Accumulated water area video segmentation method and system based on deep learning
Through the video segmentation method of water accumulation area based on deep learning, the adaptive feature library and segmentation network are used to solve the problem of unstable segmentation of water accumulation areas in different environments, and high-precision and robust water accumulation areas are achieved, and flood disaster assessment and emergency response are supported.
Patent Information
- Application Number
- CN202510170140.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-17
- Publication Date
- 2025-07-18
AI Technical Summary
The prior art is difficult to achieve stable division of water accumulation areas under different environments, affecting the accuracy of flood disaster assessment and emergency response.
The video segmentation method of water accumulation area based on deep learning is adopted. By establishing an adaptive feature library and segmentation network, the features are automatically adjusted to adapt to the appearance changes of the water body, and segmented with image and video data.
It realizes stable water segmentation under different environments, improves the accuracy and robustness of water accumulation areas, and supports real-time monitoring of flood disaster assessment and emergency response.
Smart Images
Figure CN120339870A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of flood disaster assessment, and particularly to a method and system for video segmentation of water accumulation areas based on deep learning. Background Art
[0002] Research shows that in recent years, due to multiple factors such as abnormal climate under global change and changes in the underlying surface caused by urbanization construction, the problem of urban waterlogging has intensified globally. The frequency and intensity of urban waterlogging have increased significantly, and in the future, a larger population will be exposed to the risk of urban waterlogging, especially in some countries in Asia and Africa. Currently, urban waterlogging detection can be classified according to the different data sources used, namely through water level sensors, remote sensing images, social media images, and video data. Traditional methods usually use water level sensors as data sources for water accumulation monitoring. However, due to the high cost of water level sensors, the number of waterlogging points that can be monitored is limited. Therefore, it is very difficult to popularize water depth estimation through traditional methods across the country's cities. In addition, remote sensing data obtained by satellites or other aircraft can be used to monitor large-scale floods. However, satellite remote sensing images are affected by clouds and vegetation canopies, and are usually affected by factors such as too low image resolution and long revisit periods. Therefore, it is still difficult to obtain detailed and real-time urban flood information in space from satellite remote sensing data. As the most advanced remote sensing method, unmanned aerial vehicles can provide detailed spatio-temporal information for the detection of flood areas. However, the method based on unmanned aerial vehicles still has problems such as airspace supervision, safe operation in rainy weather, and power issues, and it is difficult to obtain real-time urban waterlogging information on a large scale. Therefore, the research on the technology of segmenting urban waterlogging water accumulation areas based on images and videos is of great significance for improving traffic management and enhancing the emergency response capabilities of relevant departments. Summary of the Invention
[0003] In view of the problems existing in the above-mentioned prior art, the present invention is proposed.
[0004] Therefore, the technical problem to be solved by the present invention is: to provide a method for video segmentation of water accumulation areas based on deep learning, which enables the segmentation model to automatically adjust the learned features to capture the changes in the appearance of water bodies by establishing an adaptive feature library and a segmentation network, so that the segmentation results of water bodies in urban waterlogging inundation images and video scenes are more robust, and can solve to a certain extent the problem of unstable water accumulation segmentation results due to environmental problems, which is beneficial to flood disaster assessment and emergency response.
[0005] To solve the above technical problem, the present invention provides the following technical solution: a method for video segmentation of water accumulation areas based on deep learning, including: collecting and screening data, and performing annotation to obtain a first-level data set; dividing the first-level data set; constructing a first type of segmentation model to output a first type of segmentation result; constructing a second type of segmentation model to update the feature library.
[0006] As a preferred solution of the waterlogging area video segmentation method based on deep learning according to the present invention, wherein: collecting and screening data and performing annotation includes collecting relevant original data and screening it to ensure the relevance of the data, and using a certain type of tool to annotate the screened data to obtain a first-level data set, providing a standardized data set for the model.
[0007] As a preferred solution of the waterlogging area video segmentation method based on deep learning according to the present invention, wherein: constructing a first type of segmentation model includes training based on image data and learning using the first-level data set to output a first type of segmentation result.
[0008] As a preferred solution of the waterlogging area video segmentation method based on deep learning according to the present invention, wherein: constructing a second type of segmentation model includes analyzing and segmenting based on video data, extracting key feature information from the video, and generating a segmentation result.
[0009] As a preferred solution of the waterlogging area video segmentation method based on deep learning according to the present invention, wherein: updating the feature library includes continuously maintaining and optimizing the process of storing feature information during data processing and training.
[0010] As a preferred solution of the waterlogging area video segmentation method based on deep learning according to the present invention, wherein: the first-level data set includes a training set, a validation set, and a test set.
[0011] As a preferred solution of the waterlogging area video segmentation method based on deep learning according to the present invention, wherein: dividing the first-level data set includes dividing the annotated data according to a ratio of 7:2:1 into subsets for different purposes.
[0012] Another object of the present invention is to provide a waterlogging area video segmentation system based on deep learning, which can automatically adjust the feature library and the segmentation network to adapt to the changes in the appearance of water bodies in different video frames, thereby achieving a stable water body segmentation effect; such a segmentation system can improve the accuracy and robustness of waterlogging area recognition, and contribute to flood disaster assessment and emergency response in complex environments.
[0013] To solve the above technical problems, the present invention provides the following technical solution: a waterlogging area video segmentation system based on deep learning, including: a data annotation module, a data division module, a segmentation model module, and a feature update module;
[0014] The data annotation module collects and screens data and performs annotation to obtain a first-level data set;
[0015] The data division module divides the first-level data set;
[0016] The segmentation model module constructs a type of segmentation model and outputs a type of segmentation result.
[0017] The feature update module constructs a second type of segmentation model and updates the feature library.
[0018] A computer device includes a memory and a processor. The memory stores a computer program. It is characterized in that when the processor executes the computer program, the steps of the above-mentioned deep learning-based waterlogging area video segmentation method are implemented.
[0019] A computer-readable storage medium stores a computer program. It is characterized in that when the computer program is executed by a processor, the steps of the above-mentioned deep learning-based waterlogging area video segmentation method are implemented.
[0020] The beneficial effects of the present invention: Through the deep learning-based waterlogging area video segmentation method, the efficient identification and segmentation of urban waterlogging areas are realized, and significant beneficial effects are achieved. First, the method introduces an adaptive feature library, which can automatically adjust the feature descriptor to cope with the dynamic changes in the appearance of water bodies in video frames, enabling the segmentation model to maintain stable segmentation performance in different scenarios and conditions. Second, by combining the image-based segmentation model and the video segmentation model, the accuracy and generalization ability of segmentation are effectively improved, meeting the requirements of real-time monitoring. In addition, through preprocessing steps such as data cleaning, annotation, and data augmentation, the robustness of the model is further enhanced, reducing the overfitting phenomenon, and making it show high accuracy and reliability in different environments. This technology can be widely applied to flood disaster monitoring and early warning, improve the response speed of urban emergency management, and provide a scientific basis and technical support for disaster prevention and control. Description of the Drawings
[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.
[0022] Figure 1 It is a flowchart of the deep learning-based waterlogging area video segmentation method provided by an embodiment of the present invention.
[0023] Figure 2 It is a schematic diagram of the image-based urban waterlogging area segmentation method of the deep learning-based waterlogging area video segmentation method provided by an embodiment of the present invention.
[0024] Figure 3Module diagram of the video-based urban waterlogging ponding area segmentation method for the ponding area video segmentation method based on deep learning provided by an embodiment of the present invention.
[0025] Figure 4 Example diagram of ponding area segmentation for the ponding area video segmentation method based on deep learning provided by an embodiment of the present invention. Detailed implementation manners
[0026] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.
[0027] Example 1, referring to Figure 1 , which is an embodiment of the present invention. This embodiment provides a ponding area video segmentation method based on deep learning, including:
[0028] S1: Collect and screen data, and perform annotation to obtain a primary data set.
[0029] It should be noted that, as shown in S1 of Figure 1 , collecting and screening data and performing annotation include collecting relevant raw data and screening it to ensure the relevance of the data, and using a certain type of tool to annotate the screened data to obtain a primary data set, providing a standardized data set for the model.
[0030] Furthermore, the data screening step aims to ensure that the collected data is highly relevant and meets the identification requirements of the ponding area, removing non-flood scenarios or data with poor quality, thereby improving the quality of the data set. For the screened data, use the Roboflow tool for annotation. Specifically, mark the ponding area on each image in the form of a label to form a standardized data set.
[0031] Even further, to increase the diversity of the data and enhance the generalization ability of the model, perform data augmentation operations on the annotated data; specifically, the data augmentation process includes various processing methods such as rotating, flipping, and blurring the images, thereby generating a more diverse data set and avoiding overfitting phenomena during the training process of the model.
[0032] In the embodiments of the present application, the primary dataset consists of a training set, a validation set, and a test set. The training set is used for the initial training of the model, the validation set is used to adjust the model parameters during the training process, and the test set is used to finally evaluate the performance of the model. To meet the requirements of model training, these filtered datasets are labeled. The Roboflow tool is used to accurately label the water accumulation areas in each image, forming standardized data labels, so that the dataset is more suitable for deep learning segmentation tasks.
[0033] In an alternative embodiment, the primary dataset can also be implemented in other ways. Some flood-related image and video data can be obtained from public datasets or open data of government departments, without relying on data collection from social media platforms. This method can provide basic image and video data of water accumulation areas and reduce the dependence on social media data. However, since the sources of these public datasets are relatively scattered and the update frequency is low, they may not fully cover flood scenarios in different regions and time periods, and the diversity and timeliness of the data are insufficient, which may affect the generalization ability of the model.
[0034] In an alternative embodiment, the primary dataset can also be implemented in other ways. For example, the dataset can be obtained by collaborating with different data sources (such as geographic information systems or hydrological stations). The image data provided by these data sources usually has high accuracy, but its data update frequency is low and the data volume is limited. Although the image quality of these data sources is high, due to the limitations of the acquisition frequency and angle, the diversity of the data is not as rich as that of social media images, resulting in the model may have difficulty adapting to changes in water accumulation areas in different scenarios and environments, affecting the generalization ability.
[0035] In the embodiments of the present application, one type of tool is to perform labeling through Roboflow, and perform data augmentation on the image data by rotating, blurring, etc. to prevent model overfitting and improve the generalization ability of the model, obtaining the augmented image data. Through the Roboflow labeling tool, the filtered flood images are accurately labeled one by one for the water accumulation areas, so that the water accumulation areas in each image can be clearly identified. The methods of data augmentation are not limited to rotation and blurring, but also include various operations such as flipping, scaling, and color adjustment of the images to generate more diverse samples.
[0036] In an alternative embodiment, a type of tool can also be implemented in other ways. For example, for the annotation tool, simple image editing software such as Photoshop or GIMP can be selected to manually annotate each image of the water accumulation area. Although this method can complete the annotation of the water accumulation area, due to the lack of a dedicated annotation platform, the operation efficiency may be low. Especially in the case of a large amount of data, the annotation process will consume a lot of manpower and time. In addition, the accuracy of manual annotation may be affected by human factors, and it is difficult to ensure the consistency of annotation, which may lead to errors in the model during training and affect the final segmentation effect.
[0037] In an alternative embodiment, a type of tool can also be implemented in other ways. For example, during the data annotation process, semi-automated annotation tools can be used. These tools help users quickly delineate the water accumulation area through simple segmentation algorithms, but still require manual correction. Although such semi-automated tools can reduce the annotation workload to a certain extent, the accuracy of their automatic segmentation is low, and a large amount of manual intervention is required to ensure the accuracy of annotation. Eventually, the standardization and consistency of the final annotation may not be as good as that of fully automated annotation tools; insufficient quality of the annotated data may cause errors in subsequent model training, resulting in unstable performance of the segmentation model in complex scenarios.
[0038] Embodiment 2, referring to Figures 1 - 3 , is an embodiment of the present invention. This embodiment provides a method for video segmentation of water accumulation areas based on deep learning, including:
[0039] S2: Divide the first-level dataset.
[0040] It should be noted that, as shown in S2 of Figure 1 , dividing the first-level dataset includes dividing the annotated data into subsets for different purposes according to a ratio of 7:2:1.
[0041] Furthermore, according to the annotation, the expanded image data is 5900 pieces. Finally, the expanded dataset is divided into a training set, a validation set, and a test set according to a ratio of 7:2:1, which are 4193 pieces, 1198 pieces, and 599 pieces respectively.
[0042] S3: Construct a type of segmentation model and output a type of segmentation result.
[0043] It should be noted that, as shown in S3 of Figure 1 , constructing a type of segmentation model includes training based on the image data and learning using the first-level dataset to output a type of segmentation result.
[0044] Further, first, a pre-trained model such as ResNet101 is used as the backbone network, and the feature maps of the intermediate layers are taken from the backbone network and named C3, C4, and C5 respectively; through the Feature Pyramid Network (FPN) structure, multi-scale feature maps P3, P4, and P5 are generated from the C3, C4, and C5 feature maps, and higher-resolution P6 and P7 feature maps are generated from the P5 feature map; the P3 feature map is input into the prototype network sub-network, and the prototype network will output k original feature maps with a size of 138×138; these original feature maps are used as shared segmentation features; for each feature map P3 - P7, the classification sub-network is used to predict the location information, mask coefficients, and classification confidence respectively; non-maximum suppression is performed on the location, mask, and confidence rate information generated in the previous step, and duplicate detection targets are efficiently removed through non-maximum suppression, and the most significant detection results are retained; finally, the location and mask coefficient information retained through non-maximum suppression is combined with the k original feature maps generated by the prototype network, including operations such as overlay, cropping, and threshold segmentation, and accurate instance segmentation results, that is, one type of segmentation results, are finally output.
[0045] Furthermore, as Figure 2 shown, step one, a pre-trained model such as ResNet101 is used as the backbone network, and the feature maps of the intermediate layers are taken from the backbone network and named C3, C4, and C5 respectively;
[0046] Step two, through the Feature Pyramid Network (FPN) structure, multi-scale feature maps P3, P4, and P5 are generated from the C3, C4, and C5 feature maps, and higher-resolution P6 and P7 feature maps are generated from the P5 feature map;
[0047] Step three, the P3 feature map is input into the prototype network sub-network, and the prototype network will output k original feature maps with a size of 138×138, and these original feature maps are used as shared segmentation features;
[0048] Step four, for each feature map P3 - P7, the classification sub-network is used to predict the location information, mask coefficients, and classification confidence respectively;
[0049] Step five, non-maximum suppression is performed on the location, mask, and confidence rate information generated in the previous step, and duplicate detection targets are efficiently removed through non-maximum suppression, and the most significant detection results are retained;
[0050] Step six, the location and mask coefficient information retained through non-maximum suppression is combined with the k original feature maps generated by the prototype network, including operations such as overlay, cropping, and threshold segmentation, and accurate instance segmentation results are finally output.
[0051] S4: Build a two-class segmentation model and update the feature library.
[0052] It should be noted that, as Figure 1 shown in S4 of
[0053] , the updated feature library includes the process of continuously maintaining and optimizing the storage of feature information during data processing and training. obj and F background which store water body and non-water body features respectively; once a new frame is segmented, its features are compared with the feature library, and then a water body mask is generated according to the similarity score; in order to adapt to the changes in the appearance of objects in the video, the feature library can be updated through three operations: merge, append, and delete; when a new feature f new is significantly different from the existing features, the new feature is appended to the feature library; however, when a new feature is close to the existing features in the feature library, the nearest feature descriptor f closest in the feature library is updated by weighted averaging, so as to merge them, as shown in formula (1) specifically:
[0054] f′ closest =αf closest +(1 - α)f new (1)
[0055] where f' closest is the updated feature descriptor, and the weight coefficient α = 10 is set in the experiment. In addition, the feature library discards outdated features according to the least frequently used index; during video segmentation, since the feature library can adaptively adjust the feature descriptor according to the changing scene, such a feature library can enable the entire segmentation framework to effectively remember the appearance of background objects changing over time.
[0056] The segmentation network follows the encoder-decoder architecture. The encoder aims to encode the current frame It into its feature map f t for segmentation, as shown in formula (2) specifically:
[0057] f t =Encoder(I t ) (2)
[0058] where I t ∈R h×w×3 is an RGB image with height h and width w. We establish an attention module Atten to calculate the similarity between the current frame feature f t and F obj and F backgroundSimilarity between features; The attention module is used to enable the network to learn and focus more on important information rather than useless features; We use this attention module to calculate the most relevant feature g from the feature library obj and g background :
[0059] g obj = Atten(f t , F obj ) (3)
[0060] g background = Atten(f t , F background ) (4)
[0061] Then, the decoder network obtains the feature map f of the current frame t and combines it with the similarity information calculated from the attention module to estimate the water body segmentation mask S of the current frame t ; In the decoder module, the size of g t and g obj is gradually enlarged by utilizing the information in the feature map f background ; By checking the maximum score in the object or background map, the class label of each pixel is assigned as the estimated segmentation mask S t ; Specifically, as shown in formula (5):
[0062] S t = Decoder(f t , g obj , g background ) (5)
[0063] where S t has the same resolution as the original input frame and will be used to update the feature library
[0064] Furthermore, Precision (P), Recall (R), and Average Precision (AP) are used to evaluate the accuracy of the segmentation model; Precision, also known as the precision rate, refers to the proportion of correctly predicted positives among all predicted positives, Recall refers to the proportion of correctly predicted positives among all actual positives, AP is the average of all Precisions corresponding to Recall values between 0 and 1, AP50 refers to the IOU value of 50%, and AP@50:95 refers to the IOU value ranging from 50% to 95%, and then the mean of AP under these IOUs is calculated; The three accuracy evaluation methods complement each other for the errors of sample imbalance, false detection, and missed detection; The calculation formulas for the accuracy evaluation indicators are as follows:
[0065]
[0066] Among them, TP represents that the actual positive samples are predicted as positive, FP represents that the negative samples are predicted as positive, and FN represents that the positive samples are predicted as negative; r1, r 2, r n is the recall value corresponding to the first interpolation point of the precision interpolation segment arranged in ascending order.
[0067] Specifically, as Figure 3 shown: The red area estimates the preliminary segmentation based on the matching module and proposes an adaptive feature library to organize the feature space. The blue area designs a novel uncertain region refinement mechanism to achieve fine-grained segmentation; ① is used to encode the current frame (referred to as the query frame) into a feature map for segmentation; by using ResNet-50 as the backbone network and taking the output of the third layer as the feature map. ② To segment the t-th frame, the past frames from 1 to t - 1 are regarded as reference frames; a reference encoder to memorize the features of the target object (water body); assuming there are L objects of interest, the reference frames need to be encoded object by object and L feature maps are output. ③ The feature maps are encoded into two embedding spaces through two convolutional modules, respectively called key k and value v; the feature maps are matched through their keys k while allowing their values v to be different to retain as much semantic information as possible. ④ An adaptive feature library (AFB) to more effectively manage the key features of the object; AFB contains two operations: receiving new features and removing outdated features; when extracting new features, if it is close enough to an existing feature, they are merged; and the old features are evaluated by the LFU metric to remove the feature. ⑤ A local refinement mechanism to refine the fuzzy regions; by given two adjacent points in the space, if they belong to the same object, their features are usually similar; that is, using the pixels with high confidence in classification to refine other uncertain points in their neighborhood.
[0068] The above is a schematic solution of a deep learning-based waterlogging area video segmentation method according to this embodiment. It should be noted that the technical solution of the system of the deep learning-based waterlogging area video segmentation method belongs to the same concept as the above-mentioned deep learning-based waterlogging area video segmentation method. For the details not described in detail in the technical solution of the deep learning-based waterlogging area video segmentation system in this embodiment, reference can be made to the description of the technical solution of the above-mentioned deep learning-based waterlogging area video segmentation method.
[0069] Example 3, referring to Figure 4 , an embodiment of the present invention provides a deep learning-based waterlogging area video segmentation method. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.
[0070] 1. Data preprocessing
[0071] The data used in the present invention mainly includes: social media flood images and video data. After screening, 2305 image data were obtained. Then, they were labeled through Rokgflgw, and the image data were enhanced by rotation, blurring, etc. to prevent model overfitting and improve the generalization ability of the model, resulting in 5990 augmented image data. Finally, the augmented dataset was divided into a training set, a validation set, and a test set according to the ratio of 7:2:1, with 4193, 1198, and 599 images respectively.
[0072] 2. Result verification
[0073] For the method for segmenting urban waterlogging areas based on images and videos involved in the present invention, from Table 1, the accuracy of the waterlogging segmentation mask is 0.935, the recall rate is 0.903, the average accuracy is estimated to be 0.949 with 0.5 as the IoU threshold, the IoU values range from 50% to 95%, and the mean of AP calculated at these IoUs is 0.86. Figure 4 It can be seen that in different waterlogging environments, the boundaries of the waterlogging segmentation areas are clear and the accuracy is relatively high. Therefore, we believe that the method for segmenting urban waterlogging areas based on images and videos proposed in this study has a high credibility. Table 1 is as follows:
[0074] Table 1 Training results
[0075]
[0076]
[0077] 3. Conclusion
[0078] Through the verification of the experimental results, we believe that the method for segmenting urban waterlogging areas based on images and videos proposed in this study can better segment the waterlogging areas in different urban waterlogging scenarios, and the waterlogging segmentation models under different weather and lighting conditions can operate reliably, which solves to a certain extent the problem of unstable waterlogging segmentation results due to environmental problems and is conducive to flood disaster assessment and emergency response.
[0079] Example 4 is an embodiment of the present invention, which provides a video segmentation system for waterlogging areas based on deep learning, including: a data annotation module, a data division module, a segmentation model module, and a feature update module;
[0080] The data annotation module collects and screens data and performs annotation to obtain a primary dataset;
[0081] The data division module divides the primary dataset;
[0082] The segmentation model module constructs a type of segmentation model and outputs a type of segmentation result;
[0083] The feature update module constructs a binary segmentation model and updates the feature library.
[0084] This embodiment also provides a computing device applicable to the case of a waterlogging area video segmentation method based on deep learning, including:
[0085] A memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the waterlogging area video segmentation method based on deep learning as proposed in the above embodiment.
[0086] This embodiment also provides a storage medium on which a computer program is stored, and when the program is executed by a processor, it implements the waterlogging area video segmentation method based on deep learning as proposed in the above embodiment.
[0087] The storage medium proposed in this embodiment and the waterlogging area video segmentation method based on deep learning proposed in the above embodiment belong to the same inventive concept. Technical details not described in detail in this embodiment can be referred to in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.
[0088] If the above functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.
[0089] The logic and / or steps represented in the flowchart or otherwise described herein can, for example, be considered as a definitional sequence of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.
[0090] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), and the like.
[0091] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.
Claims
1. A video segmentation method for waterlogging areas based on deep learning, characterized in that: Including: Collect and screen data, and perform annotation to obtain a primary dataset; Divide the primary dataset; Construct a first-class segmentation model and output a first-class segmentation result; Construct a second-class segmentation model and update the feature library.
2. The method for video segmentation of water accumulation areas based on deep learning according to claim 1, characterized in that: The collecting and screening data and using them for annotation include collecting relevant original data and screening them to ensure the relevance of the data, and using a first-class tool to annotate the screened data to obtain a primary dataset, providing a standardized dataset for the model.
3. The method for segmenting water accumulation area videos based on deep learning according to claim 2, wherein: The constructing of the first-class segmentation model includes training based on image data and learning using the primary dataset to output a first-class segmentation result.
4. The method for video segmentation of waterlogging areas based on deep learning according to claim 3, characterized in that: The constructing of the second-class segmentation model includes analyzing and segmenting based on video data, extracting key feature information from the video, and generating a segmentation result.
5. The method for video segmentation of waterlogging areas based on deep learning according to claim 4, characterized in that: The updating of the feature library includes continuously maintaining and optimizing the process of storing feature information during data processing and training.
6. The method for video segmentation of waterlogging areas based on deep learning according to claim 5, characterized in that: The primary dataset includes a training set, a validation set, and a test set.
7. The method for video segmentation of water accumulation areas based on deep learning according to claim 6, characterized in that: The dividing of the primary dataset includes dividing the annotated data in a ratio of 7:2:1 into subsets for different purposes.
8. A system for video segmentation of waterlogging areas based on deep learning according to any one of claims 1-7, characterized in that: Including: A data annotation module, a data division module, a segmentation model module, and a feature update module; The data annotation module collects and screens data and performs annotation to obtain a primary dataset; The data division module divides the primary dataset; The segmentation model module constructs a first-class segmentation model and outputs a first-class segmentation result; The feature update module constructs a second-class segmentation model and updates the feature library.
9. A computer device, including a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning-based video segmentation method for waterlogging areas described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based video segmentation method for waterlogging areas described in any one of claims 1 to 7.