and identity mapping branches are added to each convolution block during training to form a ResMobileBlock, and then the ResMobileBlock is converted into a convolution.

2. The method according to claim 1, characterized in that The training image sample set includes exposure compensation training image samples, focus mode training image samples and white balance training image samples.

3. The method according to claim 1, characterized in that The generating of the image quality model including the parameters comprises: When the number of iterations of the convolutional neural network model to be trained reaches a set value or the loss value of the loss function reaches a target value, an image quality model including the parameters is generated.

4. A method for evaluating an image quality model, characterized in that: The evaluation method is applied to the image quality model according to any one of claims 1 to 3, and the method comprises: inputting a test image into the image quality model; Obtaining an output result of the image quality model, the output result including reference scores corresponding to N types of image features of the test image respectively; The reference scores corresponding to the N types of image features are weightedly summed to obtain an image quality score for the test image.

5. A system for generating an image quality model, used to implement the method for generating an image quality model according to any one of claims 1 to 3, characterized in that: The generation system comprises: A construction unit, used to construct a training image sample set having N types of image features, where N is a positive integer; a processing unit, electrically connected to the construction unit, and configured to input training image samples in the training image sample set into the convolutional neural network model to be trained, and perform feature extraction on the training image samples using the convolutional neural network model to be trained to obtain training image feature information; An acquisition unit, electrically connected to the processing unit, and configured to acquire standard image feature information of a verification image sample; A calculation unit, electrically connected to the acquisition unit, wherein the calculation unit calculates a loss value between the training image feature information and the standard image feature information using a loss function; The processing unit also adjusts the parameters in the convolutional neural network model to be trained according to the loss value to generate an image quality model including the parameters.

6. The system according to claim 5, characterized in that The convolutional neural network model to be trained consists of GoogleNet, ResNet, MobileNet and ResMobile; the ResMobile includes a residual network module and a depth-separable convolution module.

7. The system according to claim 5 or 6, characterized in that: The training image sample set constructed by the construction unit includes exposure compensation training image samples, focus mode training image samples and white balance training image samples.

8. The system according to claim 7, characterized in that The processing unit is used to generate an image quality model including the parameters when the number of iterations of the convolutional neural network model to be trained reaches a set value or the loss value of the loss function reaches a target value.

9. An image quality model evaluation system, characterized in that: The system is applied to the system for generating the image quality model according to any one of claims 5 to 8, and the system comprises: An input unit, used for inputting a test image into the image quality model; an acquisition unit, electrically connected to the input unit, and configured to acquire an output result of the image quality model, wherein the output result includes reference scores corresponding to N types of image features of the test image; A calculation unit is electrically connected to the acquisition unit and is used to perform weighted summation on the reference scores corresponding to the N types of image features to obtain an image quality score of the test image.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 3 or claim 4 is implemented.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 3 or claim 4 is implemented.

Images