Monitoring picture-oriented multi-scale image super-resolution method and device
Through the multi-scale image super-resolution method, the degraded network and GAN are used to process low-resolution monitoring images, and combined with the dictionary blurred core image enhancement network, the problem of low resolution of the monitoring image is solved, the image quality and detail restoration ability are improved, and the different Requirements for magnification.
Patent Information
- Application Number
- CN202411023918.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-11-29
AI Technical Summary
Due to the low resolution of the monitoring screen, the details are lost, blurred and distorted, which affects the accuracy of target recognition. The prior art cannot effectively restore high-resolution details.
Using multi-scale image super-resolution method, high-resolution images are processed through four-layer degradation networks and adversarial generation networks (GANs), low-resolution images are formed after adversarial learning, and super-resolution processing is performed through dictionary blurred kernel image enhancement networks. , generate high-resolution images with different magnifications.
The image quality and detail restoration capabilities of the monitoring system are improved, and clearer and more accurate monitoring images are generated, which can better capture and identify monitoring objects and adapt to the monitoring image needs of different magnifications.
Smart Images

Figure CN119048349A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of image processing, and in particular to a multi-scale image super-resolution method and device for monitoring images. Background Art
[0002] In the field of image processing, image super-resolution is a technique that improves image resolution by increasing the details and clarity of an image. Traditional image super-resolution methods include interpolation techniques (such as bilinear interpolation, bicubic interpolation, etc.) and edge-preserving filtering techniques (such as super-resolution methods based on adaptive filtering). However, these traditional methods often fail to restore true high-resolution details because they do not fully utilize the complex texture and structural information in the image. In addition, these methods have limited super-resolution processing effects on complex scenes such as surveillance images. The development of deep learning technology has brought new breakthroughs in image super-resolution. Image super-resolution methods based on deep learning have achieved remarkable results in various fields, including natural images, medical images, and surveillance images. By applying these methods to super-resolution processing of surveillance images, the image quality and detail information of the surveillance system can be improved, thereby improving the accuracy and reliability of surveillance tasks such as target recognition and behavior analysis.
[0003] Due to the resolution limitations of surveillance cameras and the need for real-time transmission, surveillance images often have a low resolution. This results in the inability to obtain clear images and sufficient detail information when observing and analyzing scenes in the surveillance system. Low-resolution surveillance images may cause the following problems: Loss of detail information: The low resolution of the surveillance image makes it difficult to see the detail information clearly, such as the inability to recognize important information such as facial features and license plate numbers. Blurring and distortion: Low-resolution images may appear blurry and distorted, making the image unclear and difficult to accurately analyze the scene. Difficult target recognition: Low-resolution images may affect the accuracy of the target recognition algorithm, resulting in false alarms or missed alarms. Therefore, solving the problems of low resolution and lack of detail information in surveillance images and improving the image quality and detail restoration capabilities of the surveillance system are technical challenges that need to be solved urgently. Summary of the invention
[0004] The present invention provides a multi-scale image super-resolution method and device for monitoring images, so as to solve the problems of low monitoring image resolution and poor detail restoration capability in the prior art.
[0005] In a first aspect, the present invention provides a multi-scale image super-resolution method for monitoring images, which specifically comprises the following steps:
[0006] Step S1, obtaining a high-resolution image HR and a corresponding real monitoring image;
[0007] Step S2: Process the high-resolution image HR and the real monitoring image through a four-layer degradation network and a generative adversarial network (GAN) to form a low-resolution image LR′ after adversarial learning. 8 ;
[0008] Step S3: the real monitoring image and LR′ 8 Input the first layer of super-resolution network to form a super-resolution image Will The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified twice.
[0009] Preferably, the method further comprises the following steps:
[0010] Step S4: Input the second layer super-resolution network to form a super-resolution image Will Input the dictionary blur kernel image enhancement network to form a high-resolution image that is magnified four times;
[0011] Step S5: Input the third layer super-resolution network to form a super-resolution image Will The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified eight times.
[0012] Preferably, in step S2, the four-layer degenerate network includes a first-layer degenerate network, a second-layer degenerate network, a third-layer degenerate network and a fourth-layer degenerate network.
[0013] Among them, the first-layer degradation network uses a four-layer neural network to blur and process the image with noise; the second-layer degradation network, the third-layer degradation network and the fourth-layer degradation network have the same structure, and all use a four-layer neural network to blur, downsample and process the image with noise.
[0014] Preferably, in step S2, the high-resolution image HR and the real monitoring image are processed by a four-layer degradation network and GAN to form a low-resolution image LR 8 , specifically including the following steps:
[0015] Step S201: Input HR into the first layer of degradation network to form a low-resolution image LR 1 ;
[0016] Step S202: LR 1 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 1 ;
[0017] Step S203: LR′1 Input the second layer degradation network to form a low-resolution image LR 2 ;
[0018] Step S204: LR 2 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 2 ;
[0019] Step S205: LR′ 2 Input the third layer degradation network to form a low-resolution image LR 4 ;
[0020] Step S206: LR 4 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 4 ;
[0021] Step S207: LR′ 4 Input the fourth layer degradation network to form a low-resolution image LR 8 ;
[0022] Step S208: LR 8 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 8 .
[0023] Preferably, in step S3, step S4 and step S5, the first layer super-resolution network, the second layer super-resolution network and the third layer super-resolution network have the same structure, and all include a four-layer convolutional neural network, an upsampling layer and a four-layer convolutional neural network.
[0024] Preferably, by minimizing LR 1 and LR 2 and and LR 4 and The L1 loss function value between them is used to optimize the degradation network and the super-resolution network to ensure that the content information of the image is preserved.
[0025] In a second aspect, the present invention further provides a multi-scale image super-resolution device for monitoring images, which specifically includes the following modules:
[0026] An image acquisition module is used to acquire a high-resolution image HR and a corresponding real monitoring image;
[0027] The image degradation and adversarial learning module is used to process the high-resolution image HR and the real monitoring image through a four-layer degradation network and GAN to form a low-resolution image LR′ after adversarial learning.8 ;
[0028] A double high resolution image generation module is provided for the real monitoring image and LR′ 8 Input the first layer of super-resolution network to form a super-resolution image Will The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified twice.
[0029] Preferably, the device also includes the following modules:
[0030] Quadruple high-resolution image generation module is used to Input the second layer super-resolution network to form a super-resolution image Will Input the dictionary blur kernel image enhancement network to form a high-resolution image that is magnified four times;
[0031] Eight times higher resolution image generation module is used to Input the third layer super-resolution network to form a super-resolution image Will The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified eight times.
[0032] Preferably, in the image degradation and adversarial learning module, the four-layer degradation network includes a first-layer degradation network, a second-layer degradation network, a third-layer degradation network and a fourth-layer degradation network.
[0033] Among them, the first-layer degradation network uses a four-layer neural network to blur and process the image with noise; the second-layer degradation network, the third-layer degradation network and the fourth-layer degradation network have the same structure, and all use a four-layer neural network to blur, downsample and process the image with noise.
[0034] Preferably, the image degradation and adversarial learning module specifically includes the following submodules:
[0035] The first submodule of image processing is used to input HR into the first layer of degradation network to form a low-resolution image LR 1 ;
[0036] The second submodule of image processing is used to convert LR 1 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 1 ;
[0037] The third submodule of image processing is used to convert LR′ 1 Input the second layer degradation network to form a low-resolution image LR 2 ;
[0038] The fourth submodule of image processing is used to convert LR 2 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 2 ;
[0039] The fifth image processing submodule is used to convert LR′ 2 Input the third layer degradation network to form a low-resolution image LR 4 ;
[0040] The sixth submodule of image processing is used to convert LR 4 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 4 ;
[0041] The seventh image processing submodule is used to convert LR′ 4 Input the fourth layer degradation network to form a low-resolution image LR 8 ;
[0042] The eighth submodule of image processing is used to convert LR 8 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 8 .
[0043] Preferably, in the double high-resolution image generation module, the quadruple high-resolution image generation module and the eightfold high-resolution image generation module, the first-layer super-resolution network, the second-layer super-resolution network and the third-layer super-resolution network have the same structure, all including a four-layer convolutional neural network, an upsampling layer and a four-layer convolutional neural network.
[0044] Preferably, the device also includes the following modules:
[0045] The loss function processing module is used to minimize LR 1 and LR 2 and and LR 4 and The L1 loss function value between them is used to optimize the degradation network and the super-resolution network to ensure that the content information of the image is preserved.
[0046] In a third aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a multi-scale image super-resolution method for surveillance images as described in any one of the first aspects of the present application.
[0047] In a fourth aspect, the present invention further provides an electronic device, comprising: a memory storing a computer program; and a processor communicatively connected to the memory, for executing a multi-scale image super-resolution method for surveillance images as described in any one of the first aspects of the present application when the computer program is called.
[0048] Compared with the prior art, the present invention has the following obvious outstanding substantive features and significant advantages:
[0049] The present invention provides a multi-scale image super-resolution method and device for monitoring images, which solves the problems of low monitoring image resolution and poor detail restoration capability in the prior art. It has the following advantages:
[0050] (1) Image super-resolution for surveillance images: This paper focuses on image super-resolution methods for surveillance images. Compared with traditional image super-resolution methods, it focuses more on processing the particularities and requirements of surveillance images. The degradation of surveillance images is learned through the Generative Adversarial Network (GAN), and the dictionary fuzzy kernel image enhancement network is trained to enhance and restore the degraded image, thereby improving the image quality and detail restoration capabilities.
[0051] (2) Multi-scale super-resolution image output: The present invention constructs a multi-scale super-resolution network that can process targets of different magnifications and generate corresponding super-resolution images of different magnifications at one time. This design improves the adaptability and flexibility of the network, enabling it to adapt to surveillance images of different magnifications.
[0052] (3) Good scalability: The scalability of the multi-scale super-resolution network is also a major advantage of the present invention. The number of network branches can be further increased or decreased according to the needs of specific application scenarios to adapt to more different magnification requirements. In addition, the network's processing capabilities for surveillance images of different scales and sizes can be further improved by introducing more network layers or modules. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] The accompanying drawings, which constitute part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention. In the accompanying drawings:
[0054] Figure 1 It is a flow chart of a multi-scale image super-resolution method for monitoring images according to a preferred embodiment of the present invention.
[0055] Figure 2 It is a flowchart of a multi-scale image super-resolution method for monitoring images according to a preferred embodiment of the present invention.
[0056] Figure 3 Schematic diagram of a training dictionary fuzzy kernel image enhancement network according to a preferred embodiment of the present invention.
[0057] Figure 4 It is a schematic diagram of the structure of a multi-scale image super-resolution device for monitoring screens according to a preferred embodiment of the present invention. DETAILED DESCRIPTION
[0058] The present invention provides a multi-scale image super-resolution method and device for monitoring screens. In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0059] 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 interchanged under appropriate circumstances. 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.
[0060] Embodiment 1:
[0061] In this embodiment, a low-resolution image taken by a surveillance camera is used as input, and a high-resolution surveillance image is generated through a multi-scale super-resolution network.
[0062] like Figure 1-Figure 3 As shown, the multi-scale image super-resolution method for monitoring images described in this embodiment specifically includes the following steps:
[0063] Step S1: Acquire a high-resolution image HR and a corresponding real monitoring image.
[0064] Step S2: Process the high-resolution image HR and the real monitoring image through a four-layer degradation network and GAN to form a low-resolution image LR′ after adversarial learning. 8 .
[0065] Preferably, in step S2, the four-layer degenerate network includes a first-layer degenerate network, a second-layer degenerate network, a third-layer degenerate network and a fourth-layer degenerate network.
[0066] Among them, the first-layer degradation network uses a four-layer neural network to blur and process the image with noise; the second-layer degradation network, the third-layer degradation network and the fourth-layer degradation network have the same structure, and all use a four-layer neural network to blur, downsample and process the image with noise.
[0067] Optionally, step S2 specifically includes the following steps:
[0068] Step S201: Input the high-resolution image HR of size 256×256 into the first layer of degradation network to form a degraded low-resolution image LR of size 256×256. 1 .
[0069] At the same time, the cropped real surveillance images of the same size are used as the real images for adversarial learning.
[0070] The first layer of degradation network uses four layers of neural network to simulate the process of image blur and noise addition. At the same time, in order to ensure the stability of image content, L 1 The loss function (L 1 loss). By minimizing the difference between the generated image and the input image and its downsampled image, it ensures that the image content information is preserved.
[0071] Using LR 1 Training samples of image enhancement network with HR as dictionary blur kernel.
[0072] Preferably, the dictionary blur kernel image enhancement network uses a dictionary to store representations of 72 blur kernels. These blur kernels can represent different types and degrees of image blur effects. The network enhances the degraded image by learning how to select and apply appropriate blur kernels from the dictionary, using the pixels around each pixel for enhancement to achieve same-size image enhancement.
[0073] Step S202: LR 1 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 1 .
[0074] The degraded image is subjected to adversarial learning with the real surveillance image to improve the perceptual similarity between the generated image and the real surveillance image. The degraded image is enhanced using a dictionary fuzzy kernel image enhancement neural network to restore it to a high-definition image. The image enhancement network is saved for subsequent use.
[0075] Step S203: Convert the low-resolution image LR′ of size 256×256 1 Input the second layer degradation network to form a low-resolution image LR of size 128×128 2 .
[0076] Step S204: convert the low-resolution image LR of size 128×128 2 The real monitoring image is trained through GAN to form a low-resolution image LR′ of size 128×128 after adversarial learning. 2 .
[0077] Step S205: Convert the low-resolution image LR′ of size 128×128 2 Input the third layer degradation network to form a low-resolution image LR of size 64×64 4 .
[0078] Step S206: Convert the low-resolution image LR of size 64×64 4 The real monitoring image is trained through GAN to form a low-resolution image LR′ of size 64×64 after adversarial learning 4 .
[0079] Step S207: The low-resolution image LR′ with a size of 64×64 is 4 Input the fourth layer of degradation network to form a low-resolution image LR of size 32×32 8 .
[0080] Step S208: Convert the low-resolution image LR of size 32×32 8 The real monitoring image is trained through GAN to form a low-resolution image LR′ of size 32×32 after adversarial learning 8 Step S3: The real monitoring image and the low-resolution image LR′ with a size of 32×32 are processed. 8 Input the first layer of super-resolution network to form a super-resolution image of size 64×64 The super-resolution image of size 64×64 The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified twice.
[0081] Step S4: super-resolution image with a size of 64×64 Input the second layer super-resolution network to form a super-resolution image of size 128×128 The super-resolution image of size 128×128 The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified four times.
[0082] Step S5: super-resolution image with a size of 128×128 Input the third layer super-resolution network to form a super-resolution image of size 256×256 The super-resolution image of size 256×256 The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified eight times.
[0083] Among them, in step S3, step S4 and step S5, the first layer super-resolution network, the second layer super-resolution network and the third layer super-resolution network have the same structure, all including a four-layer convolutional neural network, an upsampling layer and a four-layer convolutional neural network.
[0084] Optionally, by minimizing LR 1 and LR 2 and and LR 4 and The L1 loss function value between them is used to optimize the degradation network and the super-resolution network to ensure that the content information of the image is preserved.
[0085] In the above step S2 of this embodiment, a multi-layer degradation network is used to generate training samples, the resolution of the image is gradually degraded, and adversarial learning is used to generate higher resolution monitoring images. Each degradation network uses a four-layer convolutional neural network to simulate the blurring, downsampling and noise adding process of the image. In order to maintain the stability of the image content, the L1 loss function value is introduced to retain the content information of the image by minimizing the difference between the generated image and the high-resolution downsampled image. Through layer-by-layer degradation and adversarial learning, this embodiment can generate monitoring images with higher perceptual similarity, thereby improving image quality and detail restoration capabilities. This method can be applied to subsequent super-resolution network training to improve the clarity and accuracy of the image.
[0086] In the above steps S3, S4 and S5 of this embodiment, the low-resolution image is enhanced layer by layer, and the clarity and quality of the image are further improved through the dictionary fuzzy kernel image enhancement network. Through the layer-by-layer super-resolution and dictionary fuzzy kernel image enhancement, a higher-resolution monitoring image can be generated, and the content stability of the image can be maintained. This method can improve the image quality and detail restoration effect of the monitoring system, make the monitoring picture clearer and more delicate, and can more accurately capture and identify the monitored object.
[0087] Embodiment 2:
[0088] Based on the same design concept, this embodiment also provides a multi-scale image super-resolution device for monitoring images. Figure 4 As shown, the device specifically includes the following modules:
[0089] The image acquisition module is used to acquire the high-resolution image HR and the corresponding real monitoring image.
[0090] The image degradation and adversarial learning module is used to process the high-resolution image HR and the real monitoring image through a four-layer degradation network and GAN to form a low-resolution image LR′ after adversarial learning. 8 .
[0091] The four-layer degenerate network includes a first-layer degenerate network, a second-layer degenerate network, a third-layer degenerate network and a fourth-layer degenerate network.
[0092] Among them, the first-layer degradation network uses a four-layer neural network to blur and process the image with noise; the second-layer degradation network, the third-layer degradation network and the fourth-layer degradation network have the same structure, and all use a four-layer neural network to blur, downsample and process the image with noise.
[0093] Preferably, the image degradation and adversarial learning module specifically includes the following submodules:
[0094] The first submodule of image processing is used to input HR into the first layer of degradation network to form a low-resolution image LR 1 .
[0095] The second submodule of image processing is used to convert LR 1 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 1 .
[0096] The third submodule of image processing is used to convert LR′ 1 Input the second layer degradation network to form a low-resolution image LR 2 .
[0097] The fourth submodule of image processing is used to convert LR 2 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 2 .
[0098] The fifth image processing submodule is used to convert LR′ 2 Input the third layer degradation network to form a low-resolution image LR 4 .
[0099] The sixth submodule of image processing is used to convert LR 4 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 4 .
[0100] The seventh image processing submodule is used to convert LR′ 4 Input the fourth layer degradation network to form a low-resolution image LR 8 .
[0101] The eighth submodule of image processing is used to convert LR 8 The real monitoring image is trained through GAN to form a low-resolution image LR′ after adversarial learning 8 .
[0102] A double high resolution image generation module is provided for the real monitoring image and LR′ 8 Input the first layer of super-resolution network to form a super-resolution image Will The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified twice.
[0103] Quadruple high-resolution image generation module is used to Input the second layer super-resolution network to form a super-resolution image Will The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified four times.
[0104] Eight times higher resolution image generation module is used to Input the third layer super-resolution network to form a super-resolution image Will The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified eight times.
[0105] In the double-high-resolution image generation module, the quadruple-high-resolution image generation module and the eight-fold high-resolution image generation module, the first-layer super-resolution network, the second-layer super-resolution network and the third-layer super-resolution network have the same structure, all including a four-layer convolutional neural network, an upsampling layer and a four-layer convolutional neural network.
[0106] Optionally, the device also includes the following modules:
[0107] The loss function processing module is used to minimize LR 1 and LR 2 and and LR 4 and The L1 loss function value between them is used to optimize the degradation network and the super-resolution network to ensure that the content information of the image is preserved.
[0108] In summary, the present application provides a multi-scale image super-resolution method and device for monitoring images, aiming to improve the quality and detail restoration capability of low-resolution monitoring images. The method includes two stages: degradation network and super-resolution network. In the degradation network stage, the present application uses a multi-layer degradation network to gradually reduce the resolution of the image, and generates a real monitoring image through adversarial learning to maintain the stability of the image content. Each degradation network uses a convolutional neural network structure to simulate the blurring, downsampling and noise adding process of the image, and retains the content information of the image through the L1 loss function value. In the super-resolution network stage, the present application uses a multi-layer super-resolution network to gradually enhance the resolution of the image, and further improves the quality of the image through a dictionary fuzzy kernel image enhancement network. Each super-resolution network also uses a convolutional neural network structure and retains the content information of the image by minimizing the difference with the low-resolution image. Through the layer-by-layer degradation and super-resolution process, the present application can generate a higher resolution monitoring image, and improve the clarity of the image through a dictionary fuzzy kernel image enhancement network. This method can be applied to a monitoring system to improve the accuracy and clarity of the image to better capture and analyze the target object in the monitoring screen. In general, this method combines the degradation network and super-resolution network, as well as the techniques of adversarial learning and dictionary blur kernel image enhancement, which can effectively improve the quality and detail restoration ability of surveillance images, which is of great significance for improving the performance and reliability of surveillance systems.
[0109] The specific embodiments of the present invention are described in detail above, but they are only examples, and the present invention is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions made to the present invention are also within the scope of the present invention. Therefore, the equalization changes and modifications made without departing from the spirit and scope of the present invention should be included in the scope of the present invention.
Claims
1. A multi-scale image super-resolution method for monitoring images, characterized in that: The specific steps include: Step S1, obtaining a high-resolution image HR and a corresponding real monitoring image; Step S2, processing the high-resolution image HR and the real monitoring image through a four-layer degradation network and an adversarial generation network to form a low-resolution image LR′8 after adversarial learning; Step S3: Input the real monitoring image and LR′8 into the first layer super-resolution network to form a super-resolution image Will The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified twice.
2. The method for multi-scale image super-resolution for monitoring images according to claim 1, characterized in that: The following steps are also included: Step S4: Input the second layer super-resolution network to form a super-resolution image Will Input the dictionary blur kernel image enhancement network to form a high-resolution image that is magnified four times; Step S5: Input the third layer super-resolution network to form a super-resolution image Will The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified eight times.
3. The method for multi-scale image super-resolution for monitoring images according to claim 1, characterized in that: In step S3, step S4 and step S5, the first layer super-resolution network, the second layer super-resolution network and the third layer super-resolution network have the same structure, and all include a four-layer convolutional neural network, an upsampling layer and a four-layer convolutional neural network.
4. The method for multi-scale image super-resolution for monitoring images according to claim 1, characterized in that: In step S2, the four-layer degenerate network includes a first layer degenerate network, a second layer degenerate network, a third layer degenerate network and a fourth layer degenerate network. Among them, the first-layer degradation network uses a four-layer neural network to blur and process the image with noise; the second-layer degradation network, the third-layer degradation network and the fourth-layer degradation network have the same structure, and all use a four-layer neural network to blur, downsample and process the image with noise.
5. The method for multi-scale image super-resolution for monitoring images according to claim 2, characterized in that: In step S2, the high-resolution image HR and the real monitoring image are processed by a four-layer degradation network and a GAN to form a low-resolution image LR8, which specifically includes the following steps: Step S201, inputting HR into the first layer degradation network to form a low-resolution image LR1; Step S202, training LR1 and the real monitoring image through GAN to form a low-resolution image LR′1 after adversarial learning; Step S203, inputting LR′1 into the second layer degradation network to form a low-resolution image LR2; Step S204, training LR2 and the real monitoring image through GAN to form a low-resolution image LR′2 after adversarial learning; Step S205, input LR′2 into the third layer degradation network to form a low-resolution image LR4; Step S206, training LR4 and the real monitoring image through GAN to form a low-resolution image LR′4 after adversarial learning; Step S207, input LR′4 into the fourth layer degradation network to form a low-resolution image LR8; Step S208: LR8 and the real monitoring image are trained through GAN to form a low-resolution image LR′8 after adversarial learning.
6. The method for multi-scale image super-resolution for monitoring images according to claim 5, characterized in that: By minimizing LR1 and LR2 and As well as LR4 and The L1 loss function value between them is used to optimize the degradation network and the super-resolution network to ensure that the content information of the image is preserved.
7. A device for multi-scale image super-resolution for monitoring images, characterized in that: The modules include: An image acquisition module is used to acquire a high-resolution image HR and a corresponding real monitoring image; An image degradation and adversarial learning module, used for processing the high-resolution image HR and the real monitoring image through a four-layer degradation network and GAN to form a low-resolution image LR′8 after adversarial learning; A double high-resolution image generation module is provided to input the real monitoring image and LR′8 into the first layer super-resolution network to form a super-resolution image Will The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified twice.
8. The device for multi-scale image super-resolution for monitoring images according to claim 7, characterized in that: Also includes the following modules: Quadruple high-resolution image generation module is used to Input the second layer super-resolution network to form a super-resolution image Will Input the dictionary blur kernel image enhancement network to form a high-resolution image that is magnified four times; Eight times higher resolution image generation module is used to Input the third layer super-resolution network to form a super-resolution image Will The dictionary blur kernel image enhancement network is input to form a high-resolution image that is magnified eight times.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for multi-scale image super-resolution for monitoring images as described in any one of claims 1 to 6 is implemented.
10. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, a multi-scale image super-resolution method for monitoring screens as described in any one of claims 1 to 6 is implemented.
