Deep learning model training method and artificial intelligence-based image optimization method

Abstract

The invention discloses a deep learning model training method and an artificial intelligence-based image optimization method. The training method comprises the following steps: based on a plurality of convolution kernels with different sizes, respectively extracting global features and local features from a sample image; performing noise interference suppression on the global features through residual learning to obtain the global features after noise interference suppression, and performing super-resolution enhancement and detail compensation on the local features by using a generative adversarial network to obtain the local features after detail enhancement; and performing multi-layer fusion processing on the global features after noise interference suppression and the local features after detail enhancement to obtain an optimized sample image, comparing the optimized sample image with a noise-free image, and updating network parameters of the deep learning model according to a comparison result to train the deep learning model. The technical problem that image details are blurred due to the fact that the image denoising effect is not good is solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a training method for a deep learning model and an image optimization method based on artificial intelligence. Background Art

[0002] In recent years, with the continuous development of image processing technology, image optimization methods based on deep learning have been widely used in the fields of image denoising, resolution improvement and detail enhancement. However, existing image optimization technologies still face many challenges, especially in noise suppression and detail recovery. The noise in the image often interferes with the extraction of global features, causing the overall clarity of the image to decrease. While suppressing noise, existing denoising methods are often accompanied by the loss of image details, resulting in blurred image details and unable to meet the requirements of high-quality image optimization. In addition, the lack of image detail compensation is also one of the bottlenecks of current technology. For complex scenes or high-resolution images, it is often difficult for the model to accurately restore the details, resulting in insufficient clarity and fineness of the output image.

[0003] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0004] The embodiments of the present invention provide a deep learning model training method and an artificial intelligence-based image optimization method to at least solve the technical problem of blurred image details caused by poor image denoising effect.

[0005] According to one aspect of an embodiment of the present invention, a training method for a deep learning model for image optimization is provided, comprising: extracting global features and local features from a sample image based on a plurality of convolution kernels of different sizes; performing noise interference suppression on the global features through residual learning to obtain the global features after noise interference suppression, and performing super-resolution enhancement and detail compensation on the local features using a generative adversarial network to obtain the local features after detail enhancement; The global feature after noise interference suppression and the local feature after detail enhancement are subjected to multi-layer fusion processing to obtain an optimized sample image, the optimized sample image is compared with the noise-free image, and the network parameters of the deep learning model are updated according to the comparison result to train the deep learning model; wherein, the global feature is subjected to noise interference suppression by residual learning, including: calculating the minimum Euclidean distance between the global feature and the reference feature of the noise-free image, and obtaining the Euclidean distance ratio based on the minimum Euclidean distance; when the difference between the Euclidean distance ratio and 1 is greater than a preset difference threshold, the feature points of the global feature are sorted based on the Euclidean distance ratio, and the top N feature points with the highest quality are retained as a hypothetical generated set; when the difference between the Euclidean distance ratio and 1 is less than the preset difference threshold, the quality of the feature point is evaluated by using a quality factor, and the hypothetical generated set is generated based on the evaluated feature point, wherein N is an integer greater than 1; based on the hypothetical generated set, the residual value between the global feature and the reference feature is determined by using residual learning, and the global feature is subjected to noise interference suppression based on the residual value.

[0006] According to another aspect of an embodiment of the present invention, an artificial intelligence-based image optimization method is also provided, including: obtaining an image to be processed; using the above-mentioned deep learning model to optimize the image to be processed to obtain an optimized image.

[0007] According to another aspect of an embodiment of the present invention, a training device for a deep learning model for image optimization is also provided, comprising: an extraction template, configured to extract global features and local features from a sample image based on multiple convolution kernels of different sizes; a denoising module, configured to suppress noise interference on the global features through residual learning to obtain the global features after noise interference suppression, and to perform super-resolution enhancement and detail compensation on the local features using a generative adversarial network to obtain the local features after detail enhancement; an updating module, configured to perform multi-layer fusion processing on the global features after noise interference suppression and the local features after detail enhancement to obtain an optimized sample image, compare the optimized sample image with a noise-free image, and update the network parameters of the deep learning model according to the comparison result to train the deep learning model. Type; wherein the denoising module is further configured to: calculate the minimum Euclidean distance between the global feature and the reference feature of the noise-free image, and obtain the Euclidean distance ratio based on the minimum Euclidean distance; when the difference between the Euclidean distance ratio and 1 is greater than a preset difference threshold, sort the feature points of the global feature based on the Euclidean distance ratio, and retain the top N feature points with the highest quality as a hypothetical generation set; when the difference between the Euclidean distance ratio and 1 is less than the preset difference threshold, use the quality factor to evaluate the quality of the feature point, and generate the hypothetical generation set based on the evaluated feature point, wherein N is an integer greater than 1; based on the hypothetical generation set, use residual learning to determine the residual value between the global feature and the reference feature, and suppress noise interference on the global feature based on the residual value.

[0008] In an embodiment of the present invention, global features and local features are extracted from a sample image based on multiple convolution kernels of different sizes; noise interference is suppressed on the global features through residual learning to obtain the global features after noise interference suppression, and super-resolution enhancement and detail compensation are performed on the local features using a generative adversarial network to obtain the local features after detail enhancement; multi-layer fusion processing is performed on the global features after noise interference suppression and the local features after detail enhancement to obtain an optimized sample image, and the optimized sample image is compared with the noise-free image, and the network parameters of the deep learning model are updated according to the comparison results to train the deep learning model. The above scheme solves the technical problem of image detail blurring caused by poor image denoising effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0010] Figure 1 is a flowchart of a training method for a deep learning model for image optimization according to an embodiment of the present invention;

[0011] Figure 2 is a flowchart of another method for training a deep learning model for image optimization according to an embodiment of the present invention;

[0012] Figure 3 is a flow chart of a method for extracting features according to an embodiment of the present invention;

[0013] Figure 4 is a flow chart of a method for suppressing noise interference based on global features according to an embodiment of the present invention;

[0014] Figure 5 is a flow chart of a method for super-resolution enhancement and detail compensation of local features according to an embodiment of the present invention;

[0015] Figure 6 is a structural diagram of a training device for a deep learning model for image optimization according to an embodiment of the present invention;

[0016] Figure 7 A schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0017] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not 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 should fall within the scope of protection of the present invention.

[0018] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0019] According to an embodiment of the present invention, a method embodiment of a training method for a deep learning model for image optimization is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in an order different from that shown here.

[0020] Figure 1 is a training method for a deep learning model for image optimization according to an embodiment of the present invention, such as Figure 1 As shown, the method comprises the following steps:

[0021] Step S102, based on a plurality of convolution kernels of different sizes, extracting global features and local features from the sample image respectively.

[0022] Step S104, performing noise interference suppression on the global features through residual learning to obtain the global features after noise interference suppression, and performing super-resolution enhancement and detail compensation on the local features using a generative adversarial network to obtain the local features after detail enhancement.

[0023] First, the global feature is subjected to noise interference suppression by residual learning to obtain the global feature after noise interference suppression. For example, the minimum Euclidean distance between the global feature and the reference feature of the noise-free image is calculated, and the Euclidean distance ratio is obtained based on the minimum Euclidean distance; when the difference between the Euclidean distance ratio and 1 is greater than a preset difference threshold, the feature points of the global feature are sorted based on the Euclidean distance ratio, and the top N feature points with the highest quality are retained as the assumed generation set; when the difference between the Euclidean distance ratio and 1 is less than the preset difference threshold, the quality of the feature point is evaluated by using a quality factor, and the assumed generation set is generated based on the evaluated feature point, wherein N is an integer greater than 1, for example, the quality factor is calculated based on the Euclidean distance ratio and the minimum Euclidean distance; the feature points whose quality factor is greater than the preset factor threshold are retained as the assumed generation set; based on the assumed generation set, the residual value between the global feature and the reference feature is determined by residual learning, and the global feature is subjected to noise interference suppression based on the residual value. For example, by utilizing the residual learning, the residual value between the global feature and the reference feature is determined by analyzing the feature points in the assumed generation set one by one; and by utilizing an adaptive learning rate, a weight greater than a preset weight threshold is applied to the global feature whose residual value is greater than a preset residual threshold, so as to suppress noise interference to the global feature.

[0024] Next, the local features are super-resolution enhanced and detail compensated using a generative adversarial network to obtain the local features after detail enhancement. For example, in the generator of the generative adversarial network, the local features are input into a multi-layer convolution layer, and features are gradually extracted from the local features through multi-layer convolution operations to form an intermediate feature representation; a deconvolution operation is applied to the intermediate feature representation to reconstruct the super-resolution features layer by layer to form a resolution feature; in the discriminator of the generative adversarial network, the similarity between the resolution feature and the true resolution local feature is evaluated to obtain a discriminant loss, and based on the discriminant loss, the resolution feature is compensated for details to obtain the local features after detail enhancement.

[0025] Step S106, performing multi-layer fusion processing on the global features after noise interference suppression and the local features after detail enhancement to obtain an optimized sample image, comparing the optimized sample image with the noise-free image, and updating the network parameters of the deep learning model according to the comparison result to train the deep learning model.

[0026] First, a fusion process is performed. The global features after noise interference suppression and the local features after detail enhancement are matched and aligned using a spatial pyramid pooling method; the matched and aligned global features and the local features are fused in multiple layers using an attention mechanism. Next, the network parameters are updated. The optimized sample image is compared with the noise-free image, and the network parameters of the deep learning model are updated according to the comparison results.

[0027] This application uses convolution kernels of different sizes to extract global features and local features respectively, thus taking into account both the overall structure and detail performance of the image. At the same time, the residual learning method is used to suppress noise on global features to ensure the overall clarity of the image, while the generative adversarial network is used to perform super-resolution enhancement and detail compensation on local features to achieve accurate restoration of details.

[0028] This application also provides another method for training a deep learning model for image optimization, which can effectively improve the overall quality and detail performance of the image. Figure 2 As shown, the method comprises the following steps:

[0029] Step S202, extracting features.

[0030] Useful global features and local features are extracted from the sample image. In this embodiment, multiple convolution kernels of different sizes are used for feature extraction to take into account both the overall structure and detail performance of the image. Specifically, Figure 3 As shown, the following steps are included:

[0031] Step S2022, extracting global features.

[0032] The extraction of global features mainly focuses on the overall morphology and structure of the image, and requires a larger convolution kernel for processing. A large convolution kernel can capture a larger receptive field and extract the macro information of the image. In this embodiment, the convolution kernel size used can be 5x5 and 7x7, etc. The convolution kernel is slid on the sample image, a dot product operation is performed, and a feature map is generated. By applying convolution kernels of different sizes multiple times, multiple global feature maps can be obtained. After obtaining the global feature map, an activation function (such as ReLU) is applied to introduce nonlinear features so that the model can learn more complex image features. The global feature maps extracted by multiple convolution kernels are spliced ​​or weighted averaged to form a comprehensive global feature.

[0033] Step S2024, extracting local features.

[0034] The extraction of local features mainly focuses on the details in the image, such as textures, edges, and small objects. Since local features usually have a small spatial range, small-sized convolution kernels (such as 1x1 and 3x3) are selected for extraction. Small convolution kernels can accurately capture image details and handle changes in local information. For example, a 1x1 convolution kernel can be used to adjust the number of channels of a feature map, while a 3x3 convolution kernel can effectively extract edge information. When extracting local features, deeper local features can be extracted layer by layer by stacking multiple convolution layers to form a multi-level representation of local features. Afterwards, the extracted local features are normalized to ensure that the dimensions of each feature are consistent, which promotes the effectiveness of subsequent fusion processing.

[0035] Step S204, suppressing noise interference of global features.

[0036] Residual learning is used to suppress noise interference on the extracted global features to ensure the improvement of the overall clarity of the image. Figure 4 As shown, the noise interference suppression of global features includes the following steps:

[0037] Step S2042, calculating the Euclidean distance ratio.

[0038] First, calculate the minimum Euclidean distance between the global feature and the reference feature of the noise-free image to measure the quality of the current global feature. Calculate the Euclidean distance between the feature points of the global feature and the feature points of the reference feature of the noise-free image, and find the minimum Euclidean distance and the second smallest Euclidean distance , and calculate the Euclidean distance ratio according to the following formula:

[0039] .

[0040] Step S2044, evaluating the quality of the feature points.

[0041] After obtaining the Euclidean distance ratio, the quality of the feature points is evaluated based on the difference between the Euclidean distance ratio and 1.

[0042] When the difference between the Euclidean distance ratio and 1 is greater than a preset difference threshold (eg, 0.01), the feature points of the global feature are sorted based on the Euclidean distance ratio, and the top N feature points with the highest quality are retained as the hypothetical generation set.

[0043] When the difference between the Euclidean distance ratio and 1 is less than the preset difference threshold (i.e. ), the quality factor is used to evaluate the quality of the feature points, and a hypothesis generation set is generated based on the evaluated feature points. When the quality of feature points is not significantly different, this may lead to feature points of different qualities being treated equally. In this case, the quality factor can be introduced Realize the evaluation of feature point quality. The larger the value, the smaller the Euclidean distance of the feature point, and the higher the quality. In some embodiments, feature points whose quality factor is greater than a preset factor threshold (e.g., a value between 1 and 3) can be retained as a hypothesis generation set. It can be obtained by:

[0044] .

[0045] Step S2046, residual learning and noise suppression.

[0046] After the feature points are screened, the residual learning method is used to suppress noise interference.

[0047] First, the residual value is calculated. Using the feature points in the assumed generated set, the residual value between the global feature and the reference feature is calculated. Then, for the global features whose residual value is greater than the preset residual threshold, an adaptive learning rate strategy is applied to set the weight value to strengthen the learning of high residual features and ensure effective suppression of noise interference: W = α⋅(R res >T res ), where R res is the calculated residual value, T res is the preset residual threshold, α is the weight adjustment factor, and W is the weight.

[0048] Step S206: super-resolution enhancement and detail compensation of local features.

[0049] After suppressing the global feature noise interference, the Generative Adversarial Network (GAN) is used to perform super-resolution enhancement and detail compensation on local features. Figure 5 As shown, the following steps are included:

[0050] Step S2062: super-resolution enhancement.

[0051] The generative adversarial network consists of a generator and a discriminator. The generator is used to generate high-quality images, while the discriminator is used to evaluate the similarity between the generated image and the real image. In the generator, local features are input into multiple convolutional layers, and features are gradually extracted using convolution operations to form intermediate feature representations. Deconvolution operations are applied to the intermediate feature representations to reconstruct high-resolution features layer by layer to form the final resolution features.

[0052] Step S2064, perform detail compensation.

[0053] After the generator generates high-resolution features, the resolution features are evaluated by the discriminator. The discriminator evaluates the similarity between the generated resolution features and the true resolution local features and calculates the discriminant loss. Specifically, the similarity is calculated based on the generated resolution features and the true resolution local features; then, the cross entropy loss is calculated based on the number of samples, similarity, and true labels; at the same time, the regularization term is calculated based on the generated features and the true features; then, the discriminant loss is calculated based on the cross entropy loss and the regularization term. Finally, according to the discriminant loss, the resolution features are compensated for details to ensure that the generated image is close to the true features in terms of detail level.

[0054] Step S208, updating network parameters.

[0055] After the global features are noise suppressed and the local features are detail enhanced, feature fusion and network parameter update are performed next.

[0056] Through multi-layer fusion processing, the global features after noise interference suppression are combined with the local features after detail enhancement to obtain the final optimized feature representation. For example, the spatial pyramid pooling method is used to match and align global features and local features. This method can effectively capture feature information at different levels through multi-scale processing. Then, the attention mechanism is used to fuse the matched and aligned global features with the local features, and the different features are integrated together through weighted summation to obtain the optimized sample image. Finally, the parameters of the network are updated through the back-propagation algorithm. For example, the loss function is calculated based on the output after feature fusion and the target output: the back-propagation algorithm is used to update the parameters in the network according to the loss function to minimize the loss function and improve the performance of the model.

[0057] Through the above method, a significant improvement in image quality is achieved through noise suppression of global features and super-resolution enhancement of local features, combined with feature fusion and network parameter updating.

[0058] The present application also provides an image optimization method based on artificial intelligence. First, an image to be processed is obtained, and the image to be processed is input into a deep learning model trained using the method of the above embodiment. The deep learning model is used to optimize the image to be processed to obtain an optimized image.

[0059] The present application also provides a training device for a deep learning model for image optimization, such as Figure 6 As shown, it includes: an extraction template 62, which is configured to extract global features and local features from the sample image based on multiple convolution kernels of different sizes; a denoising module 64, which is configured to suppress noise interference on the global features through residual learning to obtain the global features after noise interference suppression, and use a generative adversarial network to perform super-resolution enhancement and detail compensation on the local features to obtain the local features after detail enhancement; an updating module 66, which is configured to perform multi-layer fusion processing on the global features after noise interference suppression and the local features after detail enhancement to obtain an optimized sample image, compare the optimized sample image with the noise-free image, and update the network parameters of the deep learning model according to the comparison result to train the deep learning model.

[0060] Among them, the denoising module 64 is also configured to: calculate the minimum Euclidean distance between the global feature and the reference feature of the noise-free image, and obtain the Euclidean distance ratio based on the minimum Euclidean distance; when the difference between the Euclidean distance ratio and 1 is greater than a preset difference threshold, sort the feature points of the global feature based on the Euclidean distance ratio, and retain the top N feature points with the highest quality as the assumed generation set; when the difference between the Euclidean distance ratio and 1 is less than the preset difference threshold, use the quality factor to evaluate the quality of the feature points, and generate the assumed generation set based on the evaluated feature points, wherein N is an integer greater than 1; based on the assumed generation set, use residual learning to determine the residual value between the global feature and the reference feature, and suppress noise interference on the global feature based on the residual value.

[0061] It should be noted that the training device for the deep learning model for image optimization provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the training device for the deep learning model for image optimization provided in the above embodiment and the training method embodiment of the deep learning model for image optimization belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0062] Figure 7 FIG. 1 shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiment of the present disclosure. It should be noted that: Figure 7 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0063] like Figure 7 As shown, the electronic device includes a central processing unit (CPU) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage part 1008 into a random access memory (RAM) 1003. In the RAM 1003, various programs and data required for system operation are also stored. The CPU 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0064] The following components are connected to the I / O interface 1005: an input section 1006 including a keyboard, a mouse, etc.; an output section 1007 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, a modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as needed. A removable medium 1011, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1010 as needed, so that a computer program read therefrom is installed into the storage section 1008 as needed.

[0065] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A training method for a deep learning model for image optimization, characterized in that: include: Based on multiple convolution kernels of different sizes, global features and local features are extracted from the sample image respectively; Performing noise interference suppression on the global features through residual learning to obtain the global features after noise interference suppression, and performing super-resolution enhancement and detail compensation on the local features through a generative adversarial network to obtain the local features after detail enhancement; Performing multi-layer fusion processing on the global features after noise interference suppression and the local features after detail enhancement to obtain an optimized sample image, comparing the optimized sample image with the noise-free image, and updating the network parameters of the deep learning model according to the comparison result to train the deep learning model; The noise interference suppression of the global feature is performed through residual learning, including: Calculating a minimum Euclidean distance between the global feature and a reference feature of the noise-free image, and obtaining a Euclidean distance ratio based on the minimum Euclidean distance; In the case where the difference between the Euclidean distance ratio and 1 is greater than a preset difference threshold, the feature points of the global feature are sorted based on the Euclidean distance ratio, and the top N feature points with the highest quality are retained as the assumed generated set; in the case where the difference between the Euclidean distance ratio and 1 is less than the preset difference threshold, the quality of the feature points is evaluated using a quality factor, and the assumed generated set is generated based on the evaluated feature points, wherein N is an integer greater than 1; Based on the hypothesis generation set, residual learning is used to determine a residual value between the global feature and the reference feature, and noise interference suppression is performed on the global feature based on the residual value.

2. The method according to claim 1, characterized in that: Using a quality factor to evaluate the quality of the feature points, and generating the hypothesis generation set based on the evaluated feature points, comprising: Calculating the quality factor based on the Euclidean distance ratio and the minimum Euclidean distance; The feature points whose quality factors are greater than a preset factor threshold are retained as the hypothesis generation set.

3. The method according to claim 1, characterized in that Based on the hypothesis generation set, residual learning is used to determine a residual value between the global feature and the reference feature, and noise interference suppression is performed on the global feature based on the residual value, including: Determine the residual value between the global feature and the reference feature by analyzing the feature points in the hypothetical generation set one by one using the residual learning; By utilizing an adaptive learning rate, a weight greater than a preset weight threshold is applied to the global feature whose residual value is greater than a preset residual threshold, so as to suppress noise interference on the global feature.

4. The method according to claim 1, characterized in that Using a generative adversarial network to perform super-resolution enhancement and detail compensation on the local features to obtain the local features after detail enhancement, including: In the generator of the generative adversarial network, the local features are input into a multi-layer convolution layer, and features are gradually extracted from the local features through multi-layer convolution operations to form an intermediate feature representation; Applying a deconvolution operation to the intermediate feature representation to reconstruct super-resolution features layer by layer to form resolution features; In the discriminator of the generative adversarial network, the similarity between the resolution feature and the true resolution local feature is evaluated to obtain the discriminant loss, and based on the discriminant loss, the resolution feature is compensated for details to obtain the local feature with enhanced details.

5. The method according to claim 1, characterized in that: The global features after noise interference suppression and the local features after detail enhancement are subjected to multi-layer fusion processing, including: Using a spatial pyramid pooling method to match and align the global features after noise interference suppression and the local features after detail enhancement; The attention mechanism is used to perform multi-layer fusion processing on the matched and aligned global features and the local features.

6. An image optimization method based on artificial intelligence, characterized in that: include: Obtain the image to be processed; Utilize the deep learning model described in any one of claims 1 to 5 to optimize the image to be processed to obtain an optimized image.

7. A training device for a deep learning model for image optimization, characterized in that: include: The extraction template is configured to extract global features and local features from the sample image based on multiple convolution kernels of different sizes; A denoising module is configured to suppress noise interference on the global feature through residual learning to obtain the global feature after noise interference suppression, and to perform super-resolution enhancement and detail compensation on the local feature by using a generative adversarial network to obtain the local feature after detail enhancement; An updating module is configured to perform multi-layer fusion processing on the global features after noise interference suppression and the local features after detail enhancement to obtain an optimized sample image, compare the optimized sample image with the noise-free image, and update the network parameters of the deep learning model according to the comparison result to train the deep learning model; Wherein, the denoising module is further configured as follows: Calculating a minimum Euclidean distance between the global feature and a reference feature of the noise-free image, and obtaining a Euclidean distance ratio based on the minimum Euclidean distance; In the case where the difference between the Euclidean distance ratio and 1 is greater than a preset difference threshold, the feature points of the global feature are sorted based on the Euclidean distance ratio, and the top N feature points with the highest quality are retained as the assumed generated set; in the case where the difference between the Euclidean distance ratio and 1 is less than the preset difference threshold, the quality of the feature points is evaluated using a quality factor, and the assumed generated set is generated based on the evaluated feature points, wherein N is an integer greater than 1; Based on the hypothesis generation set, residual learning is used to determine a residual value between the global feature and the reference feature, and noise interference suppression is performed on the global feature based on the residual value.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 5.

9. A computer device, characterized in that: include: Memory and processor, The memory stores a computer program; The processor is used to execute the computer program stored in the memory, and when the computer program is run, the processor is enabled to execute the method according to any one of claims 1 to 5.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images