Image enhancement method, system, device and medium

Abstract

The invention relates to the field of image processing, and particularly discloses an image enhancement method and system, computer equipment and a medium, and the method comprises the steps: obtaining an original image with first quality; obtaining a scene description vector based on the original image; obtaining a reference image of second quality based on a historical image and the scene description vector, wherein the second quality is higher than the first quality; and enhancing the original image based on the original image, the scene description vector and the reference image to obtain an enhanced image. According to the scheme provided by the invention, the historical high-quality picture and the guide vector related to the scene are combined, so that when the image of a specific scene is restored, the information related to the scene can be more concerned, for example, the number of lanes, the positions of signal lamps, lane lines and the like can be more concerned in a traffic scene, and the model can focus on restoring the image.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing, in particular to an image enhancement method, system, device and medium. BACKGROUND

[0002] With the continuous promotion of vehicle-road-cloud integration, the deployment scale of roadside cameras as key perception devices is expanding, and they need to run uninterruptedly for 7x24 hours to continuously collect road scene video information. These video information contains rich traffic data, which will produce significant social and economic value if effectively utilized. However, roadside cameras face many challenges during long-term operation, among which the influence of adverse weather is particularly prominent. Adverse weather (such as rain, snow, haze, etc.) will interfere with the normal work of the camera in many ways, resulting in a significant decline in the quality of the collected images. For example, water droplets, ice crystals or particulate matter in the air can scatter light, causing the picture to be blurred and the contrast to be reduced, which is more obvious at night or in low light conditions.

[0003] The decline in image quality directly affects subsequent data processing and analysis, and is easy to cause target misidentification problems, such as misjudging rain streaks as lane lines, snow cover leading to traffic sign recognition errors, etc., thereby adversely affecting the stable operation and function realization of the vehicle-road-cloud integrated system. Therefore, how to enhance the low-quality images collected by roadside cameras is a problem to be solved. SUMMARY

[0004] In view of this, in order to overcome at least one aspect of the above problems, the embodiments of the present application propose an image enhancement method, comprising the following steps: obtaining an original image of a first quality; obtaining a scene description vector based on the original image; obtaining a reference image of a second quality based on historical images and the scene description vector, the second quality being higher than the first quality; enhancing the original image based on the original image, the scene description vector and the reference image to obtain an enhanced image.

[0005] In some embodiments, obtaining a scene description vector based on the original image further comprises: preprocessing the original image based on a preprocessing module; encoding the preprocessed original image based on a first encoder module to obtain encoded features; processing the encoded features based on an adapter module to obtain the scene description vector.

[0006] In some embodiments, further comprising: constructing a training set using a plurality of first-quality images and a plurality of second-quality images; performing first-stage training on the preprocessing module and the adapter module based on the training set.

[0007] In some embodiments, the enhancing the original image based on the original image, the scene description vector and the reference image to obtain an enhanced image further comprises: encoding the original image and the reference image based on a second encoder to obtain a first encoded feature and a second encoded feature, respectively; calculating the first encoded feature, the second encoded feature and the scene description vector based on a neural network to obtain a target feature; obtaining the enhanced image based on the target feature.

[0008] In some embodiments, the method further comprises: performing second-stage training on the neural network based on the training set and adjusting the preprocessing module and the adapter module after the first-stage training.

[0009] In some embodiments, the obtaining the second-quality reference image based on the historical image and the scene description vector further comprises: constructing a database based on a plurality of second-quality historical images of the target region and generating a corresponding vector for each image; obtaining the reference image based on the similarity between the scene description vector and the vector of each image in the database.

[0010] In some embodiments, the constructing the database based on a plurality of second-quality images further comprises: obtaining a plurality of historical images of the target region; selecting a plurality of second-quality images based on a quality evaluation index to construct the database; updating the database every predetermined time period.

[0011] Based on the same inventive concept, according to another aspect of the present application, embodiments of the present application also provide an image enhancement system, comprising: an acquisition module configured to acquire an original image of a first quality; an encoding module configured to obtain a scene description vector based on the original image; a search module configured to obtain a reference image of a second quality based on a historical image and the scene description vector, the second quality being higher than the first quality; an enhancement module configured to enhance the original image based on the original image, the scene description vector and the reference image to obtain an enhanced image.

[0012] Based on the same inventive concept, according to another aspect of the present application, embodiments of the present application further provide a computer device, comprising: at least one processor; and a memory storing a computer program capable of running on the processor, wherein the processor executes the program to perform the steps of any one of the image enhancement methods described above.

[0013] Based on the same inventive concept, according to another aspect of the present application, embodiments of the present application further provide a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to perform the steps of any one of the image enhancement methods described above.

[0014] The present application has one of the following beneficial technical effects: the scheme provided by the present application combines historical high-quality pictures and scene-related guide vectors, so that when restoring images of a specific scene, more attention can be paid to scene-related information, such as the number of lanes, the position of traffic lights, lane lines, etc. in a traffic scene, so that the model can restore the image with emphasis. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other embodiments can also be obtained from these drawings without creative labor.

[0016] Figure 1 a flowchart of the image enhancement method provided by the embodiments of the present application; Figure 2 a flowchart of the image enhancement method provided by the embodiments of the present application; Figure 3 a training schematic diagram of the Scenario Encoder module provided by the embodiments of the present application; Figure 4 a structural schematic diagram of the image enhancement system provided by the embodiments of the present application; Figure 5 a structural schematic diagram of the computer device provided by the embodiments of the present application; Figure 6 a structural schematic diagram of the computer readable storage medium provided by the embodiments of the present application. DETAILED DESCRIPTION

[0017] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] It should be noted that all the expressions of "first" and "second" in the embodiments of the present application are used to distinguish two same name different entities or different parameters, and "first" and "second" are only used for the convenience of description, and should not be understood as the limitation of the embodiments of the present application. The subsequent embodiments will not be described one by one.

[0019] According to one aspect of the present application, the embodiments of the present application provide an image enhancement method, as shown in Figure 1 which can include the steps of: S1, obtaining a first quality original image; S2, obtaining a scene description vector based on the original image; S3, obtaining a second quality reference image based on the scene description vector and historical images, the second quality being higher than the first quality; S4, enhancing the original image based on the original image, the scene description vector and the reference image to obtain an enhanced image.

[0020] Specifically, the first-quality original image can be an image in a specific scene, such as a low-quality traffic scene image captured by a roadside camera (for example, an image quality decrease caused by bad weather, water droplets, ice crystals or particulate matter scattering light, insufficient light, etc.), which contains basic visual information of the specific scene, but some detailed information (such as lane lines, traffic signs) may be distorted or missing. After obtaining the original image, a model can be used to generate a scene description vector based on the original image. The scene description vector can include specific scene descriptions of the original image, and can also include key detailed information, such as “license plate area”, “signal light color”, “lane line identification” and the like in the traffic scene image. After generating the scene description vector, a second-quality reference image similar to the original image in scene can be matched from a high-quality scene database, wherein the high-quality scene database is constructed based on the characteristics of the fixed shooting position of the camera, and stores high-quality images captured by the same camera at different times, in different weather, and in the same scene, so that the reference image can include the key details missing in the original image, providing a real sample basis for subsequent enhancement. Finally, the original image, the scene description vector and the reference image are fused to complete image enhancement by means of a generation model. The final output enhanced image not only repairs the quality defects of the original image, but also retains the key detailed information in the specific scene. In this way, the scheme proposed in the present application combines historical high-quality pictures and scene-related guide vectors, so that when restoring images in a specific scene, more attention can be paid to scene-related information, such as the number of lanes, the position of the signal light, the lane line and the like in a traffic scene, so that the model can focus on restoring the image.

[0021] In some embodiments, obtaining the scene description vector based on the original image further includes: preprocessing the original image based on a preprocessing module; encoding the preprocessed original image based on a first encoder module to obtain encoded features; processing the encoded features based on an adapter module to obtain the scene description vector.

[0022] Specifically, as shown in Figure 2 the scene description vector can be obtained by using a Scenario Encoder to calculate the original image. The Scenario Encoder can include three sub-modules, one is a fixed CLIP Image Encoder, and the other two are sub-modules participating in training (Pre-process Module and Scenario Adapter).

[0023] The pre-processing module (Pre-process Module) can perform targeted pre-processing operations on the original image of the first quality (low-quality image captured by the roadside camera), thereby eliminating irrelevant interference factors in the scene characteristics of the original image, providing more regular input for subsequent encoding, for example, correcting the tilt distortion of the image caused by the installation angle of the camera, removing noise points caused by device noise, performing normalization processing, or standardizing the image size and resolution (such as uniform scaling to the input size required by the CLIP Image Encoder).

[0024] The first encoder module (CLIP Image Encoder) can encode the pre-processed image and convert visual information into semantically associated encoded features. As a pre-trained general visual encoder, the CLIP Image Encoder can capture rich visual features in the image and project the image into a text feature space, so that the output encoded features not only contain visual information of the image, but also include semantic information.

[0025] The adapter module (Scenario Adapter) is used to perform targeted processing on the encoded features output by the first encoder module to generate a description vector adapted to a specific scene, for example, a roadside camera scene. During training, the image data of a specific scene is used for optimization to learn the unique features of the specific scene, thereby adjusting and enhancing the encoded features, for example, the key information weight of the specific scene can be strengthened, or the general features irrelevant to the specific scene can be weakened, so that the output scene description vector more accurately reflects the attributes of the specific scene.

[0026] The scenario description vector generated by the Scenario Encoder finally has the semantic richness of CLIP encoding and can better adapt to a specific scene through the scene adaptation of the adapter module.

[0027] In some embodiments, further comprising: Constructing a training set using a plurality of first-quality images and a plurality of second-quality images; Based on the training set, the pre-processing module and the adapter module are trained in the first stage.

[0028] Specifically, the second quality image, i.e., the high-quality image (HR), can be used as the target true value for training. The second quality image can be a clear image captured by the camera when the device is normal and the environment is good. The first quality image, i.e., the low-quality image (LR), can be generated based on the high-quality image, generated by simulating common camera degradation problems, such as downsampling the HR image (to simulate the detail blur caused by insufficient camera resolution), adding noise (to simulate the noise points generated by low light at night or device aging), and superimposing fog / rain effects (to simulate the picture quality degradation caused by bad weather), thereby realizing accurate pairing of LR-HR in the same scene and the same perspective, avoiding training target deviation caused by scene differences. At the same time, real low-quality images captured by the camera (such as rainwater fog blurred images and low-light images at night) can be supplemented and paired with corresponding high-quality images to further improve the adaptability of the training set to real scenes.

[0029] As shown in Figure 3 , the training can adopt a two-stage strategy. The first stage trains the pre-processing module (Pre-process Module) and the adapter module (Scenario Adapter), and the second stage trains the (Denoising U-Net) in the diffusion model (Diffusion Model). The specific training process of the first stage can be to optimize the parameters of the pre-processing module and the adapter module by combining the loss function and the gradient backpropagation strategy (the first encoder module parameter is fixed and does not participate in training). The goal of training is to make the image obtained after inputting the scenario description vector (Scenario Embedding) output by the adapter module into the diffusion model (Diffusion Model) close to the high-quality image as the true value. Therefore, the reconstruction loss can be used as the loss function, and the loss function expression can be: Loss[Diffusion(Adapter(CLIP(Pre(LR)))),HR] wherein, Pre(LR) represents processing the low-quality image using the pre-processing module; CLIP(Pre(LR)) represents encoding the pre-processed image using the fixed-parameter CLIP Image Encoder to obtain the encoding feature; Adapter(CLIP(Pre(LR))) represents processing the encoding feature using the adapter module to obtain the scenario description vector; Diffusion(...) represents restoring the scenario description vector using the diffusion model to obtain the generated image; Loss[generated image, HR] represents calculating the difference (such as L1, L2 loss, etc.) between the generated image and the paired high-quality image. The smaller the loss value is, the more the features output by the pre-processing and adapter modules can guide the model to generate high-quality images close to the true value.

[0030] In some embodiments, enhancing the original image based on the original image, the scene description vector, and the reference image to obtain an enhanced image further includes: The original image and the reference image are encoded based on the second encoder to obtain the first encoded feature and the second encoded feature, respectively. The target features are calculated based on the first encoded features, the second encoded features, and the scene description vector using a neural network. The enhanced image is obtained based on the target features.

[0031] Specifically, such as Figure 2 As shown, in the image enhancement process, firstly, for the low-quality original image (LR), scene parsing is performed using a Scenario Encoder to obtain a scene description vector (Scenario Embedding) containing specific scene descriptions and key details. Next, based on the scene description vector, matching is performed in a high-quality image database to select reference images highly similar to the current scene. Then, the second encoder (VQGAN Encoder) of the Diffusion Model encodes both the low-quality original image and the high-quality reference image. Specifically, the original image is encoded to generate first encoded features, preserving the basic structure of the scene; the reference image is encoded to generate second encoded features, obtaining clearer details and common scene features. Finally, the first encoded features, second encoded features, and scene description vector are input into the Denoising U-Net neural network of the Diffusion Model for image enhancement. The neural network uses these three inputs as conditions to process random noise Z... t The reverse diffusion iteration is performed. In each iteration, the neural network gradually suppresses random noise Z based on the second encoded features of the reference image. t The noise and defects are reduced, while key details are enhanced under the guidance of the scene description vector, such as "license plate area," "traffic light color," and "lane markings" in traffic scenes. After multiple iterations, the random noise Z... t The process gradually transforms the target features Z into a composite of the original scene structure, reference image details, and scene guidance information. Finally, the VQGAN Decoder is used to decode the target features Z and generate an enhanced high-quality image (HR). The resulting enhanced image not only corrects the quality defects of the original image but also ensures the clarity and accuracy of key details while maintaining scene consistency.

[0032] In some embodiments, it also includes: The neural network is trained in the second stage based on the training set, and the preprocessing module and the adapter module are adjusted after the first stage of training.

[0033] Specifically, such as Figure 2 As shown, in the second stage of model training, the neural network (DenoisingU-Net) is subjected to targeted reinforcement training based on the training set, and the preprocessing module and adapter module, which were initially optimized in the first stage, are finely adjusted to further improve the robustness of the model in complex scenarios. The specific training process in the second stage can combine the loss function and gradient backpropagation strategy to optimize the neural network (the second encoder and decoder are fixed and do not participate in training). The goal of the training is to make the generated image output by the neural network as close as the high-quality image used as the ground truth.

[0034] In some embodiments, obtaining a second-quality reference image based on historical images and the scene description vector further includes: A database is constructed based on multiple second-quality historical images of the target region, and a corresponding vector is generated for each image. The reference image is obtained based on the similarity between the scene description vector and the vector of each image in the database.

[0035] Specifically, a dedicated database can be built based on a large number of second-quality images captured by cameras in the target area. The target area can refer to a specific traffic area (such as an urban intersection, highway toll station, tunnel entrance / exit, etc.) that is fixedly monitored by roadside cameras, and its spatial range, shooting angle, and covered traffic elements (such as the number of lanes, traffic light positions, and fixed buildings) remain stable. The second-quality images in the database can cover the full scene dimensions within the fixed shooting range of the cameras, including clear images at different times, in different weather conditions, and under different traffic conditions, and include complete key traffic elements (such as the number of lanes, traffic light positions, lane lines, etc.). The construction of the database relies on the fixed position of the roadside cameras to ensure that the stored images share the same physical scene as the original images to be enhanced (such as the same intersection or the same road section), providing a spatial dimension matching basis for subsequent image enhancement. To facilitate subsequent filtering, a scene description vector (Scenario Embedding) is generated for each image in the database using ScenarioEncoder.

[0036] After obtaining the scene description vector of the original image to be enhanced, the most suitable reference image is selected by calculating the similarity of the vector with all the vectors of the images in the database. For example, cosine similarity, Euclidean distance, etc. can be used to quantify the matching degree of the scene description vector and the database image vector. The smaller the angle or the closer the distance in the vector space, the more similar the scene features (such as time, weather, key elements) of the two. Compared with randomly selecting a reference image, matching based on vector similarity can ensure that the reference image is highly related to the original image in terms of scene logic and key elements, so that the enhanced image not only conforms to the physical scene rules (such as fixed position of signal lights), but also accurately repairs defects. Moreover, the efficiency of vector representation and similarity calculation enables the system to quickly locate the optimal reference image in a large amount of historical data, meeting the real-time needs of specific scenarios.

[0037] In some embodiments, constructing the database based on the plurality of second-quality images further comprises: obtaining a plurality of historical images of the target area; selecting the plurality of second-quality images based on the quality evaluation index to construct the database; updating the database every predetermined time period.

[0038] Specifically, when obtaining historical images, the fixed shooting characteristics of roadside cameras can be relied on to collect image data under different times and different environmental conditions in a specific area. The goal is to cover all-dimensional changes of the scene, such as covering all 24-hour time nodes to ensure that the database contains scene images under different light intensities and different traffic flows. Images under different weather conditions such as sunny, rainy, foggy, snowy, and cloudy days, as well as scene data in different seasons (such as strong light in summer and snow cover in winter), are collected to cover common image interference scenarios in traffic scenes. Regular traffic scenes (vehicles driving normally, pedestrians crossing the road) and special event scenes (such as temporary vehicle parking, signal light failure, and minor scratches) are collected to ensure that the database can provide references for low-quality images in various traffic scenarios. In this way, through targeted collection, historical images can fully reflect the visual feature changes of the target area, avoiding the problem that reference images cannot match original images under special working conditions due to single data.

[0039] The image quality can be quantitatively scored in combination with a no-reference quality evaluation index (such as MANIQA and MUSIQ), and N frames of clear images with higher scores are screened out to construct an initial high-quality image database. In order to ensure the timeliness and quality stability of the high-quality images, the database can adopt a dynamic updating mechanism, for example, the database is updated every day or every week, the quality of the existing images is re-evaluated through the MANIQA and MUSIQ indexes, the images with low scores are removed, and higher-quality images collected at the same period are supplemented, while the total capacity (N frames) of the database is kept unchanged, thereby providing a continuous and reliable high-quality true value source for image enhancement.

[0040] The scheme provided by the application combines historical high-quality pictures and scene-related guide vectors, so that when the image restoration of a specific scene is performed, more attention can be paid to scene-related information, for example, more attention can be paid to the number of lanes, the position of a signal lamp and lane lines in a traffic scene, so that the model can restore the image with emphasis.

[0041] Based on the same inventive concept, according to another aspect of the application, embodiments of the application further provide an image enhancement system 400, as shown in the figure, comprising: Figure 4 an acquisition module 401 configured to acquire an original image of a first quality; an encoding module 402 configured to obtain a scene description vector based on the original image; a search module 403 configured to obtain a reference image of a second quality based on a historical image and the scene description vector, the second quality being higher than the first quality; an enhancement module 404 configured to enhance the original image based on the original image, the scene description vector and the reference image to obtain an enhanced image. In some embodiments, obtaining the scene description vector based on the original image further comprises:

[0042] preprocessing the original image based on a preprocessing module; encoding the preprocessed original image based on a first encoder module to obtain encoded features; processing the encoded features based on an adapter module to obtain the scene description vector. In some embodiments, further comprising:

[0043] constructing a training set using a plurality of images of a first quality and a plurality of images of a second quality; performing first-stage training on the preprocessing module and the adapter module based on the training set.

[0044] ​In some embodiments, the enhancing the original image based on the original image, the scene description vector and the reference image to obtain an enhanced image further comprises: encoding the original image and the reference image based on a second encoder to obtain a first encoding feature and a second encoding feature respectively; calculating the first encoding feature, the second encoding feature and the scene description vector based on a neural network to obtain a target feature; obtaining the enhanced image based on the target feature.

[0045] In some embodiments, the method further comprises: performing a second stage training on the neural network based on the training set and adjusting the pre-processing module and the adapter module after the first stage training.

[0046] In some embodiments, the obtaining the reference image of the second quality based on the historical image and the scene description vector further comprises: constructing a database based on a plurality of historical images of the target region of the second quality and generating a corresponding vector for each image; obtaining the reference image based on a similarity between the scene description vector and the vector of each image in the database.

[0047] In some embodiments, the constructing the database based on a plurality of historical images of the second quality further comprises: obtaining a plurality of historical images of the target region; selecting a plurality of historical images of the second quality based on a quality evaluation index to construct the database; updating the database every predetermined time period.

[0048] The scheme provided by the application combines historical high-quality pictures and scene-related guide vectors, so that when the image of a specific scene is restored, more attention can be paid to scene-related information, such as the number of lanes, the position of traffic lights, lane lines, etc. in a traffic scene, so that the model can restore the image with emphasis.

[0049] Based on the same inventive concept, according to another aspect of the application, as Figure 5 shown, the embodiments of the application also provide a computer device 501, comprising: at least one processor 520; and a memory 510, the memory 510 storing a computer program 511 executable on the processor, and the processor 520 executes the steps of any of the image enhancement methods as above when executing the program.

[0050] Based on the same inventive concept, according to another aspect of the application, asFigure 6 As shown, the embodiments of the present application also provide a computer readable storage medium 601, which stores a computer program 610, and the computer program 610 is executed by a processor to perform the steps of any of the above image enhancement methods.

[0051] Finally, it should be noted that those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by a computer program to instruct related hardware, and the program can be stored in a computer readable storage medium, and the program can include the processes of the above-mentioned embodiments when executed.

[0052] In addition, it should be understood that the computer readable storage medium (for example, memory) herein can be a volatile memory or a non-volatile memory, or can include both volatile memory and non-volatile memory.

[0053] Those skilled in the art will also appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the disclosure herein can be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the particular application and design constraints imposed on the overall system. Skilled persons can implement the functions described above in various ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the embodiments disclosed herein.

[0054] The above is the exemplary embodiment disclosed by the present application, but it should be noted that various changes and modifications can be made without departing from the scope of the embodiments disclosed by the present application defined by the claims. The functions, steps and / or acts of the method claims described in the embodiments disclosed herein need not be performed in any particular order. In addition, although the elements of the embodiments disclosed by the present application can be described or claimed in individual form, they can also be understood as plural unless explicitly limited to singular.

[0055] It should be understood that, as used herein, the singular forms "a", "an" and "the" are intended to include plural forms unless the context clearly dictates otherwise. It should also be understood that "and / or" as used herein is intended to include any and all possible combinations of one or more of the associated listed items.

[0056] The above-mentioned embodiment number of the embodiments disclosed by the present application is only for description, not representing the advantages and disadvantages of the embodiments.

[0057] Those skilled in the art can understand that all or part of the steps of the above-mentioned embodiments can be completed by hardware, or by program instructing relevant hardware to complete, and the program can be stored in a computer readable storage medium. The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0058] Those skilled in the art shall understand that the above discussion of any embodiment is only exemplary, and is not intended to imply that the scope (including claims) of the embodiments of the present application is limited to these examples; the technical features in the above embodiments or different embodiments can also be combined, and there are many other changes of different aspects of the embodiments of the present application as above. In order to be brief, they are not provided in details. Therefore, any omission, modification, equivalent replacement, improvement, etc. made in the spirit and principle of the embodiments of the present application shall be included in the protection scope of the embodiments of the present application.

Claims

1. An image enhancement method characterized by, The method comprises the following steps: obtaining a first-quality original image; obtaining a scene description vector based on the original image; obtaining a second-quality reference image based on a historical image and the scene description vector, the second quality being higher than the first quality; enhancing the original image based on the original image, the scene description vector and the reference image to obtain an enhanced image.

2. The method of claim 1, wherein, The step of obtaining a scene description vector based on the original image further comprises: preprocessing the original image based on a preprocessing module; encoding the preprocessed original image to obtain encoded features based on a first encoder module; processing the encoded features to obtain the scene description vector based on an adapter module.

3. The method of claim 2, wherein, The method further comprises: constructing a training set using a plurality of first-quality images and a plurality of second-quality images; performing first-stage training on the preprocessing module and the adapter module based on the training set.

4. The method of claim 3, wherein, The step of enhancing the original image based on the original image, the scene description vector and the reference image to obtain an enhanced image further comprises: encoding the original image and the reference image to obtain first encoded features and second encoded features respectively based on a second encoder; calculating the first encoded features, the second encoded features and the scene description vector to obtain target features based on a neural network; obtaining the enhanced image based on the target features.

5. The method of claim 4, wherein, The method further comprises: performing second-stage training on the neural network and adjusting the preprocessing module and the adapter module after the first-stage training based on the training set.

6. The method of claim 1, wherein, The step of obtaining a second-quality reference image based on a historical image and the scene description vector further comprises: constructing a database based on a plurality of second-quality historical images of a target region and generating a corresponding vector for each image; obtaining the reference image based on the similarity between the scene description vector and the vector of each image in the database.

7. The method of claim 6, wherein, The step of constructing a database based on a plurality of second-quality images further comprises: obtaining a plurality of historical images of a target region; selecting a plurality of second-quality images based on a quality evaluation index to construct the database; updating the database every predetermined time period.

8. An image enhancement system characterized by, The method comprises: an obtaining module configured to obtain a first-quality original image; an encoding module configured to obtain a scene description vector based on the original image; a searching module configured to obtain a second-quality reference image based on a historical image and the scene description vector, the second quality being higher than the first quality; an enhancing module configured to enhance the original image based on the original image, the scene description vector and the reference image to obtain an enhanced image.

9. A computer device, comprising: at least one processor; and a memory storing a computer program executable on the processor, wherein the processor executes the program to perform the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to perform the steps of the method according to any one of claims 1-7.

Images