Water area range identification method and device and water area identification model training method and device

Abstract

The invention relates to a water area range recognition method and device and a water area recognition model training method and device. The method comprises the following steps: acquiring to-be-identified water area data; and inputting the to-be-identified water area data into a pre-trained water area identification model, and outputting a boundary range and a water surface line range of the to-be-identified water area through the water area identification model. The water area recognition model is obtained by performing resolution enhancement processing and image noise removal processing on various different types of river water area data, performing data labeling and training by using a labeling result of the data labeling; the water area recognition model comprises an encoder network and a decoder network, and the encoder network comprises a backbone network for extracting deep semantic features and a cavity pyramid pooling layer for performing cavity convolution; the decoder network comprises a bottom feature decoding layer for decoding features output by the backbone network and a deep feature decoding layer for decoding features output by the void pyramid pooling layer. By adopting the method, the water area can be accurately identified.

Description

Technical Field

[0001] The present disclosure relates to the technical field of image data processing, and particularly to a method for identifying a water area range, a method for training a water area recognition model, and an apparatus. Background Art

[0002] With the development of image data processing technology, more and more industries apply image recognition technology. Traditional flash flood monitoring methods mainly rely on manual observation and simple water level monitoring devices, which have problems such as lagging response and insufficient accuracy. With the development of computer vision and artificial intelligence technologies, flood monitoring methods based on image recognition have gradually attracted attention.

[0003] However, existing flood monitoring systems based on video recognition still have some problems: the ability to identify water areas in complex environments is insufficient, such as the recognition accuracy is relatively low in cases of fog, rain, or when there are reflections or ripples on the water surface. Summary of the Invention

[0004] Based on this, it is necessary to provide a method for identifying a water area range, a method for training a water area recognition model, and an apparatus for solving the above technical problems.

[0005] In a first aspect, the present disclosure provides a method for identifying a water area range. The method includes:

[0006] Obtain water area data to be recognized; the water area data to be recognized includes video data and / or image data;

[0007] Input the water area data to be recognized into a pre-trained water area recognition model, and output the boundary range and water surface line range of the water area to be recognized through the water area recognition model.

[0008] Wherein, the water area recognition model is obtained by performing resolution enhancement processing and image noise removal processing on various different types of river water area data and then performing data annotation, and training with the annotation results of the data annotation;

[0009] The water area recognition model includes: an encoder network and a decoder network. The encoder network includes: a backbone network for extracting deep semantic features and an atrous spatial pyramid pooling layer for performing atrous convolution, and the decoder network includes: a bottom layer feature decoding layer for decoding the features output by the backbone network and a deep layer feature decoding layer for decoding the features output by the atrous spatial pyramid pooling layer.

[0010] In one embodiment, the method further includes:

[0011] Obtain various different types of river water area data;

[0012] Perform resolution enhancement processing on the different types of river channel water area data with a resolution less than a preset resolution threshold using resolution enhancement technology;

[0013] Perform image denoising processing on the different types of river channel water area data after resolution enhancement processing and before resolution enhancement processing to obtain the river channel water area data to be labeled;

[0014] Perform water area boundary annotation and water surface line range annotation on the river channel water area data to be labeled;

[0015] Train a neural network model based on the annotation results of water area boundary annotation and water surface line range annotation in the river channel water area data to be labeled to obtain a water area recognition model.

[0016] In one embodiment, the performing resolution enhancement processing on the different types of river channel water area data with a resolution less than a preset resolution threshold using resolution enhancement technology includes:

[0017] Perform resolution enhancement processing on the different types of river channel water area data with a resolution less than a preset resolution threshold using a super-resolution reconstruction algorithm.

[0018] In one embodiment, the image denoising processing includes: image de-raining and image de-fogging processing. The performing image denoising processing on the different types of river channel water area data after resolution enhancement processing and before resolution enhancement processing to obtain the river channel water area data to be labeled includes one or more of the following:

[0019] Perform image de-raining and image de-fogging processing on the different types of river channel water area data after resolution enhancement processing and before resolution enhancement processing using a global histogram equalization algorithm;

[0020] Perform image de-raining and image de-fogging processing on the different types of river channel water area data after resolution enhancement processing and before resolution enhancement processing using a homomorphic filtering algorithm;

[0021] Accelerate the Retinex algorithm using recursive Gaussian filtering and perform image de-raining and image de-fogging processing on the different types of river channel water area data after resolution enhancement processing and before resolution enhancement processing using a linear stretching method;

[0022] Perform image de-raining and image de-fogging processing on the different types of river channel water area data after resolution enhancement processing and before resolution enhancement processing using curvelet transform.

[0023] In one embodiment, the image processing includes: image de-inversion and ripple processing, and performing image denoising processing on different types of river water area data after and before resolution enhancement processing to obtain river water area data to be labeled, including:

[0024] Using histogram equalization image enhancement technology and the deeplabv3+ algorithm, perform deconstruction analysis on different types of river water area data after and before resolution enhancement processing from multiple scales, and perform image de-inversion and ripple processing;

[0025] And / or, use a pre-trained inversion and ripple recognition model to identify inversions and ripples in different types of river water area data after and before resolution enhancement processing, and remove the inversions and ripples.

[0026] In one embodiment, the method further includes:

[0027] In response to the river water area data in the multiple different types of river water area data including river water area data in foggy weather, perform image enhancement processing on the river water area data in foggy weather based on the atmospheric modulation transfer function.

[0028] In a second aspect, the present disclosure also provides a method for training a water area recognition model, the method including:

[0029] Obtain multiple different types of river water area data;

[0030] Use resolution enhancement technology to perform resolution enhancement processing on the different types of river water area data with a resolution less than a preset resolution threshold;

[0031] Perform image denoising processing on different types of river water area data after and before resolution enhancement processing to obtain river water area data to be labeled;

[0032] Perform water area boundary annotation and water surface line range annotation on the river water area data to be labeled;

[0033] Train a neural network model according to the annotation results of water area boundary annotation and water surface line range annotation in the river water area data to be labeled to obtain a water area recognition model.

[0034] In a third aspect, the present disclosure also provides a water area range recognition device. The device includes:

[0035] A data acquisition module for acquiring water area data to be recognized; the water area data to be recognized includes video data and / or image data;

[0036] A model processing module, configured to input water area data to be recognized into a pre-trained water area recognition model, and output the boundary range and water surface line range of the water area to be recognized via the water area recognition model;

[0037] Wherein, the water area recognition model is obtained by performing resolution enhancement processing and image noise removal processing on various different types of river water area data and then performing data annotation, and training using the annotation results of the data annotation;

[0038] The water area recognition model includes: an encoder network and a decoder network. The encoder network includes: a backbone network for extracting deep semantic features and an atrous pyramid pooling layer for performing atrous convolution. The decoder network includes: a low-level feature decoding layer for decoding the features output by the backbone network and a deep-level feature decoding layer for decoding the features output by the atrous pyramid pooling layer.

[0039] Fourthly, the present disclosure also provides a water area recognition model training device, and the device includes:

[0040] A training data acquisition module, configured to acquire various different types of river water area data;

[0041] A resolution enhancement module, configured to perform resolution enhancement processing on the different types of river water area data with a resolution less than a preset resolution threshold by using resolution enhancement technology;

[0042] An image denoising module, configured to perform image denoising processing on the different types of river water area data after resolution enhancement processing and without resolution enhancement processing to obtain river water area data to be annotated;

[0043] A annotation module, configured to perform water area boundary annotation and water surface line range annotation on the river water area data to be annotated;

[0044] A training module, configured to train a neural network model according to the annotation results of the water area boundary annotation and water surface line range annotation in the river water area data to be annotated to obtain a water area recognition model.

[0045] Fifthly, the present disclosure also provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps in any of the above method embodiments are implemented.

[0046] Sixthly, the present disclosure also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in any of the above method embodiments are implemented.

[0047] In a seventh aspect, the present disclosure also provides a computer program product. The computer program product includes a computer program which, when executed by a processor, implements the steps in any of the above method embodiments.

[0048] In the above embodiments, the water area recognition model performs resolution enhancement processing on various different types of river water area data, enabling water area features that may originally be blurred due to low resolution to become clearer, thereby allowing the model to more accurately identify and distinguish water areas from non-water areas, and precisely determine the boundary range and water surface line range. For example, in the monitoring of small rivers in some remote mountainous areas, after processing low-resolution video or image data, the model can better capture subtle water flow changes and the blurred transition areas of water area boundaries, reducing misjudgments. The image noise removal processing effectively reduces the impact of noise generated by environmental factors (such as bad weather, light interference, etc.) on the image. In actual scenarios, rain, fog, reflection, etc. may all cause image noise. After processing, the model can focus on real water area features, further improving the recognition accuracy and ensuring stable output of reliable results in complex environments. Training with various different types of river water area data, covering samples in different dimensions such as watershed size, river width, geomorphic conditions, time, and light, enables the model to learn rich and diverse water area feature patterns, thereby enhancing its generalization ability in different actual scenarios. Whether it is a wide river or a narrow mountain stream, and whether it is a water area under sunny or bad weather conditions, the model can more effectively identify and analyze, with a wider application range, reducing over-reliance on specific scenarios and improving the practicality and reliability of the system. The backbone network in the encoder network uses an efficient convolutional neural network for deep semantic feature extraction, and uses depthwise separable convolutions to reduce the computational amount and the number of model parameters, improving the computational efficiency while ensuring the model performance. The atrous spatial pyramid pooling layer performs atrous convolution to increase the feature receptive field, which helps to capture feature information in a wider area and better understand the overall structure and change trend of the water area. The bottom feature decoding layer and the deep feature decoding layer in the decoder network decode and combine features from different sources respectively, enabling more refined reconstruction of image features, improving the accuracy of pixel-level classification results, and thus more accurately outputting the boundary range and water surface line range of the water area. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the specific embodiments of the present disclosure or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present disclosure. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0050] Figure 1 Schematic diagram of the application environment of the water area range recognition method in an embodiment;

[0051] Figure 2 Schematic flowchart of the water area range recognition method in an embodiment;

[0052] Figure 3 Schematic diagram of dilated convolution in an embodiment;

[0053] Figure 4 Schematic flowchart of the model training steps in an embodiment;

[0054] Figure 5 Schematic flowchart of one process of step S306 in an embodiment;

[0055] Figure 6 Schematic flowchart of another process of step S306 in an embodiment;

[0056] Figure 7 Schematic block diagram of the structure of the water area range recognition device in an embodiment;

[0057] Figure 8 Schematic block diagram of the structure of the domain recognition model training device in an embodiment

[0058] Figure 9 Schematic diagram of the internal structure of a computer device in an embodiment;

[0059] Figure 10 Schematic diagram of the internal structure of a computer device in an embodiment. Detailed implementation manners

[0060] In order to make the purpose, technical solutions and advantages of the present disclosure clearer, the present disclosure will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present disclosure, and are not used to limit the present disclosure.

[0061] It should be noted that the terms "first", "second", etc. in the specification and claims of this article and the above-mentioned drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in an order different from those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, device, product or equipment that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or equipment.

[0062] In this text, the term "and / or" is merely a description of the association relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. Additionally, in this text, the character " / " generally indicates that the associated objects before and after are in an "or" relationship.

[0063] The embodiments of the present disclosure provide a method for identifying a water area range, which can be applied to an application environment as Figure 1 shown. Among them, the terminal 102 communicates with the server 104 through a network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed in the cloud or on other network servers. The terminal 102 can obtain the water area data to be identified stored in the server 104, and the water area data to be identified includes video data and / or image data. The terminal 102 can input the water area data to be identified into the water area identification model pre-trained by the terminal 102 or the server 104, and the boundary range and water surface line range of the water area to be identified are output via the water area identification model. Among them, the water area identification model is obtained by the terminal 102 or the server 104 performing resolution enhancement processing and image noise removal processing on various different types of river water area data and then performing data annotation, and training with the annotation results of the data annotation. The water area identification model includes: an encoder network and a decoder network. The encoder network includes: a backbone network for extracting deep semantic features and an atrous spatial pyramid pooling layer for performing atrous convolution. The decoder network includes: a bottom feature decoding layer for decoding the features output by the backbone network and a deep feature decoding layer for decoding the features output by the atrous spatial pyramid pooling layer. Among them, the terminal 102 can be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, etc. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0064] In one embodiment, as Figure 2 shown, a method for identifying a water area range is provided. Taking the method applied to the Figure 1 terminal 102 as an example for illustration, the method includes the following steps:

[0065] S202, obtain the water area data to be identified; the water area data to be identified includes video data and / or image data.

[0066] Among them, the water area data to be recognized are the data of the water area to be recognized, including video data and image data. In actual application scenarios, these data mainly come from video monitoring devices arranged in key monitoring areas such as mountain flood ditches. These devices continuously shoot real-time pictures of the ditch, thus forming video data. At the same time, image data at specific moments or in specific scenarios can also be obtained by intercepting the video according to certain rules. For example, the automatic interception method of video pictures can be adopted to collect according to dimensions such as different basin sizes, river widths, different geomorphic conditions, and different time lights.

[0067] S204, input the water area data to be recognized into the pre-trained water area recognition model, and output the boundary range and water surface line range of the water area to be recognized through the water area recognition model.

[0068] Among them, the boundary range refers to the area enclosed by the demarcation line between the water area and the surrounding land or other non-water environments. In actual scenarios, it clearly defines the coverage range of water. For example, in a mountain flood channel, when a flood occurs, there is an obvious boundary between the area occupied by the flood and the areas such as riverbanks and slopes that are not flooded by water, and the range enclosed by this boundary is the boundary range. The water surface line range generally refers to the position where the water level is located on the water surface and the range involved in its changes. It reflects the height of the water level and its changing trend at different positions. In the channel, the water surface line may fluctuate with changes in water flow velocity and water volume. For example, during the rising process of a flood, the water surface line gradually rises and its range correspondingly expands; when the flood recedes, the water surface line drops and the range shrinks. The water area recognition model is obtained by performing resolution enhancement processing and image noise removal processing on various different types of river channel water area data and then performing data annotation, and training using the annotation results of the data annotation. Image noise removal processing can be to remove the interference of various noises on the image noise, for example, defogging, de-raining, removing reflections, removing ripples, and so on. The water area recognition model is usually an end-to-end model. By inputting the original image, the model calculates and outputs pixel-level prediction results, realizing the recognition of the water area boundary in natural scenarios, enabling 24-hour real-time monitoring of the water area range, being able to recognize the water area boundary range, and automatically analyzing the change of the water surface line range. The water area recognition model includes: an encoder network and a decoder network. The encoder network includes: a backbone network for extracting deep semantic features and an atrous spatial pyramid pooling layer for performing atrous convolution. The decoder network includes: a low-level feature decoding layer for decoding the features output by the backbone network and a deep-level feature decoding layer for decoding the features output by the atrous spatial pyramid pooling layer. The encoder network is responsible for extracting high-level deep semantic features of the original image, for extracting effective features of different categories. The deep semantic features are beneficial to the image pixel classification and reconstruction of the subsequent decoder. The encoder part consists of a backbone network and an ASPP (atrous spatial pyramid pooling) layer. The backbone network uses an efficient convolutional neural network to achieve the extraction of deep semantic features. The depthwise separable convolution adopted by the convolutional neural network can effectively reduce the computational amount and the number of model parameters. The backbone network performs deep feature extraction on the original image. In order to perform fine reconstruction work on the edge part, the ASPP layer is used to extract effective edge features. The ASPP uses atrous convolution to extract a deep feature map. Atrous convolution is a convolution that can increase the receptive field of features. By adjusting the rate of atrous convolution, different sizes of convolutional receptive fields can be adjusted. Such as Figure 3As shown, several are schematic diagrams of dilated convolutions corresponding to rate = 1, 2, and 4 respectively. The gray shaded part represents the receptive field, and the dot part represents the convolutional branches with weight parameters. Dilated convolution can extract features in a larger range. In the ASPP layer, a global feature map pooling operation is used to generate global features. Feature decoder network: The decoder network is responsible for reconstructing the encoded image features to the original image size and generating a predicted pixel-level classification result map. The decoder is mainly composed of a low-level feature decoding layer and a deep-level feature decoding layer, and finally combines the two decoded feature maps to generate a prediction result map. The feature map of the low-level feature decoding comes from the intermediate feature map of the backbone network of the encoder, and 1×1 convolution is used to extract low-level features and reduce the channel dimension of the output feature map. The features of the deep-level feature decoding come from the output feature map of the ASPP layer of the encoder, and the feature output of the encoder is enlarged in size through an upsampling operation. The features of the low-level feature decoding and the deep-level decoding are combined into a decoder feature map through a channel-level concatenation operation. The decoded feature map is further processed by a 3×3 convolution to extract effective features, and finally, an upsampling operation is used to restore the feature map to the original image size, and finally, a prediction result map is output.

[0069] Specifically, first, for video data, the real-time images of the corresponding video monitoring points of mountain flood ditches can be retrieved. These video data will extract key frames at a certain frame rate and convert them into an image sequence because the model usually processes based on images. For image data, if it is obtained during the sample data collection phase, it needs to go through processes such as dataset management and data annotation to ensure the quality and usability of the data. Before inputting the data into the model, it also needs to be standardized, such as unifying the image size, adjusting the pixel value range, etc., to make it meet the input requirements of the model. Then, the processed data is sent to the server where the model is located through network transmission. The water area recognition model is based on AI image recognition technology, such as using architectures like convolutional neural networks (CNNs). When receiving the input data, the model processes it according to its internal calculation process. In the front-end deployment link of the model, the trained model is evaluated and deployed to the front end to be able to quickly respond to data input. The encoder network in the model is responsible for extracting high-level deep semantic features of the original image. For example, for the input water area image, it can identify feature information related to the water area such as color, texture, shape, etc. Through continuous operations such as convolution and pooling, more abstract and representative features are gradually extracted. Then, the feature decoder network reconstructs the encoded image features to the original image size and generates a predicted pixel-level classification result map. In this process, the model uses a large number of feature patterns of water areas and non-water areas learned during the training phase to classify and judge each pixel of the input image to determine whether it belongs to the water area category. After the calculation and analysis of the model, the boundary range of the water area to be identified and the water surface line range are finally output. The output results are usually presented in the form of image annotations or data files. For the boundary range, the model will mark the boundary line between the water area and the surrounding environment on the image with specific lines or colors, and determine which continuous pixel areas belong to the water area boundary by analyzing the pixel-level classification results. For the water surface line range, the model will give the position and approximate change range of the water surface line on the image or in the data file according to the relative position of the water level in the image and the depth information of the water area.

[0070] In the above-mentioned water area recognition method, the water area recognition model can enhance the resolution of various types of river channel water area data, making the water area features that may have been blurred due to low resolution clearer, so that the model can more accurately identify and distinguish the water area and non-water area parts, and accurately determine the boundary range and water surface line range. For example, in the monitoring of small river channels in some remote mountainous areas, after processing the low-resolution video or image data, the model can better capture the subtle water flow changes and the blurred transition areas of the water area boundary, reducing misjudgments. The image noise removal process effectively reduces the impact of noise generated by environmental factors (such as bad weather, light interference, etc.) on the image. In actual scenarios, rain, fog, reflection, etc. may all cause image noise. After processing, the model can focus on the real water area features, further improving the recognition accuracy and ensuring reliable results can be stably output under complex environments. Training with various types of river channel water area data, covering samples in different dimensions such as different basin sizes, river channel widths, geomorphic conditions, and time and light, enables the model to learn rich and diverse water area feature patterns, thereby enhancing its generalization ability in different actual scenarios. Whether it is a wide river or a narrow mountain stream, whether it is a water area under sunny or bad weather conditions, the model can more effectively identify and analyze, with a wider scope of application, reducing the over-reliance on specific scenarios and improving the practicality and reliability of the system. The backbone network in the encoder network uses an efficient convolutional neural network for deep semantic feature extraction, and uses depthwise separable convolution to reduce the computational amount and the number of model parameters, improving the computational efficiency while ensuring the model performance. The atrous spatial pyramid pooling layer performs atrous convolution to increase the feature receptive field, which helps to capture feature information in a wider area and better understand the overall structure and change trend of the water area. The underlying feature decoding layer and the deep feature decoding layer in the decoder network decode and combine features from different sources respectively, which can reconstruct image features more precisely, improve the accuracy of pixel-level classification results, and thus more accurately output the boundary range and water surface line range of the water area.

[0071] In one embodiment, as Figure 4 shown, the method further includes:

[0072] S302, obtaining various types of river channel water area data.

[0073] Among them, various types of river channel water area data refer to image data related to river channel water areas with various different characteristics, such as river channel water area images in different seasons, different time periods (daytime, night), different weather conditions (sunny, rainy, foggy, etc.), different shooting angles, and different resolutions. The diversity of these data helps to comprehensively analyze and process various situations related to river channel water areas.

[0074] Specifically, since there are various river flood patterns and it is difficult to achieve general annotation, the sample data of the water area recognition model should cover different types of river water areas as much as possible, and sample data collection should be carried out according to dimensions such as different basin sizes, river widths, different geomorphic conditions, and different time lights. The sample data collection can adopt the method of automatic video and picture interception, regularly and automatically obtain the sample data sources of different rivers at different times, and conduct data classification management through manual integration.

[0075] S304, use resolution enhancement technology to perform resolution enhancement processing on the different types of river water area data with a resolution less than the preset resolution threshold.

[0076] Among them, resolution enhancement is for some images with low resolution. Through technical means, the resolution is enhanced, and the information required in the image can be intelligently identified.

[0077] Specifically, some resolution enhancement algorithms can be used, such as interpolation-based methods, deep learning-based methods, multi-scale analysis methods, etc., to perform resolution enhancement processing on different types of river water area data with a smaller resolution (less than the resolution threshold).

[0078] S306, perform image denoising processing on the different types of river water area data after resolution enhancement processing and those without resolution enhancement processing to obtain the river water area data to be annotated.

[0079] Among them, image denoising processing can include: processing methods such as rain removal, fog removal, and reflection removal that remove the impact on image quality.

[0080] Specifically, appropriate image denoising methods can be selected, such as mean filtering, median filtering, wavelet transform, etc., to perform image denoising processing on the different types of river water area data after resolution enhancement processing and those without resolution enhancement processing to obtain the river water area data to be annotated.

[0081] S308, perform water area boundary annotation and water surface line range annotation on the river water area data to be annotated.

[0082] Specifically, edge detection algorithms can be used to perform water area boundary annotation and water surface line range annotation on the river water area data to be annotated.

[0083] In some exemplary embodiments, for the river channel water area data to be annotated, data annotation work can be carried out through a data annotation tool. The data annotation work should support selective annotation work on indicators such as image acquisition time, affiliated project, measuring station, and channel. At the same time, the user can mark the image quality (normal, blurred, insufficient light, damaged) and set the calibration type (image classification, object detection, semantic segmentation, key point detection) during the annotation process. The result file after the image is annotated is stored in the corresponding dataset of the project model, and a function for viewing and verifying the annotated data is provided to effectively ensure the quality of data annotation.

[0084] S310. Train a neural network model based on the annotation results of the water area boundary annotation and the water surface line range annotation in the river channel water area data to be annotated to obtain a water area recognition model.

[0085] Specifically, a convolutional neural network (CNN) is used for training the water area recognition model. It is an abstraction of the way the human brain neurons think. Through learning a large number of classified images, CNN can automatically extract image features to achieve image classification and object recognition. By defining a CNN network model with up to dozens of layers, an accurate defect detection rate can be achieved. The dataset formed by the annotation results of the water area boundary annotation and the water surface line range annotation in the river channel water area data to be annotated is divided into multiple subsets, including a training set, a validation set, and a test set. For CNN, data segmentation helps ensure that the model can learn and evaluate in multiple stages, thereby improving the generalization ability of the model. When performing data segmentation, the cross-validation method is used. Ensuring the randomness and representativeness during data segmentation helps the trained CNN model have better performance and stability in practical applications. First, the input data is calculated through convolutional layers, pooling layers, and fully connected layers, etc. The final prediction value is obtained. These calculations include convolutional operations, the application of activation functions, pooling, etc., which can effectively extract the spatial features in the data. Then, the gradient of the loss with respect to the parameters (weights, biases) of each layer is calculated through the backpropagation algorithm, and optimization algorithms such as the gradient descent method are used to adjust the parameters. Backpropagation includes calculating the gradient of each neuron and updating the network weights. Hyperparameters include learning rate, batch size, network structure (such as convolutional kernel size, number of layers, etc.), regularization techniques (such as Dropout, L2 regularization), etc. The selection of these hyperparameters will directly affect the training efficiency and generalization ability of the model. Hyperparameter optimization is carried out through methods such as grid search, random search, and Bayesian optimization. Through these methods, the best combination of hyperparameters can be found within a reasonable search space to improve the final performance of the CNN model. Finally, the water area recognition model is obtained after training is completed.

[0086] In one embodiment, the use of the resolution enhancement technology to perform resolution enhancement processing on the different types of river channel water area data with a resolution less than a preset resolution threshold includes:

[0087] Use a super-resolution reconstruction algorithm to perform resolution enhancement processing on the different types of river channel water area data with a resolution less than a preset resolution threshold.

[0088] Specifically, since the water area river channel data involves multiple elements, river channel images usually contain rich details, such as the vegetation texture on the riverbank, the ripples on the water surface, the stones in the river, etc. Therefore, a super-resolution reconstruction algorithm based on deep learning (such as SRGAN, ESRGAN) can be used to perform resolution enhancement processing on the different types of river channel water area data with a resolution less than a preset resolution threshold.

[0089] In one embodiment, both SRGAN (Super-Resolution Generative Adversarial Network) and ESRGAN (Enhanced Super-Resolution Generative Adversarial Network) are deep learning-based methods. They are trained with a large number of low-resolution and high-resolution image pairs and can learn the complex mapping relationship from low-resolution to high-resolution images. During the training process, the network will automatically learn how to recover the missing detail information in the low-resolution images. When processing river channel water area data, these algorithms can learn details such as the texture of the water flow in the river channel, the vegetation details on the riverbank, the shape of the stones on the riverbed, and the ripples on the water surface. These details may be blurred or invisible in low-resolution images due to pixel loss. For river channel water area images, they can generate clearer and more natural high-resolution images, enabling observers to see the detail information in the river channel more clearly. In addition, such algorithms can include multiple convolutional layers in the network structure, and different convolutional layers can learn features at different scales. For example, shallow convolutional layers can learn local detail information of the image, such as local texture, small objects (such as floating objects in the water); deep convolutional layers can learn more abstract and global information, such as the outline of the entire river channel, the shape of large water areas, etc.

[0090] In one embodiment, the image denoising processing includes: image de-raining and image de-fogging processing. Image de-raining and image de-fogging processing are two image restoration techniques for images affected by bad weather, aiming to eliminate the interference of rain and fog on image quality and improve image clarity and visual effects. As Figure 5 shown, perform image denoising processing on the different types of river channel water area data after resolution enhancement processing and before resolution enhancement processing to obtain the river channel water area data to be labeled, including one or more of the following:

[0091] S402, use the global histogram equalization algorithm to perform image de-raining and image de-fogging processing on the different types of river channel water area data after resolution enhancement processing and before resolution enhancement processing.

[0092] Among them, the Global Histogram Equalization algorithm is an image enhancement technology used in the field of image processing, which aims to improve the contrast of the image and make the image details clearer and more discernible.

[0093] Specifically, a global histogram equalization algorithm can be used to transform the histogram of foggy and / or rainy images into a form of approximately uniform distribution, thereby increasing the dynamic range of pixel grayscale values, thereby achieving the effect of enhancing the overall contrast of foggy and / or rainy images, thereby achieving defogging and / or rain removal.

[0094] Furthermore, on rainy days, the global histogram equalization algorithm is first applied. This step can enhance the overall contrast of the image, making the difference between raindrops and the background more obvious, which is convenient for subsequent processing. For example, the river image that was originally gray and blurred in detail due to rainfall may be presented as clearer bright spots or dark spots after histogram equalization, providing a better basis for subsequent raindrop detection and removal. Morphological operations (such as corrosion, dilation, opening operation, and closing operation) are used to detect and remove raindrops. After the contrast is enhanced by histogram equalization, the features of raindrops in the image are easier to identify. For example, the opening operation can be used to remove small raindrop highlights, and the shape of the background object can be restored by closing operation. Since the histogram equalization increases the difference between raindrops and the background, the morphological operation can more accurately separate raindrops from the background. Select appropriate filters, such as Gaussian filtering, median filtering, etc. The raindrop features are more prominent in the image after histogram equalization. Gaussian filtering can smooth the image to a certain extent and reduce the influence of raindrop noise; median filtering can effectively remove raindrops similar to salt and pepper noise.

[0095] In foggy days, the global histogram equalization algorithm is applied to the fog image of the river water area (regardless of whether it has been resolution enhanced) to increase the image contrast. This helps to calculate the dark channel more accurately. In foggy images, the overall image contrast is low due to the influence of fog, which may lead to inaccurate calculation of the dark channel. After histogram equalization, the low grayscale values ​​in the non-sky area are more obvious, and the dark channel can be calculated more accurately. Based on the dark channel prior algorithm, the transmittance is estimated using the dark channel with enhanced contrast. Since histogram equalization makes the difference between the scenery and the fog in the image more prominent, the transmittance estimation is more accurate. Combined with the atmospheric light value, the fog-free image is restored through the foggy image degradation model. For example, in the river scene, the details such as the river bank and water surface obscured by fog can be restored more clearly.

[0096] S404, using a homomorphic filtering algorithm to perform image deraining and image defogging processing on different types of river water area data that have been subjected to resolution enhancement processing and those that have not been subjected to resolution enhancement processing.

[0097] Among them, the homomorphic filtering algorithm is a technique for processing images in the frequency domain. Further, it is an image enhancement processing method that combines frequency filtering and gray-scale transformation. It is also a processing technique that uses the illumination reflection model as the basis for frequency domain processing and improves image quality by compressing the brightness range and enhancing contrast. It is mainly used to enhance the contrast of images and simultaneously improve the brightness and details of images, and is particularly suitable for processing images with uneven illumination.

[0098] Specifically, for the river channel water area rain image without resolution enhancement, homomorphic filtering can be used as a preprocessing step. Since rainfall may cause uneven overall illumination of the image, with some areas being too bright or too dark, homomorphic filtering corrects the uneven illumination by suppressing the low-frequency illumination component and simultaneously enhances the high-frequency reflection component, highlighting the detail differences between raindrops and the background. For example, raindrops often appear as local brightness changes in the image, and homomorphic filtering can enhance these local features, making raindrops easier to be recognized by subsequent algorithms.

[0099] For the river channel water area rain image that has undergone resolution enhancement processing, although the resolution has been improved, there may still be problems with illumination and contrast. Homomorphic filtering can further optimize the image quality and enhance the contrast between raindrops and the background, providing a clearer image basis for the rain removal operation. For example, some deep learning-based rain removal algorithms may be sensitive to the illumination and contrast of the image. After preprocessing with homomorphic filtering, the rain removal effect of these algorithms can be improved. After homomorphic filtering enhances the image features, the subsequent methods for rain removal and fog removal can be the same as in step S402 above, and will not be repeated here.

[0100] S406, Use recursive Gaussian filtering to accelerate the Retinex algorithm and use linear stretching to perform image rain removal and image fog removal on different types of river channel water area data with and without resolution enhancement processing.

[0101] Among them, Retinex is a model that describes color constancy. It has the characteristics of dynamic range compression and color constancy, and has a good enhancement effect on low-contrast color images caused by uneven illumination. The Retinex algorithm aims to separate the reflection component to achieve image enhancement because the reflection component contains the inherent attribute information of the object and is not affected by illumination changes. Recursive Gaussian filtering is a method for quickly implementing Gaussian filtering. It calculates the filtering result recursively, greatly reducing the computational amount. In the Retinex algorithm, it is usually necessary to perform Gaussian filtering on the image multiple times to estimate the illumination component. Using recursive Gaussian filtering to replace traditional Gaussian filtering can significantly reduce the calculation time.

[0102] Specifically, the Retinex algorithm can be accelerated by using recursive Gaussian filtering, and the contrast of the image can be enhanced by using linear stretching. Then, rain and fog removal can be performed in the same way as in step S402 described above.

[0103] In some exemplary embodiments, rainfall can reduce the contrast of the river water area image, and some areas are too bright or too dark. For such an image, the Retinex algorithm is accelerated by using recursive Gaussian filtering and linear stretching is performed. First, the gray level range of the image is statistically analyzed. For example, in some rainfall images, the gray levels may be concentrated between [50, 100], while the entire gray level range is [0, 255]. Through linear stretching, it is mapped to [0, 255], so that the bright and dark parts of the image can be better displayed. Then, rain and fog removal are performed in the same way as in step S402 described above.

[0104] S408, perform image rain removal and image fog removal on different types of river water area data that have undergone resolution enhancement processing and those that have not undergone resolution enhancement processing by using curvelet transform.

[0105] Among them, the curvelet transform is a multi-scale geometric analysis tool. Curvelets are a new multi-scale analysis method developed on the basis of wavelet transform. Since it is particularly suitable for signal processing of anisotropic singularity features, it can well make up for the limitations of wavelet transform in enhancing the curve edges of images.

[0106] Specifically, perform wavelet transform on the river channel water area rain image (including different types of river channel water area data after resolution enhancement processing and without resolution enhancement processing). Based on its multi-scale and multi-directional characteristics, decompose the image into wavelet coefficients of different scales and directions. Small raindrops are similar to small, local singularities in the image, and large raindrops or rain lines may present curve characteristics. Wavelet transform can capture these characteristics at different scales. Small-scale coefficients reflect raindrop details, and large-scale coefficients present the overall rainfall distribution. Rainfall in the image is equivalent to noise, interfering with the original information of the river channel. Due to the sparse representation of the image by wavelet transform, rainfall noise is usually reflected in some wavelet coefficients. By analyzing the coefficient distribution and setting an appropriate threshold, the coefficients representing rainfall noise can be separated from the coefficients representing the river channel scene. For example, using the hard threshold method, coefficients less than the threshold are set to 0, and coefficients greater than the threshold are retained. These retained coefficients mainly contain river channel scene information. Perform inverse wavelet transform on the processed wavelet coefficients to reconstruct the image after removing rainfall noise. This process restores the original scene of the river channel water area image, making information such as the river channel and riverbank clearer. Although there may be some detail losses, overall, rainfall interference is removed. Fog reduces the overall contrast of the image and blurs details. Decompose the fog image into coefficients of different scales and directions through wavelet transform. At a large scale, the wavelet coefficients reflect the overall distribution of fog, such as the approximate range of areas with high fog concentration; at a small scale, capture the subtle changes at the boundary between fog and the scenery. Coefficients in different directions can detect the influence of fog on the scenery in different directions, such as the blurring effect of fog on the riverbank line in the horizontal direction. Analyze the statistical characteristics of the wavelet coefficients. The coefficients in the fog-affected area often have a specific distribution. For example, at certain scales and directions, the coefficient values may be relatively concentrated and small. By adjusting these coefficients, such as enhancing the coefficients related to contrast and suppressing the coefficients representing fog interference, the influence of fog on the image can be changed. The wavelet coefficients can be weighted according to statistical quantities such as the mean and variance of the coefficients to highlight the scenery features and weaken the fog influence. The adjusted wavelet coefficients are reconstructed into an image through inverse wavelet transform to restore the clarity of the river channel water area scenery. The reconstructed image can show information such as the river channel contour and water surface details more clearly.

[0107] In this embodiment, rain and fog are removed in multiple ways, which can ensure the accuracy and clarity of the image.

[0108] In one embodiment, the image processing includes: image de-reflection and ripple processing, such as Figure 6 shown, perform image denoising processing on different types of river channel water area data after resolution enhancement processing and without resolution enhancement processing to obtain river channel water area data to be labeled, including:

[0109] S502, use histogram equalization image enhancement technology and the deeplabv3+ algorithm to perform deconstruction analysis on different types of river water area data that have been resolution-enhanced and those that have not been resolution-enhanced from multiple scales, and perform image reflection and ripple removal processing.

[0110] Among them, histogram equalization image enhancement technology is usually a technology that redistributes the gray values of an image to make the gray-level distribution of the image more uniform, thereby expanding the gray dynamic range of the image and enhancing the image contrast. Deeplabv3+ algorithm: This is a deep learning algorithm for semantic segmentation. Based on the encoder-decoder architecture and combined with technologies such as dilated convolution and multi-scale feature fusion, it can effectively segment different object categories in the image. In the processing of river water area data, it can be used to identify and segment different regions such as water areas, reflections, and ripples. Deconstruction analysis: Here it refers to analyzing the image from multiple scales, decomposing the image into different levels of features to comprehensively understand the content of the image. For example, observing a river image at different scales, the large scale can grasp the overall water area range and layout, and the small scale can focus on the detailed features of ripples and reflections. Image reflection and ripple removal processing: Aims to remove the reflections formed due to the water surface reflection and the ripples generated by the water surface fluctuation in the image, so that the image can more clearly present the real scene of the river channel, facilitating subsequent analysis and processing.

[0111] Specifically, in image recognition analysis applications, situations such as image reflection, having reflections or water surface ripples often occur, and these situations will all affect the image recognition effect. Therefore, it is necessary to remove the influencing problems such as reflections and ripples through technical means. Based on histogram equalization image enhancement technology and the deeplabv3+ semantic segmentation algorithm, fully excavate the context content information at multiple scales, gradually reconstruct the spatial information to better capture the water area boundary, and conduct targeted training on a large number of reflection, reflection, and ripple regions, which can effectively identify the water area range in complex situations.

[0112] In some exemplary embodiments, for each river channel water area image, its grayscale histogram is calculated, and the frequency of each grayscale level is statistically counted. The cumulative distribution function (CDF) is calculated based on the histogram, and the original grayscale values are mapped to new grayscale values through the CDF to achieve the redistribution of grayscale levels, making the grayscale distribution of the image more uniform and enhancing the contrast of the image. This step helps subsequent algorithms better identify different regions in the image because the reflections and ripples may have a higher distinguishability from the background in the image with enhanced contrast. The image processed by histogram equalization is input into the encoder of deeplabv3+. The encoder uses a convolutional neural network (CNN) structure. Through a series of convolutional layers and pooling layers, the image is downsampled to gradually extract the high-level semantic features of the image. At the same time, the dilated convolution technique is used to expand the receptive field of the convolutional kernel without reducing the resolution, enabling the network to capture multi-scale context information. For example, when processing river channel images, large-scale information of the entire water area and local small-scale information of reflections and ripples can be obtained simultaneously. The deeplabv3+ algorithm uses dilated convolutions with different rates in the encoder to obtain feature maps of different scales. These feature maps contain information from fine-grained details to coarse-grained global information. Through a specific fusion method, such as concatenating or weighted summing feature maps of different scales, the multi-scale features are fused together, enabling the network to comprehensively consider various scale information for subsequent segmentation. The fused multi-scale features are input into the decoder. The decoder upsamples the feature maps through transposed convolutional layers or upsampling layers to restore the spatial resolution of the image to the same size as the original image. During this process, combined with the high-level semantic information passed from the encoder, each pixel is classified to determine whether it belongs to the water area, reflection, ripple, or other background categories. After obtaining the segmentation result of the deeplabv3+ algorithm, for the regions identified as reflections and ripples, various methods can be used for removal. For example, for the reflection region, it can be filled with surrounding background information according to features such as its symmetry relationship with real objects; for the ripple region, methods such as smoothing filtering can be used to make it more integrated with the surrounding water surface region, thereby achieving the processing of removing reflections and ripples from the image.

[0113] And / or, S504, use the pre-trained reflection and ripple recognition model to identify reflections and ripples in different types of river channel water area data after and without resolution enhancement processing, and remove the reflections and ripples.

[0114] Specifically, collect a large number of image data of river channel waters of different types, including images under different seasons, weather, lighting conditions, and shooting angles. These images should include both those after resolution enhancement processing and those without resolution enhancement processing. For the collected images, manually label the reflection and ripple areas. This step is for training the reflection and ripple recognition model, and the accuracy of the labeling directly affects the performance of the model. Professional image annotation tools such as LabelMe and VGG ImageAnnotator can be used for annotation. Select a model suitable for reflection and ripple recognition, such as a semantic segmentation model based on convolutional neural network (CNN), like U-Net, SegNet, deeplabv3+, etc. These models perform excellently in image segmentation tasks and can effectively identify different regions in the image. Divide the labeled image data into a training set, a validation set, and a test set. Usually, the training set is used for parameter learning of the model, the validation set is used for adjusting the hyperparameters of the model, and the test set is used for evaluating the final performance of the model. Generally, it is divided according to the ratio of 70%, 15%, and 15%. Use the training set data to train the selected model. During the training process, the model will learn the feature patterns of reflections and ripples, and continuously adjust the weights of the model through the backpropagation algorithm to minimize the difference between the predicted results of the model and the true labeled results. Commonly used loss functions include cross-entropy loss function, etc. Appropriate hyperparameters such as learning rate, number of iterations, batch size, etc. need to be set during the training process to ensure that the model can converge to a better solution. Load the pre-trained reflection and ripple recognition model into memory to prepare for processing river channel water data.

[0115] Data preprocessing: For the river channel water area image data to be processed, whether it has been subjected to resolution enhancement processing or not, preprocessing is required. The preprocessing steps usually include image normalization (scaling the pixel values of the image to the interval [0, 1] or [-1, 1]), resizing the image to the size required by the model input, etc. These preprocessing operations can enable the model to better process image data of different specifications. The preprocessed image is input into the loaded model, and the model will output the probability that each pixel belongs to the categories of reflection, ripple, water area, or other background. By setting an appropriate threshold (such as 0.5), the probability values are converted into category labels to determine which areas in the image are reflections and which areas are ripples. Image inpainting algorithms can be used, such as the PatchMatch-based image inpainting method. The basic idea of this method is to find image patches similar to the reflection area in the non-reflection area of the image, and then fill the reflection area with these similar image patches. When operating specifically, for each pixel point in the reflection area, search for similar image patches in the surrounding non-reflection area, and select the most similar image patch for filling according to certain similarity measurement criteria (such as color, texture, etc.). Image inpainting algorithms can also be used. Since the ripple area is relatively small and the texture is relatively regular, the features of the ripple area can be extracted first, and then similar texture patterns can be found in the relatively smooth water area around for filling. Some deep learning-based image inpainting models can also be used, such as the Context Encoder, which encodes the image into a feature vector through an encoder, and then the decoder generates the inpainted image area according to the surrounding context information.

[0116] In this embodiment, by removing reflections and ripples in different ways, the accuracy of water area recognition can be guaranteed.

[0117] In one embodiment, the method further includes:

[0118] In response to the river channel water area data of multiple different types including the river channel water area data in foggy weather, perform image enhancement processing on the river channel water area data in foggy weather based on the atmospheric modulation transfer function.

[0119] Among them, the atmospheric modulation transfer function describes the impact of the atmosphere on the transmission of image information and reflects the attenuation degree of the atmosphere on signals with different spatial frequencies. The spatial frequency can be understood as the speed of gray-scale change in an image. For example, the fine texture part in an image corresponds to a higher spatial frequency, while the large-area uniform region corresponds to a lower spatial frequency. The atmospheric modulation transfer function can quantify the modulation effect of the atmosphere on different frequency components in an image, thereby helping us understand and compensate for the degradation of image quality caused by the atmosphere. The principle of this method is as follows: First, through the prediction of the atmospheric modulation transfer function, the degradation process of the atmosphere on image quality is approximately estimated. When prior information is obtained, the corresponding turbulent modulation transfer function and aerosol modulation transfer function are calculated through the prediction formula, and then the total atmospheric modulation transfer function is obtained by multiplying the former two. Then, the weather-degraded image is restored in the frequency domain using the atmospheric modulation transfer function, and the attenuation caused by the atmospheric modulation transfer function in the outdoor scene image is compensated.

[0120] Specifically, first, image data taken under foggy weather conditions are screened out from a large number of various types of river channel water area data. These data may come from different monitoring devices or collection projects, and there may be differences in formats, resolutions, etc. Therefore, it is necessary to perform unified preprocessing on them. For example, the images are adjusted to the same resolution, converted to a unified image format (such as the common RGB format), and the pixel values are normalized to make their ranges unified for subsequent algorithm processing. To perform image enhancement based on the atmospheric modulation transfer function, it is necessary to first estimate this function, which requires using some prior knowledge about atmospheric characteristics and information about the image itself. The influence of the atmosphere on different spatial frequencies can be estimated based on the statistical characteristics of the image. For example, by analyzing the gray-scale changes in different regions of the foggy image and combining the physical models of atmospheric scattering and absorption, the atmospheric modulation transfer function can be approximately calculated. Additionally, some known atmospheric parameters (such as visibility, humidity, etc.) can be used to assist in the estimation. If there is relevant meteorological data support, a relationship model between atmospheric parameters and the atmospheric modulation transfer function can also be established to obtain this function more accurately. The foggy river channel water area image is processed in the frequency domain. First, the image is transformed from the spatial domain to the frequency domain, and the commonly used method is through Fourier transform. In the frequency domain, different frequency components correspond to different features of the image. The low-frequency components mainly represent the general outline and background information of the image, while the high-frequency components are related to the details of the image. According to the estimated atmospheric modulation transfer function, the frequency domain image is corrected. Since the atmosphere attenuates the high-frequency components severely, resulting in blurred details in the foggy image, appropriate enhancement operations can be performed on the high-frequency part to compensate for the attenuation caused by the atmosphere. For example, the frequency domain image is multiplied point by point with the inverse function of the atmospheric modulation transfer function (or a properly adjusted function to avoid over-enhancement), so that the high-frequency components are relatively enhanced, and the low-frequency components remain basically unchanged or are moderately adjusted, thereby improving the overall frequency distribution of the image. After completing the frequency domain processing, the image is transformed back to the spatial domain through inverse Fourier transform to obtain the enhanced foggy river channel water area image. At this time, the contrast and details of the image will be improved to a certain extent, becoming clearer, which is beneficial for subsequent analysis and processing of the river channel water area.

[0121] In this embodiment, the fog reduces the overall contrast of the image and blurs the boundaries between scenes. Through the processing based on the atmospheric modulation transfer function, the frequency distribution of the image can be adjusted, the high-frequency part can be enhanced, thereby highlighting the boundaries and details between different objects in the image and improving the contrast of the image. The attenuation of the high-frequency information by the atmosphere causes many details in the foggy image to be lost. This processing method compensates for the high-frequency components and restores the details covered by the fog, such as small stones and floating objects in the river channel, making the image clearer and richer, presenting more visual information.

[0122] In some embodiments, after determining the boundary range and water surface line range of the water area to be recognized, early warning recognition can be carried out, calibrating the safe water area range for specific locations, and combining the corresponding threat levels after exceeding the safe water area range, which are used together as the judgment basis for channel flood early warning. Record information such as video images, water area range, and early warning range each time the water area range reaches or exceeds the early warning range. Calibrate the threatened water area range and set corresponding danger levels, manage them as early warning rules, and provide corresponding maintenance functions such as addition, modification, and deletion. When the water area range exceeds or reaches the range defined by the early warning rules, an early warning is automatically generated and recorded, providing retrieval and detailed viewing of the early warning records. After triggering the early warning, in addition to prominently displaying it on the monitoring interface, the early warning information should also be published to the corresponding business departments and personnel through various methods and channels such as business system notifications and prompts, and external information push. Additionally, the real-time video image of the water area to be recognized can be displayed. The current water area range is displayed, and the real-time water area range information recognized by the recognition model is presented in fusion with the real-time video. Early warning prompts, prominently display and strikingly prompt the water area stations currently in the early warning state.

[0123] In one embodiment, the present disclosure also provides a method for training a water area recognition model, including: obtaining various different types of river channel water area data;

[0124] Using resolution enhancement technology to perform resolution enhancement processing on the different types of river channel water area data with a resolution less than a preset resolution threshold;

[0125] Performing image denoising processing on the different types of river channel water area data after resolution enhancement processing and those without resolution enhancement processing to obtain river channel water area data to be labeled;

[0126] Performing water area boundary annotation and water surface line range annotation on the river channel water area data to be labeled;

[0127] Training a neural network model according to the annotation results of water area boundary annotation and water surface line range annotation in the river channel water area data to be labeled to obtain a water area recognition model.

[0128] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0129] Based on the same inventive concept, an embodiment of the present disclosure further provides a water area range recognition device for implementing the water area range recognition method described above. The solution provided by this device to solve the problem is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the water area range recognition device provided below can refer to the limitations on the water area range recognition method in the above text, and will not be repeated here.

[0130] In one embodiment, as Figure 7 shown, a water area range recognition device 600 is provided, including: a data acquisition module 602 and a model processing module 604, where:

[0131] The data acquisition module 602 is used to acquire water area data to be recognized; the water area data to be recognized includes video data and / or image data;

[0132] The model processing module 604 is used to input the water area data to be recognized into a pre-trained water area recognition model, and output the boundary range and water surface line range of the water area to be recognized through the water area recognition model;

[0133] Among them, the water area recognition model is obtained by performing resolution enhancement processing and image noise removal processing on various different types of river water area data and then performing data annotation, and training with the annotation results of the data annotation;

[0134] The water area recognition model includes: an encoder network and a decoder network. The encoder network includes: a backbone network for extracting deep semantic features and an atrous spatial pyramid pooling layer for performing atrous convolution. The decoder network includes: a low-level feature decoding layer for decoding the features output by the backbone network and a deep-level feature decoding layer for decoding the features output by the atrous spatial pyramid pooling layer.

[0135] In one embodiment of the device, the device further includes:

[0136] A training data acquisition module for acquiring various different types of river water area data.

[0137] A resolution enhancement module for performing resolution enhancement processing on the different types of river water area data with a resolution less than a preset resolution threshold by using resolution enhancement technology.

[0138] An image denoising module for performing image denoising processing on the different types of river water area data after resolution enhancement processing and before resolution enhancement processing to obtain river water area data to be labeled.

[0139] A labeling module for performing water area boundary labeling and water surface line range labeling on the river water area data to be labeled.

[0140] A training module for training a neural network model according to the labeling results of water area boundary labeling and water surface line range labeling in the river water area data to be labeled to obtain a water area recognition model.

[0141] In an embodiment of the device, the resolution enhancement module is further configured to perform resolution enhancement processing on the different types of river water area data with a resolution less than a preset resolution threshold by using a super-resolution reconstruction algorithm.

[0142] In an embodiment of the device, the image denoising processing includes: image de-raining and image de-fogging processing. The denoising processing module includes: a de-raining and de-fogging module for performing image de-raining and image de-fogging processing on the different types of river water area data after resolution enhancement processing and before resolution enhancement processing by using a global histogram equalization algorithm;

[0143] Performing image de-raining and image de-fogging processing on the different types of river water area data after resolution enhancement processing and before resolution enhancement processing by using a homomorphic filtering algorithm;

[0144] Accelerating the Retinex algorithm by using recursive Gaussian filtering and performing image de-raining and image de-fogging processing on the different types of river water area data after resolution enhancement processing and before resolution enhancement processing by using a linear stretching method;

[0145] Performing image de-raining and image de-fogging processing on the different types of river water area data after resolution enhancement processing and before resolution enhancement processing by using curvelet transform.

[0146] In one embodiment of the device, the image processing includes: removing reflections and ripples from the image. The denoising processing module includes: an image reflection and ripple removal processing model, which uses histogram equalization image enhancement technology and the deeplabv3+ algorithm to perform deconstruction analysis on different types of river water area data after resolution enhancement processing and without resolution enhancement processing from multiple scales, and perform image reflection and ripple removal processing; and / or, use a pre-trained reflection and ripple recognition model to identify reflections and ripples in different types of river water area data after resolution enhancement processing and without resolution enhancement processing, and remove the reflections and ripples.

[0147] In one embodiment of the device, the rain and fog removal module includes: an image enhancement module, which is used to respond to the river water area data in foggy weather included in the multiple different types of river water area data, and perform image enhancement processing on the river water area data in foggy weather based on the atmospheric modulation transfer function.

[0148] In one embodiment, as Figure 8 shown, the present disclosure also provides a water area recognition model training device 700, and the device includes:

[0149] A training data acquisition module 702, which is used to acquire multiple different types of river water area data;

[0150] A resolution enhancement module 704, which is used to perform resolution enhancement processing on the different types of river water area data with a resolution less than a preset resolution threshold by using resolution enhancement technology;

[0151] An image denoising module 706, which is used to perform image denoising processing on different types of river water area data after resolution enhancement processing and without resolution enhancement processing to obtain river water area data to be labeled;

[0152] A labeling module 708, which is used to perform water area boundary labeling and water surface line range labeling on the river water area data to be labeled;

[0153] A training module 710, which is used to train a neural network model according to the labeling results of water area boundary labeling and water surface line range labeling in the river water area data to be labeled to obtain a water area recognition model.

[0154] Each module in the above water area range recognition device can be implemented in whole or in part by software, hardware, and their combination. Each of the above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above each module.

[0155] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structural diagram may be as shown in Figure 9 . The computer device includes a processor, a memory, and a network interface 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, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store water area data to be recognized. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a method for identifying the water area range.

[0156] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structural diagram may be as shown in Figure 10 . The computer device includes a processor, a memory, a communication 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 communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner. The wireless manner can be implemented through WIFI, a mobile cellular network, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for identifying the water area range. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device may be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the computer device housing, or an external keyboard, a touchpad, or a mouse, etc.

[0157] Those skilled in the art can understand that Figure 9 the structures shown in or 10 are merely block diagrams of some structures related to the solution of the present disclosure, and do not constitute a limitation on the computer device to which the solution of the present disclosure is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different component layout.

[0158] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0159] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0160] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0161] It should be noted that the water area data to be recognized involved in the present disclosure are all information and data authorized by the user or fully authorized by all parties.

[0162] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned method embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned method embodiments. Among them, any reference to a memory, a database, or other media used in the embodiments provided by the present disclosure can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tapes, floppy disks, flash memories, optical memories, high-density embedded non-volatile memories, resistive random access memories (ReRAM), magnetoresistive random access memories (MRAM), ferroelectric random access memories (FRAM), phase change memories (PCM), graphene memories, etc. Volatile memories can include random access memory (RAM) or external cache memories, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided by the present disclosure can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., and are not limited thereto. The processors involved in the embodiments provided by the present disclosure can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., and are not limited thereto.

[0163] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0164] The above-described embodiments merely represent several implementation manners of the present disclosure. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present disclosure. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present disclosure, several modifications and improvements can still be made, and these all belong to the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the appended claims.

Claims

1. A method for identifying a water area, characterized in that: The method comprises: Acquire water area data to be identified; the water area data to be identified includes video data and / or image data; The water area data to be identified is input into a pre-trained water area identification model, and the boundary range and water surface line range of the water area to be identified are output through the water area identification model. The water area recognition model is obtained by performing resolution enhancement processing and image noise removal processing on various types of river water area data and then annotating the data, and training using the annotation results of the data; The water area identification model includes: an encoder network and a decoder network, wherein the encoder network includes: a backbone network for extracting deep semantic features and a dilated pyramid pooling layer for performing dilated convolution, and the decoder network includes: a bottom feature decoding layer for decoding features output by the backbone network and a deep feature decoding layer for decoding features output by the dilated pyramid pooling layer.

2. The method according to claim 1, characterized in that The method further comprises: Obtain various types of river water data; Performing resolution enhancement processing on the different types of river water area data having a resolution less than a preset resolution threshold using a resolution enhancement technique; Perform image denoising on different types of river water area data after and without resolution enhancement processing to obtain river water area data to be labeled; Marking the water area boundary and water surface line range of the river water area data to be marked; The neural network model is trained according to the annotation results of water area boundary annotation and water surface line range annotation in the river water area data to be annotated to obtain the water area recognition model.

3. The method according to claim 2, characterized in that The method of performing resolution enhancement processing on the different types of river water area data having a resolution less than a preset resolution threshold by using the resolution enhancement technology includes: A super-resolution reconstruction algorithm is used to perform resolution enhancement processing on the different types of river water area data having a resolution less than a preset resolution threshold.

4. The method according to claim 2, characterized in that: The image denoising process includes: image rain removal and image defogging process. The image denoising process is performed on different types of river water area data after and without resolution enhancement process to obtain the river water area data to be labeled, including one or more of the following: The global histogram equalization algorithm is used to perform image deraining and image defogging on different types of river water area data with and without resolution enhancement processing. Homomorphic filtering algorithm is used to perform image deraining and image defogging on different types of river water area data with and without resolution enhancement processing. Retinex algorithm is accelerated by recursive Gaussian filtering and linear stretching is used to remove rain and fog from different types of river water data after or without resolution enhancement. Curvelet transform is used to perform image deraining and defogging on different types of river water area data with or without resolution enhancement.

5. The method according to claim 2, characterized in that: The image processing includes: image reflection removal and ripple removal, and image denoising is performed on different types of river water area data after resolution enhancement and without resolution enhancement to obtain river water area data to be labeled, including: Using histogram equalization image enhancement technology and deeplabv3+ algorithm, we deconstruct and analyze different types of river water data with and without resolution enhancement at multiple scales, and remove reflections and ripples from the images. And / or, using a pre-trained reflection and ripple recognition model to recognize reflections and ripples in different types of river water area data after and without resolution enhancement processing, and removing the reflections and ripples.

6. The method according to claim 2 or 4, characterized in that: The method further comprises: In response to the multiple different types of river water area data including river water area data under foggy weather, image enhancement processing is performed on the river water area data under foggy weather based on an atmospheric modulation transfer function.

7. A water area identification model training method, characterized in that: The method comprises: Obtain various types of river water data; Performing resolution enhancement processing on the different types of river water area data having a resolution less than a preset resolution threshold using a resolution enhancement technique; Perform image denoising on different types of river water area data after and without resolution enhancement processing to obtain river water area data to be labeled; Marking the water area boundary and water surface line range of the river water area data to be marked; The neural network model is trained according to the annotation results of water area boundary annotation and water surface line range annotation in the river water area data to be annotated to obtain the water area recognition model.

8. A water area identification device, characterized in that: The device comprises: A data acquisition module, used to acquire water area data to be identified; the water area data to be identified includes video data and / or image data; A model processing module, used for inputting the water area data to be identified into a pre-trained water area identification model, and outputting the boundary range and water surface line range of the water area to be identified through the water area identification model; The water area recognition model is obtained by performing resolution enhancement processing and image noise removal processing on various types of river water area data and then annotating the data, and training using the annotation results of the data; The water area identification model includes: an encoder network and a decoder network, wherein the encoder network includes: a backbone network for extracting deep semantic features and a dilated pyramid pooling layer for performing dilated convolution, and the decoder network includes: a bottom feature decoding layer for decoding features output by the backbone network and a deep feature decoding layer for decoding features output by the dilated pyramid pooling layer.

9. A water area identification model training device, characterized in that: The device comprises: A training data acquisition module is used to obtain various types of river water data; A resolution enhancement module, used for performing resolution enhancement processing on the different types of river water area data having a resolution less than a preset resolution threshold using a resolution enhancement technology; An image denoising module is used to perform image denoising on different types of river water area data after and without resolution enhancement processing to obtain river water area data to be labeled; A labeling module, used for labeling the water area boundary and water surface line range of the river water area data to be labeled; The training module is used to train the neural network model according to the labeling results of the water area boundary labeling and the water surface line range labeling in the river water area data to be labeled, so as to obtain the water area recognition model.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 or claim 7 are implemented.

Images