A method and device for fog image data set based on depth estimation
By combining depth estimation and optical fogging models, a high-quality foggy image dataset was generated, which solved the problems of scarcity and low quality of foggy datasets in existing technologies, achieved simulation of real fog effects and data support with detailed labels, and improved image processing and detection effects in foggy environments.
Patent Information
- Application Number
- CN202510948609.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-10
- Publication Date
- 2025-10-17
- Estimated Expiration
- 2045-07-10
AI Technical Summary
The number of existing foggy data sets is limited and the quality is uneven. The existing synthetic fog technology lacks modeling of natural optical laws, and the generated effect is unrealistic. The image data generated by virtual scene simulation fog technology lacks authenticity and versatility, and the lack of labels affects the application effect.
A combination of depth estimation model and optical fogging model is adopted to obtain the depth information of the target image through depth estimation for semantic deepening, and the optical fogging model is used to generate fog layers of different densities and levels according to the scene structure. The foggy visual task matching algorithm is determined in combination with the target task to generate a high-quality dataset.
The accuracy and efficiency of generating synthetic fog datasets are improved, and image datasets that conform to real-world fog effects are generated. This supports the application of image processing and computer vision models in foggy environments, and provides detailed labeled target detection and fusion task data support.
Smart Images
Figure CN120451727B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of fog data set synthesis, in particular to a fog image data set synthesis method and device based on depth estimation. BACKGROUND
[0002] Artificial intelligence is increasingly integrated into our lives, and industries such as autonomous driving and monitoring detection need to have the ability to adapt to various weather conditions. Fog, as a common natural phenomenon, is particularly common in the early morning of autumn and winter and in the mountains of the south, but existing fog data sets are limited in quantity and quality due to lighting and environmental restrictions. The inconsistency of fog density and lighting conditions leads to poor usability. After the introduction of multi-modal fusion technology, infrared and visible light images are of great concern due to their complementary nature. Due to the scarcity of fog data sets and the difficulty of real-time acquisition, synthetic fog data sets have a significant advantage in supporting fusion, detection, and other advanced visual task models.
[0003] In fog data set collection, real data acquisition is difficult and label updating is costly. To solve the problem of open source data scarcity, a virtual fog is often superimposed on the image to construct a fog image data set. However, the existing technical methods have many shortcomings in generating fog effects: first, the existing RGB channel-based synthetic fog technology only applies a uniform density fog layer to the image, lacks modeling of natural optical laws, and the generated effect is not realistic; second, the center point synthetic fog method based on the standard optical model applies the atmospheric scattering model, but due to the random selection of the center point, its usable range is limited, making it difficult to apply to complex outdoor environments; in addition, virtual scene simulation fog technology can generate paired images on a simulation platform, but the authenticity and universality of the generated image data still need to be verified, and the lack of labels limits its application in fusion and detection models. These shortcomings affect the quality and practical application effect of synthetic fog data sets. SUMMARY
[0004] Therefore, it is necessary to provide a fog image data set synthesis method and device based on depth estimation to improve the generation accuracy of synthetic fog data sets and improve the efficiency of building large-scale data sets.
[0005] A fog image data set synthesis method based on depth estimation, the method comprising:
[0006] A fog image data set synthesis model is constructed. The fog image data set synthesis model includes a depth estimation model, an optical fogging model, and a synthetic fog image model.
[0007] After optimizing and training the depth estimation model using a monocular depth estimation method based on a public dataset, the target image scene data is input into the optimized depth estimation model for semantic deepening, and a synthetic fog effect of the depth map corresponding to the target image scene data is output.
[0008] According to the synthetic fog effect, the depth map is atomized by the optical fog model to generate a fogged image.
[0009] The foggy visual task matching algorithm of the synthetic foggy image model is determined according to the foggy image and the preset target task, and the target foggy image dataset is generated according to the foggy visual task matching algorithm.
[0010] A device for synthesizing foggy image data based on depth estimation, the device comprising:
[0011] The synthesis model building module is used to build a synthesis model for foggy image data. The synthesis model for foggy image data includes: a depth estimation model, an optical fogging model, and a synthetic fog image model.
[0012] The synthetic fog effect generation module is used to optimize and train the depth estimation model using a monocular depth estimation method based on a public dataset, input the target image scene data into the optimized depth estimation model for semantic deepening, and output the synthetic fog effect of the depth map corresponding to the target image scene data.
[0013] The fog image generation module is used to generate a fog image after performing fog processing on the depth map through an optical fog model according to the synthetic fog effect.
[0014] The synthetic fog data module is used to determine the foggy visual task matching algorithm of the synthetic fog image model based on the fogged image and the preset target task, and generate the target foggy image dataset based on the foggy visual task matching algorithm.
[0015] The fog image data set synthesis method and device based on depth estimation can simulate the optical phenomena in the fog environment more accurately by introducing the combination of the depth estimation model and the optical fogging model. After optimizing and training the depth estimation model, the depth information of the target image is obtained by using the model, and semantic deepening is performed, so that a depth map that is more consistent with the real-world fog effect is generated. Then, the optical fogging model is used to fog the target image according to the depth information of the scene, to generate an image with a natural fog effect. Secondly, the fogging process of the optical fogging model is guided by the depth map information, so that a more fine and adaptive fogging effect in a complex outdoor environment is realized. Specifically, the depth map can provide spatial position information of different objects and backgrounds in the scene, so that the optical fogging model can generate fog layers with different densities and levels according to the actual structure of the scene, to simulate a more realistic fog effect. In addition, the technical solution also solves the problems of the virtual scene simulation fog technology in the generation of image data authenticity and universality. By using the depth estimation and the optical fogging model together, the fog image can be synthesized in the real scene, avoiding the limitation of the virtual scene. In addition, the synthesized fog image model can determine the fog visual task matching algorithm according to the characteristics of the original public data set, so as to generate a target fog image data set with original detailed labels, further improving the quality of the data set and the universality of the application, and successfully constructing a high-quality fog image data set with detailed labels. This data set can not only improve the application effect of the image processing and computer vision model in the fog environment, but also provide more real and efficient data support for target detection, fusion and other tasks in the fog scene. BRIEF DESCRIPTION OF DRAWINGS
[0016] Figure 1 A fogging image synthesis framework diagram for a fog image data set synthesis method based on depth estimation in an embodiment;
[0017] Figure 2 A flowchart diagram of a fog image data set synthesis method based on depth estimation in an embodiment;
[0018] Figure 3 A depth estimation model framework diagram based on ResNet in an embodiment;
[0019] Figure 4 A flowchart diagram of a standard optical model for synthesizing fogging images in an embodiment;
[0020] Figure 5 A comparison diagram of thick fog data synthesis results of the MFNet public data set in a daytime scene in an embodiment;
[0021] Figure 6Fig. 1 is a comparison diagram of the thick fog data synthesis results of the MFNet public dataset in a night scene in an embodiment;
[0022] Figure 7 Fig. 2 is a comparison diagram of the thick fog data synthesis results of the RoadScene public dataset in a day scene in an embodiment;
[0023] Figure 8 Fig. 3 is a comparison diagram of the thick fog data synthesis results of the RoadScene public dataset in a night scene in an embodiment;
[0024] Figure 9 Fig. 4 is a diagram of the dataset construction classification in an embodiment;
[0025] Figure 10 Fig. 5 is a comparison diagram of the synthetic fog images generated by different fog concentration parameters in an embodiment;
[0026] Figure 11 Fig. 6 is a structural block diagram of a foggy image data set synthesis device based on depth estimation in an embodiment;
[0027] Figure 12 Fig. 7 is an internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION
[0028] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.
[0029] The foggy image data set synthesis method based on depth estimation provided by the present application can be applied to a fogging image synthesis framework as shown in Figure 1 , which includes three modules: a depth estimation model, an optical fogging model and a synthetic fog data set, which respectively realize image scene depth estimation, optical model fogging processing and synthetic fog data pairing arrangement.
[0030] In an embodiment, as shown in Figure 2 , a foggy image data set synthesis method based on depth estimation is provided, which is described below by taking the fogging image synthesis framework in Figure 1 as an example, including the following steps:
[0031] Step 202, constructing a foggy image data synthesis model.
[0032] The foggy image data synthesis model includes a depth estimation model, an optical fogging model and a synthetic fog image model.
[0033] In step 204 , after optimizing and training the depth estimation model using a monocular depth estimation method based on a public dataset, the target image scene data is input into the optimized depth estimation model for semantic deepening, and a synthetic fog effect of a depth map corresponding to the target image scene data is output.
[0034] Step 206 : After performing a fogging process on the depth map using an optical fogging model according to the synthetic fog effect, a fogged image is generated.
[0035] Step 208 : determining a foggy visual task matching algorithm for the synthetic foggy image model according to the foggy image and the preset target task, and generating a target foggy image dataset according to the foggy visual task matching algorithm.
[0036] The above-mentioned method for synthesizing foggy image datasets based on depth estimation first introduces a combination of a depth estimation model and an optical fogging model to more accurately simulate the optical phenomena in foggy environments. After optimizing and training the depth estimation model, it is used to obtain depth information of the target image and perform semantic deepening to generate a depth map that more closely resembles real-world fog effects. Next, the optical fogging model performs fogging based on the scene's depth information, producing an image with a natural fog effect. Secondly, the depth map information guides the optical fogging model's fogging process, achieving a more refined fogging effect that is more adaptable to complex outdoor environments. Specifically, the depth map provides spatial position information of different objects and backgrounds in the scene, allowing the optical fogging model to generate fog layers of varying density and depth based on the scene's actual structure, simulating a more realistic foggy effect. Furthermore, this technical solution addresses the issues of virtual scene fog simulation technology regarding the authenticity and versatility of generated image data. By combining depth estimation with the optical fogging model, foggy images can be synthesized in real scenes, avoiding the limitations of virtual scenes. Furthermore, the synthetic fog image model also determines the fog vision task matching algorithm based on the characteristics of the original public dataset by presetting the target task. This generates a target fog image dataset with the original detailed labels, further improving the dataset's quality and versatility. This successfully constructs a high-quality, detailed fog image dataset. This dataset not only improves the application of image processing and computer vision models in foggy environments, but also provides more realistic and efficient data support for tasks such as object detection and fusion in foggy scenes.
[0037] In one embodiment, the depth estimation model is built based on ResNet, wherein the depth estimation model includes: a feature fusion module, a residual convolution module, and an adaptive output module.
[0038] It is worth noting that the depth estimation model is used to generate the depth map of the target image scene. Figure 3The model is built based on the ResNet framework, combined by feature fusion block, residual convolution block and adaptive output block, and can realize the conversion from input image to corresponding depth map. The monocular depth estimation method is adopted, and the model is trained by using large-scale data of multiple classic data sets. Specifically, the training of each data set is regarded as an independent task, and the approximate Pareto optimal solution is sought on the data set.
[0039] In one embodiment, the monocular depth estimation method is used to define the loss function of each training data set in the depth estimation model training according to different training data sets in the public data set:
[0040] ;
[0041] ;
[0042] wherein, is the size of the training data set l, is the scale and translation invariant loss, is the multi-scale and scale invariant gradient matching term, M is the number of pixels contained in the image corresponding to the training data set, is the disparity prediction value, is the label true value, is the type of loss function, is the hyperparameter. The least square method is used to minimize the loss function, and the scale and translation invariant loss function is obtained:
[0043] ;
[0044] ;
[0045] wherein, s is the resolution parameter, t is the position parameter, is the difference between the normalized disparity prediction value and the label value, is the difference of the disparity map at the scale k , is the gradient of in the x-axis direction, is the gradient of in the y-axis direction, is a multi-scale varying scale level.
[0046] By introducing 3D data set into each of the training data sets, if the scale and translation invariant loss function optimized according to the enhanced task is reduced, the minimum value of all scale and translation invariant loss functions after training is taken as the model optimization parameter:
[0047] ;
[0048] wherein, is the enhanced task being trained at present, is the model parameter. Conversely, the scale-invariant translation loss function is retained, the enhanced task being trained at present is discarded, and the task data set to be optimized is obtained. The depth estimation model is optimized according to the scale-invariant translation loss function and the task data set to be optimized, and the target image is trained by the optimized depth estimation model.
[0049] In one of the embodiments, the target image scene data and the target image are input into the optimized depth estimation model for semantic deepening, and a predicted depth map is output. The predicted depth map is subjected to gray scale conversion and brightness inversion to generate a first gray scale image.
[0050] The scene depth information of the first gray scale image is extracted by using a standard optical algorithm.
[0051] ;
[0052] ;
[0053] wherein, is the fog image, is the target image, is the transmission map, is the atmospheric light, is a pixel point in the image, is the scene depth, is the atmospheric scattering coefficient. The synthetic fog effect of the depth map corresponding to the target image scene data is generated according to the scene depth information.
[0054] It is worth noting that, based on the standard optical model, the fog image is synthesized by the depth map and the original clear image, and the specific process is as shown in Figure 4 First, the depth map obtained in the previous stage is subjected to gray scale conversion and brightness inversion to generate a corresponding gray scale image. Subsequently, the optical formula is used:
[0055] ;
[0056] wherein, is the fog image, is the target image, is the transmission map, is the atmospheric light, is a pixel point in the image. The transmission map can be obtained by the formula wherein represents the scene depth, is the atmospheric scattering coefficient.
[0057] In one of the embodiments, a fogged image corresponding to the depth map is generated after the visible light image of the depth map is superimposed and fogged according to the synthetic fog effect by the optical fogging model.
[0058] In one of the embodiments, the target task includes: constructing a foggy day visual task matching algorithm performance test dataset, constructing a foggy day infrared-visible light pixel-level paired fusion dataset, and constructing a foggy day target detection dataset.
[0059] In one of the embodiments, the foggy day visual task matching algorithm performance test dataset is constructed, and the specific steps are as follows: the semantic segmentation part of the MFNet public dataset is selected, 4-channel data is extracted and reconstructed first to obtain 3-channel RGB images. Then, the images are classified according to the original classification labels, and are divided into daytime and nighttime scenes. These classified data are input into a depth estimation model to match the corresponding fogging brightness, and different concentrations of daytime and nighttime foggy day image datasets are generated by adjusting the fog concentration parameter. The dataset retains the original image information and is paired with the depth map, and is suitable for effect test and comparative evaluation of the foggy day visual task matching algorithm, supporting subjective visual comparison and objective index evaluation. Part of the dataset is shown as Figure 5 and Figure 6 .
[0060] It is worth noting that the present method only needs to input clear images to quickly generate synthetic foggy day images and corresponding depth maps of multiple scenes and different concentrations.
[0061] In one of the embodiments, the foggy day infrared-visible light pixel-level paired fusion dataset is constructed, and the specific steps are as follows: the experiment is carried out based on the RoadScene public dataset. The dataset independently divides the infrared and visible light images. First, 221 visible light images in the dataset are classified, which are divided into daytime scenes under strong light conditions and nighttime scenes under weak light conditions, and the corresponding depth maps are generated using a depth estimation model. In view of the fact that the daytime scenes in the dataset are sunny and the nighttime scenes have more lights, we select a relatively high fogging brightness value to match the scene conditions. Finally, the fogged visible light images are paired with the original infrared images to construct a dataset containing pixel-level paired visible light images, infrared images, depth maps and fogged images, which is suitable for image fusion tasks. At the same time, retaining the original images is helpful for fusion algorithm evaluation based on subjective visual nature and objective indicators, such as comparative analysis of indicators such as visual information fidelity (VIF) and mutual information (MI). The foggy day infrared-visible light pixel-level paired fusion dataset constructed based on the RoadScene dataset is shown as Figure 7 , Figure 8As shown, the dataset realizes pixel-level pairing, and from left to right are the visible light image, the depth map, the fog image, and the infrared image.
[0062] It is worth noting that, on the basis of the existing visible-infrared paired dataset, the visible light image is superimposed with synthetic fog. Investigation shows that thin fog weather has little effect on infrared imaging, so the original infrared image data and labels are retained to construct a visible-infrared pixel-level paired dataset under foggy conditions. This dataset can effectively fill the gap in foggy multi-modal research and provide a large amount of reliable test data for related image fusion algorithms.
[0063] In one of the embodiments, a foggy target detection dataset is constructed, and the specific steps are as follows: the detection part of the MFNet public dataset is selected, and then it is classified according to the light conditions, such as Figure 9 As shown, it is divided into three categories: strong light, insufficient light, and weak light. The visible light image data in these categories are input into the depth estimation model, matched with the corresponding fog brightness, and then different fog concentration parameters are selected to generate foggy visible light image data under light conditions such as Figure 10 As shown, after pairing with the infrared image and detection labels in the original dataset, the foggy target detection dataset is obtained. Figure 10 is a partial image data comparison of the dataset: from top to bottom are the daytime thick fog, thin fog scene, night thick fog, and thin fog scene.
[0064] It is worth noting that, the existing annotated public dataset is processed with fog to generate foggy images with clear annotations to support the training and generalization of foggy target detection algorithms. At the same time, since the infrared image has less interference with fog, the infrared and visible light paired dataset is selected to evaluate the performance of the detection algorithm on different image data, thereby improving the overall accuracy.
[0065] It should be understood that, although Figure 2 , Figure 4 the flowcharts in Figure 2 , Figure 4 display each step in turn according to the direction of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless explicitly stated in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least part of the steps in
[0066] may include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps.In one embodiment, as shown in Figure 11 A foggy image data set synthesis device based on depth estimation is provided, comprising: a synthesis model construction module 1102, a synthetic fog effect generation module 1104, a fogged image generation module 1106, and a synthetic fog data module 1108, wherein:
[0067] The synthesis model construction module 1102 is configured to construct a foggy image data synthesis model. The foggy image data synthesis model comprises a depth estimation model, an optical fogging model, and a synthetic fog image model.
[0068] The synthetic fog effect generation module 1104 is configured to input target image scene data into the optimized depth estimation model for semantic deepening after the depth estimation model is optimized and trained by using a monocular depth estimation method based on a public data set, and output the synthetic fog effect of the depth map corresponding to the target image scene data.
[0069] The fogged image generation module 1106 is configured to generate a fogged image by fogging the depth map through the optical fogging model according to the synthetic fog effect.
[0070] The synthetic fog data module 1108 is configured to determine a foggy visual task matching algorithm of the synthetic fog image model according to the fogged image and a preset target task, and generate a target foggy image data set according to the foggy visual task matching algorithm.
[0071] The specific limitations of the foggy image data set synthesis device based on depth estimation can be referred to the limitations of the foggy image data set synthesis method based on depth estimation in the foregoing, which will not be repeated here. Each module in the above foggy image data set synthesis device based on depth estimation can be realized by software, hardware, and a combination thereof, in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each module.
[0072] In one embodiment, a computer device is provided, which can be a terminal, and the internal structure diagram thereof can be as shown in Figure 12As shown in the figure. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with the external terminal through the network connection. The computer program is executed by the processor to implement a depth estimation-based fog image data set synthesis method. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.
[0073] Those skilled in the art can understand that Figures 11-12 The structure shown in the figure is only a block diagram of part of the structure related to the present application scheme, and does not constitute a limitation on the computer device to which the present application scheme is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.
[0074] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the following steps:
[0075] A fog image data synthesis model is constructed. The fog image data synthesis model includes a depth estimation model, an optical fogging model, and a synthetic fog image model.
[0076] After the depth estimation model is optimized and trained by using a monocular depth estimation method according to a public data set, the target image scene data is input into the optimized depth estimation model for semantic deepening, and the synthetic fog effect of the depth map corresponding to the target image scene data is output.
[0077] After the depth map is fogged by the optical fogging model according to the synthetic fog effect, a fogged image is generated.
[0078] According to the fogged image and the preset target task, a foggy visual task matching algorithm of the synthetic fog image model is determined, and a target foggy image data set is generated according to the foggy visual task matching algorithm
[0079] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiment methods can be included. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0080] Any combination of the technical features of the above embodiments can be made, and in order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.
[0081] The above embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.
Claims
1. A method for synthesizing foggy image datasets based on depth estimation, characterized in that: The method comprises: Constructing a foggy image data synthesis model; the foggy image data synthesis model includes: a depth estimation model, an optical fogging model, and a synthetic fog image model; After optimizing and training the depth estimation model using a monocular depth estimation method based on a public dataset, the target image scene data is input into the optimized depth estimation model for semantic deepening, and a synthetic fog effect corresponding to the depth map of the target image scene data is output; Inputting target image scene data, i.e., the target image, into an optimized depth estimation model for semantic deepening, outputting a predicted depth map, and performing brightness inversion on the predicted depth map through grayscale conversion to generate a first grayscale map; The scene depth information of the first grayscale image is extracted using a standard optical algorithm: in, For the fog image, is the target image, is the transmission diagram, is atmospheric light, is the pixel in the image, is the scene depth, is the atmospheric scattering coefficient; Generate a synthetic fog effect of a depth map corresponding to the target image scene data using professional optical theory according to the scene depth information; After performing a fogging process on the depth map using the optical fogging model according to the synthetic fog effect, a fogged image is generated; The foggy visual task matching algorithm of the synthetic foggy image model is determined based on the foggy image and the preset target task, and a target foggy image dataset is generated based on the foggy visual task matching algorithm. The target tasks include: constructing a foggy visual task matching algorithm performance test dataset, constructing a foggy infrared-visible light pixel-level pairing fusion dataset, and constructing a foggy target detection dataset.
2. The method for synthesizing foggy image data based on depth estimation according to claim 1, characterized in that: The depth estimation model is built based on ResNet, wherein the depth estimation model includes: a feature fusion module, a residual convolution module and an adaptive output module.
3. The method for synthesizing foggy image data based on depth estimation according to claim 1, characterized in that: The depth estimation model is optimized and trained using a monocular depth estimation method based on a public dataset, including: According to different training data sets in the public data set, a monocular depth estimation method is used to define the loss function of each training data set in the depth estimation model training: in, For training data set The size of is the scale and translation invariance loss, is the multi-scale and scale-invariant gradient matching term, M is the number of pixels in the corresponding image of the training dataset, is the disparity prediction value, is the true value of the label, is the type of loss function, is a hyperparameter; The least squares method is used to minimize the loss function to obtain the scale and translation invariant loss function: in, s is the resolution parameter, t is a positional parameter, is the difference between the normalized disparity prediction value and the label value, For scale k The difference in the disparity map at In the x-axis direction The gradient, In the y-axis direction The gradient, is a scale level with multiple scale changes; By introducing a 3D dataset enhancement training task into each of the training datasets, if the scale and translation invariance loss function of the enhanced task optimization training is reduced, the minimum value of all scale and translation invariance loss functions after training is taken as the model optimization parameter: in, is the augmented task for current training, are model parameters; Otherwise, the scale and translation invariant loss function is retained, the enhanced task currently being trained is discarded, and the task dataset to be optimized is obtained; The depth estimation model is optimized according to the scale and translation invariant loss function and the task dataset to be optimized, and the target image is trained using the optimized depth estimation model.
4. The method for synthesizing foggy image data based on depth estimation according to claim 3, characterized in that: After performing a fogging process on the depth map using the optical fogging model according to the synthetic fog effect, a fogged image corresponding to the original target image is generated, including: After the visible light image of the depth map is superimposed and fogged using the optical fogging model according to the synthetic fog effect, a fogged image corresponding to the original target image is generated.
5. A device for synthesizing foggy image data based on depth estimation, characterized in that: A method for synthesizing foggy image data based on depth estimation according to any one of claims 1 to 4 is implemented, the device comprising: A synthesis model construction module is used to construct a foggy image data synthesis model; the foggy image data synthesis model includes: a depth estimation model, an optical fogging model and a synthetic fog image model; A synthetic fog effect generation module is used to optimize and train the depth estimation model using a monocular depth estimation method based on a public dataset, input the target image scene data into the optimized depth estimation model for semantic deepening, and output a synthetic fog effect corresponding to the depth map of the target image scene data; a fog image generation module, configured to generate a fog image by performing fogging processing on the depth map using the optical fog model according to the synthetic fog effect; The synthetic fog data module is used to determine the foggy visual task matching algorithm of the synthetic fog image model according to the fogged image and the preset target task, and generate the target foggy image dataset according to the foggy visual task matching algorithm.