A scaleless water level monitoring method based on improved bilateral network

Abstract

The present invention discloses a scaleless water level monitoring method based on an improved bilateral network, comprising the following steps: calibrating the water level; selecting water level image feature points, performing image registration on the water level image, and processing the position deviation caused by camera movement; inputting water level images collected under different conditions into an improved BiDGANet neural network for training; inputting real-time video acquisition images into a trained model, outputting image semantic segmentation results of water body areas and non-water body areas; calculating the pixel position of the water level line of the semantic segmentation image, and converting it into the actual elevation. The scaleless water level monitoring method provided by the present invention does not rely on the actual water level scale or mark, but uses the principle of photogrammetry, and can convert the water level line coordinates into the actual elevation only through the physical resolution of the template image, thereby overcoming the deficiency of scale blur caused by long-term erosion of the actual scale or mark.

Description

Technical Field

[0001] The present invention belongs to the field of hydrological and water resources monitoring technology and deep learning, and specifically relates to a scaleless water level monitoring method based on an improved bilateral network. Background Art

[0002] Water level is one of the basic hydrological elements of rivers, lakes and reservoirs. Water level data is the basic basis for flood control and drought relief, irrigation, shipping and water conservancy facilities construction and management. Timely and reliable automatic water level monitoring is of great significance to improving the early warning and forecasting level of flood control and drought relief. The water gauge records the water level by reading, which is the most intuitive and simplest water level measurement tool. However, traditional water gauge measurement requires manual regular observation, low degree of automation, and high labor intensity. Existing automatic water level gauges mainly include float type, pressure type, ultrasonic type and radar type according to different measurement principles. However, these devices have certain limitations in actual installation, use and daily maintenance. They generally have high equipment and installation costs, measurement accuracy is easily affected by ambient temperature, water sediment content and on-site control structure, measurement results are not easy to verify, and regular maintenance by staff is required.

[0003] With the development of big data technology and high-performance computing equipment, deep learning technology has been gradually applied to many computer vision tasks. Deep learning methods, with their deep structure, gradually abstract and extract deep features from low-level features of input images, and automatically complete classification in an end-to-end manner, showing excellent classification performance. At present, the use of deep learning technology for water level measurement follows two traditional image-based methods, including methods inspired by human vision and methods based on machine vision, but these methods have certain limitations.

[0004] The method inspired by human vision determines the position of the water level by detecting scale lines and identifying characters. It is very dependent on the visibility of scale lines and image features. It cannot be measured under complex conditions such as local dirt and damage of the water gauge, low image resolution, insufficient natural light, and excessive artificial light. The method based on machine vision first detects the position of the water level in the image, and then uses the transformation relationship to convert the pixel coordinates of the water level into the actual water level reading. There are three main problems with the existing machine vision-based methods. First, although it does not rely on tools such as water gauges with accurate degrees, it still has the limitation of relying on logos or artificial marks as reference objects, and cannot completely solve the problems of natural wear or floating objects. Second, the complexity of the model is not taken into account. Most of the existing models borrow mature convolutional semantic segmentation modules for segmentation, but these models are not necessarily suitable for processing real-time semantic segmentation tasks, and pure convolutional networks also have the disadvantage of insufficient receptive field. Third, it does not take into account the fact that the actual water surface is in a fluctuating state, so the water level conversion method cannot simply predict the water level as a straight line on the spot to improve accuracy. Summary of the invention

[0005] Purpose of the invention: To address the problems of water level measurement using deep learning technology, such as the visibility of scale lines and image features, the inability to measure under complex conditions such as excessive artificial light, and floating objects entangled with the water gauge at the water level line, a scaleless water level monitoring method based on an improved bilateral network is provided, which does not rely on traditional water level scales or mark detection, and avoids various problems caused by wear. The improved bilateral network solves the problems of high complexity and insufficient receptive field of traditional convolutional networks, and the water level line calculation algorithm overcomes the problem that traditional water level segmentation does not take into account water surface fluctuations, and can ensure accurate and stable operation under complex conditions, realize pixel-by-pixel classification and prediction of water level images, and monitor the actual water level of the target water level in real time.

[0006] Technical solution: To achieve the above purpose, the present invention provides a scaleless water level monitoring method based on an improved bilateral network, comprising the following steps:

[0007] S1: Traditional scales or signs are easily eroded by the natural environment, making it difficult to accurately identify the water level. The present invention uses a fixed spherical camera to shoot, divides the picture into 20 equal parts and measures the pixel coordinates of the center point of each dividing line. The camera has 1.5 million pixels, supports a maximum of 20 times optical zoom, a horizontal range of 360° and a vertical range of -15° to 45° automatic flip, and 100m infrared fill light. At the same time, the total station measures the actual elevation of each dividing line relative to the top of the embankment, so that the corresponding relationship between the pixel coordinates p of the dividing line and the actual elevation h can be established simply by using the camera;

[0008] S2: In order to solve the camera offset problem, a template water level image is established, and n pairs of matching control points are selected from the template water level image as the basis for image registration;

[0009] S3: Calculate the spatial transformation matrix using the selected n pairs of matching control points to convert the water level image to be registered into a registered standard water level image, so that the offset camera image satisfies the previously established correspondence between the pixel coordinate p and the actual elevation h;

[0010] S4: Collect orthophoto water level images under different conditions, perform data enhancement to construct a water level image dataset, where the training sample set and the test sample set are divided in a ratio of 8:2;

[0011] S5: Input the water level image dataset into the BiDGANet neural network for training;

[0012] S6: Input the registered standard water level image into the trained BiDGANet neural network to obtain a segmentation map of the water body area and the non-water body area of ​​the water level image, and then expand the pixel gray value after segmentation;

[0013] S7: In the result image obtained by enlarging the pixel grayscale value after segmentation, in order to solve the problem of water surface fluctuation, the pixel grayscale values ​​in the image are accumulated row by row, the pixel length in the vertical direction of the water body part is calculated, and the coordinate position l of the water level line is determined;

[0014] S8: According to the correspondence between the pixel coordinate p and the actual elevation h obtained in S1, the water level line coordinate position l is converted into the actual elevation h;

[0015] A further solution is that step S2 comprises:

[0016] S21: When the water level is low, use the measurement system arranged in S1 to take a water level image as a template water level image;

[0017] S22: The method for selecting n pairs of matching control points is: starting from the upper and lower ends of the template water level image and moving toward the middle, n pairs of matching control points are selected as the basis for image registration. The original pixel coordinates of the control points and the transformed pixel coordinates of the control points are respectively expressed as (x i ,y i ) and (u i , v i ), where i = 1, 2, 3,…, n.

[0018] A further solution is that the formula for calculating the spatial transformation matrix in step S3 is:

[0019]

[0020] Wherein: m1~m9 are transformation coefficients, which are calculated through the control points in the template water level image and the water level image to be registered. The coordinates of the water level image to be registered can be converted into the standard water level image after registration through the formula.

[0021] A further solution is that the data set establishment process in step S4 is specifically as follows:

[0022] Orthophoto water level images under different weather, lighting and water flow conditions in the dry season and the flood season are collected, and the images are flipped, cropped, scaled, and color transformed for data enhancement to construct a water level image dataset.

[0023] A further solution is that the BiDGANet neural network in step S5 is specifically:

[0024] Traditional convolutional networks have the problems of high computational complexity and insufficient receptive field. The present invention adopts an improved bilateral attention network (BiDGANet); this network uses the Bilateral Networ real-time semantic segmentation network as the basic network to reduce the computational complexity, and adds a guided attention module (GA) on its basis to expand the receptive field of the image and increase the accuracy of segmentation. The overall architecture is divided into two parts: the first part is a double branch composed of high and low resolution, and the image will enter the high-resolution branch and the low-resolution branch respectively to obtain feature extraction of different scales; the second part is a double guided attention module, which uses a lightweight attention mechanism to merge the features of different scales in the previous part for output; the trained network is used to test the semantic segmentation of water level images; the input image is a 24-bit 2000×2000 pixel orthophoto image, and the model can generate a pixel classification vector with a pixel grayscale value between 0 and 1.

[0025] A further solution is that the calculation method for obtaining the segmentation map of the water body area and the non-water body area of ​​the water level image in step S6 is:

[0026] After obtaining the pixel classification vector output by the model, the Softmax classifier is used to make pixel-by-pixel predictions. The Softmax multi-classification layer performs exponential calculations on the probability values ​​of each input classifier. After normalization, the category represented by the maximum value is the classification result. After determining the category to which each pixel belongs, the loss function is calculated. The output vector of the Softmax multi-classification layer and the label value of the sample data are cross-entropy processed:

[0027]

[0028] Where: H x (y) is the cross entropy, which is used to measure the difference between the model segmentation result and the label value; x i is the value of the i-th element in the label; y i Output vector (y1, y2, ..., y m )’s i-th vector data;

[0029] The calculation method of the grayscale value of the enlarged pixel after segmentation in step S6 is:

[0030] In the segmentation results, the pixel grayscale values ​​of the water area and the non-water area are 1 and 0 respectively; in the subsequent water level line detection, the pixel grayscale value of each part of the segmentation result image is expanded 100 times, that is, the grayscale values ​​of the water area and the non-water area are 100 and 0 respectively.

[0031] A further solution is that the calculation method for determining the water level coordinate position k in step S7 is:

[0032] S71: Traditional water surface segmentation algorithms do not consider that the water surface is wavy, and the water surface segmentation result will not be a straight line. Therefore, it is necessary to expand the pixel grayscale value after segmentation to obtain the result image, and accumulate the pixel grayscale values ​​in the image row by row:

[0033] b(r)=M(r,1)+M(r,2)+…+M(r,w)

[0034] Where: M(r, w) is the grayscale value of the pixel at (r, w); r is the row coordinate of the current pixel, taking r = 1, 2, ... 2 votes 0 votes; w is the column coordinate of the pixel, taking w = 1, 2, ... 2 votes votes; the fixed grayscale value threshold of the water body is set to T w .

[0035] S72: Traverse from top to bottom in the horizontal projection of the pixel values ​​of the segmentation result image, b(r)≤T w The pixel row coordinates are non-water areas, denoted by r w , b(r)>T w The pixel row coordinates are the water area, denoted by r r ;

[0036] The pixel length in the vertical direction of the water body is calculated by the following formula to determine the coordinate position l of the water level line:

[0037] l=H-(r w2 -r w1 )

[0038] Where: H is the total length of the pixel row coordinates; r w2 is the row coordinate of the top pixel of the non-water part; r w1 is the row coordinate of the bottom pixel of the non-water part.

[0039] A further solution is that the calculation method for converting the water level coordinate position l into the actual elevation h in step S8 is:

[0040] Since the registration image and the template image are in the same coordinate system, the corresponding relationship between the pixel coordinate p and the actual elevation h can be used h = p / Δd to convert the water level coordinate position l into the actual elevation:

[0041] h=l / Δd

[0042] Where: Δd is the physical resolution of the template water level image, and the corresponding unit is pixel / mm.

[0043] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0044] 1. It does not rely on auxiliary tools such as water gauges or markings to measure, which can avoid various problems caused by tools, thereby ensuring that the water level can be measured normally under conditions such as excessive light, heavy rain, strong winds, and garbage on the water surface.

[0045] 2. The BiDGANet neural network is used, and an attention module is added to the real-time semantic segmentation network. It has higher accuracy, lower requirements on equipment, and is more practical in industrial measurement.

[0046] 3. The improved water level coordinate prediction algorithm fully considers the influence of water surface fluctuations and performs cumulative summation processing on each pixel point near the water level, making the water level prediction more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the process of the present invention;

[0048] Figure 2 Schematic diagram of the BiDGANet neural network in the method of the present invention. DETAILED DESCRIPTION

[0049] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0050] This embodiment provides a scaleless water level monitoring method based on an improved bilateral network. Figure 1 As shown, it includes the following steps:

[0051] S1: Traditional scales or signs are easily eroded by the natural environment, making it difficult to accurately identify the water level. The present invention uses a fixed spherical camera to shoot, divides the picture into 20 equal parts and measures the pixel coordinates of the center point of each dividing line. The camera has 1.5 million pixels, supports a maximum of 20 times optical zoom, a horizontal range of 360° and a vertical range of -15° to 45° automatic flip, and 100m infrared fill light. At the same time, the total station measures the actual elevation of each dividing line relative to the top of the embankment, so that the corresponding relationship between the pixel coordinates p of the dividing line and the actual elevation h can be established simply by using the camera;

[0052] S2: In order to solve the camera offset problem, a template water level image is established, and n pairs of matching control points are selected from the template water level image as the basis for image registration;

[0053] S3: Calculate the spatial transformation matrix using the selected N pairs of matching control points to convert the water level image to be registered into a registered standard water level image, so that the offset camera image satisfies the previously established correspondence between the pixel coordinate p and the actual elevation h;

[0054] S4: Collect orthophoto water level images under different conditions, perform data enhancement to construct a water level image dataset, where the training sample set and the test sample set are divided in a ratio of 8:2;

[0055] S5: Input the water level image dataset into the BiDGANet neural network for training;

[0056] S6: Input the registered standard water level image into the trained BiDGANet neural network to obtain a segmentation map of the water body area and the non-water body area of ​​the water level image, and then expand the pixel gray value after segmentation;

[0057] S7: In the result image obtained by enlarging the pixel grayscale value after segmentation, in order to solve the problem of water surface fluctuation, the pixel grayscale values ​​in the image are accumulated row by row, the pixel length in the vertical direction of the water body part is calculated, and the coordinate position l of the water level line is determined;

[0058] S8: According to the correspondence between the pixel coordinate p and the actual elevation h obtained in S1, the water level line coordinate position l is converted into the actual elevation h;

[0059] Step S2 of this embodiment is specifically as follows:

[0060] S21: When the water level is low, use the measurement system arranged in S1 to take a water level image as a template water level image.

[0061] S22: The method for selecting n pairs of matching control points is: starting from the upper and lower ends of the template water level image and moving toward the middle, n pairs of matching control points are selected as the basis for image registration. The sub-pixel coordinates of the control points are expressed as

[0062] (x i ,y i ) and (u i , v i ), where i = 1, 2, 3,…, n.

[0063] The formula for calculating the space transformation matrix in step S3 of this embodiment is:

[0064]

[0065] Wherein: m1~m9 are transformation coefficients, which are calculated through the control points in the template water level image and the water level image to be registered. The coordinates of the water level image to be registered can be converted into the standard water level image after registration through this formula.

[0066] The process of establishing the data set in step S4 of this embodiment is specifically as follows:

[0067] Orthophoto water level images under different weather, lighting and water flow conditions in the dry season and the flood season are collected, and the images are flipped, cropped, scaled, and color transformed for data enhancement to construct a water level image dataset.

[0068] The BiDGANet neural network in step S5 of this embodiment is specifically:

[0069] Traditional convolutional networks have the problems of high computational complexity and insufficient receptive field. The present invention adopts a bilateral network with improved attention (BiDGANet); this network uses the Bilateral Networ real-time semantic segmentation network as the basic network to reduce the computational complexity, and adds a guided attention module (GA) on its basis to expand the receptive field of the image and increase the accuracy of segmentation. The overall architecture is divided into two parts: the first part is a dual branch composed of high and low resolutions, and the image will enter the high-resolution branch and the low-resolution branch respectively to obtain feature extraction of different scales; the second part is a dual guided attention module, which uses a lightweight attention mechanism to merge the features of different scales in the previous part for output; the trained network is used to test the semantic segmentation of water level images; the input image is a 24-bit 2000×2000 pixel orthophoto image, and the model can generate a pixel classification vector with a pixel grayscale value between 0 and 1.

[0070] The overall architecture is as follows Figure 2 As shown in the figure, there are three main parts: the high-resolution branch in the upper right box, the low-resolution branch in the upper left box, and the dual-guided attention module in the dashed box below. The high-resolution branch is divided into three stages, and the low-resolution branch has six RSU blocks. The numbers in the cubes are the ratio of the feature map size ratio to the input resolution. The last stage of the low-resolution branch is the output of the context embedding block. In the dual-guided attention part, "1 / 2×down" represents the downsampling operation, "2×up" represents the upsampling operation, the rectangle "Cat" represents the concat operation, and "GA" represents GuidedAttention.

[0071] In this embodiment, the calculation method for obtaining the segmentation map of the water body area and the non-water body area of ​​the water level image in step S6 is:

[0072] After obtaining the pixel classification vector output by the model, use the Softmax classifier to make pixel-by-pixel predictions. The Softmax multi-classification layer performs exponential calculations on the probability values ​​of each input classifier. After normalization, the category represented by the maximum value is the classification result. After determining the category to which each pixel belongs, the loss function is calculated, and the output vector of the Softmax multi-classification layer and the label value of the sample data are cross-entropy processed:

[0073]

[0074] Where: H x (y) is the cross entropy, which is used to measure the difference between the model segmentation result and the label value; x i is the value of the i-th element in the label; y i Output vector (y1, y2, ..., y m )’s i-th vector data;

[0075] The calculation method of the grayscale value of the enlarged pixel after segmentation in step S6 is:

[0076] In the segmentation results, the pixel grayscale values ​​of the water area and the non-water area are 1 and 0 respectively; in the subsequent water level line detection, the pixel grayscale value of each part of the segmentation result image is expanded 100 times, that is, the grayscale values ​​of the water area and the non-water area are 100 and 0 respectively.

[0077] In this embodiment, the calculation method for determining the water level line coordinate position l in S7 is:

[0078] S71: Traditional water surface segmentation algorithms do not consider that the water surface is wavy, and the water surface segmentation result will not be a straight line. Therefore, it is necessary to expand the pixel grayscale value after segmentation to obtain the result image, and accumulate the pixel grayscale values ​​in the image row by row:

[0079] b(r)=M(r,1)+M(r,2)+…+M(r,w)

[0080] Where: M(r, w) is the grayscale value of the pixel at (r, w); r is the row coordinate of the current pixel, taking r = 1, 2, ... 2 votes 0 votes; w is the column coordinate of the pixel, taking w = 1, 2, ... 2 votes votes; the fixed grayscale value threshold of the water body is set to T w .

[0081] S72: Traverse from top to bottom in the horizontal projection of the pixel values ​​of the segmentation result image, b(r)≤T w The pixel row coordinates are non-water areas, denoted by r w , b(r)>T w The pixel row coordinates are the water area, denoted as rr;

[0082] The pixel length in the vertical direction of the water body is calculated by the following formula to determine the coordinate position l of the water level line:

[0083] l=H-(r w2 -r w1 )

[0084] Where: H is the total length of the pixel row coordinates; r w2is the row coordinate of the top pixel of the non-water part; r w1 is the row coordinate of the bottom pixel of the non-water body;

[0085] In this embodiment, the calculation method for converting the water level coordinate position l into the actual elevation h in step S8 is:

[0086] Since the registration image and the template image are in the same coordinate system, the corresponding relationship between the pixel coordinate p and the actual elevation h can be used h = p / Δd to convert the water level coordinate position l into the actual elevation:

[0087] h=l / Δd

[0088] Where: Δd is the physical resolution of the template water level image, and the corresponding unit is pixel / mm.

[0089] According to the above examples, in order to solve the problems existing in water level measurement using deep learning technology, such as the dependence on the visibility of scale lines and image features, the inability to measure under complex conditions such as excessive artificial lighting, and the entanglement of floating objects on the water level scale, the method of the present invention performs pixel-level segmentation on the water level scale image, utilizes the relationship between the pixel position and the actual water level to detect the pixel position of the water level line, and converts it into the actual elevation; the method does not rely on auxiliary tools such as traditional water level rulers for detection, avoids various problems caused by the water level ruler, and can ensure accurate and stable operation under complex conditions, realize pixel-by-pixel classification and prediction of water level images, and monitor the actual elevation of the target water level in real time.

Claims

1. A scaleless water level monitoring method based on an improved bilateral network, characterized in that: The steps include: S1: Use a fixed spherical camera to take pictures of the water level, divide the picture into equal parts and measure the pixel coordinates of the center point of each equal-division line; at the same time, the total station measures the actual elevation of each equal-division line relative to the top of the embankment, and uses the camera to establish the corresponding relationship between the pixel coordinates p of the equal-division line and the actual elevation h; S2: Establish a template water level image to solve the camera offset problem, and select n pairs of matching control points from the template water level image as the basis for image registration; S3: Calculate the spatial transformation matrix using the selected n pairs of matching control points to convert the water level image to be registered into a registered standard water level image, so that the offset camera image satisfies the previously established correspondence between the pixel coordinate p and the actual elevation h; S4: Collect orthophoto water level images under different conditions, perform data enhancement to construct a water level image dataset, where the training sample set and the test sample set are divided in a ratio of 8:2; S5: Input the water level image dataset into the BiDGANet neural network for training; S6: Input the registered standard water level image into the trained BiDGANet neural network to obtain a segmentation map of the water body area and the non-water body area of ​​the water level image, and then expand the pixel gray value after segmentation; S7: In the result image obtained by enlarging the pixel grayscale value after segmentation, the pixel grayscale values ​​in the image are accumulated row by row to solve the problem of water surface fluctuation, the pixel length in the vertical direction of the water body part is calculated, and the coordinate position l of the water level line is determined; S8: According to the correspondence between the pixel coordinate p and the actual elevation h obtained in S1, the water level line coordinate position l is converted into the actual elevation h; The BiDGANet neural network in step S5 is specifically: The Bilateral Networ real-time semantic segmentation network is used as the basic network to reduce the computational complexity. On this basis, a guided attention module is added to expand the receptive field of the image and increase the accuracy of segmentation. The overall architecture consists of two parts: the first part is a dual branch consisting of high and low resolution. The image will enter the high-resolution branch and the low-resolution branch respectively to obtain feature extraction of different scales; The second part is the dual-guided attention module, which uses a lightweight attention mechanism to fuse the features of different scales in the previous part for output; the trained network is used to test the semantic segmentation of water level images; the input image is a 24-bit 2000×2000 pixel orthophoto, and the model can generate pixel classification vectors with pixel grayscale values ​​between 0 and 1; The calculation method for obtaining the segmentation map of the water body area and the non-water body area of ​​the water level image in step S6 is: After obtaining the pixel classification vector generated by the model, use the Softmax classifier to make pixel-by-pixel predictions; The Softmax multi-classification layer performs exponential calculations on the probability values ​​of each input classifier. After normalization, the category represented by the maximum value is the classification result. After determining the category to which each pixel belongs, the loss function is calculated. The output vector of the Softmax multi-classification layer and the label value of the sample data are subjected to cross entropy processing: Where: H x (y) is the cross entropy, which is used to measure the difference between the model segmentation result and the label value; x i is the value of the i-th element in the label; y i Output vector (y1, y2, ..., y m )’s i-th vector data; The calculation method of the grayscale value of the enlarged pixel after segmentation in step S6 is: The pixel grayscale values ​​of the water area and non-water area in the segmentation result are 1 and 0 respectively; in the subsequent water level detection, the pixel grayscale value of each part of the segmentation result image is expanded 100 times, that is, the grayscale values ​​of the water area and non-water area are 100 and 0 respectively; The calculation method for determining the water level coordinate position l in step S7 is: S71: Accumulate the grayscale values ​​of pixels in the image row by row: b(r)=M(r,1)+M(r,2)+…+M(r,w) Where: M(r, w) is the grayscale value of the pixel at (r, w); r is the row coordinate of the current pixel, taking r = 1, 2, ... 2000; w is the column coordinate of the pixel, taking w = 1, 2, ... 200; the fixed grayscale value threshold of the water body is set to T w ; S72: Traverse from top to bottom in the horizontal projection of the pixel values ​​of the segmentation result image, b(r)≤T w The pixel row coordinates are non-water areas, denoted by r w , b(r)>T w The pixel row coordinates are the water area, denoted by r r ; The pixel length in the vertical direction of the water body is calculated by the following formula to determine the coordinate position l of the water level line: l＝H-(r w2 -r w1 ) Where: H is the total length of the pixel row coordinates; r w2 is the row coordinate of the top pixel of the non-water part; r w1 is the row coordinate of the bottom pixel of the non-water part.

2. The scaleless water level monitoring method based on improved bilateral network according to claim 1 is characterized in that: Step S2 is specifically as follows: S21: When the water level is low, the water level image taken by S1 is used as a template water level image; S22: The method for selecting n pairs of matching control points is: starting from the upper and lower ends of the template water level image and moving toward the middle, n pairs of matching control points are selected as the basis for image registration. The original pixel coordinates of the control points and the transformed pixel coordinates of the control points are respectively expressed as (x i ,y i ) and (u i , v i ), where i = 1, 2, 3,…, n.

3. The scaleless water level monitoring method based on improved bilateral network according to claim 1 is characterized in that: The formula for calculating the spatial transformation matrix in step S3 is: Wherein: m1~m9 are transformation coefficients, which are calculated through the control points in the template water level image and the water level image to be registered. The coordinates of the water level image to be registered are converted into the standard water level image after registration through the formula.

4. The scaleless water level monitoring method based on improved bilateral network according to claim 1 is characterized in that: The process of establishing the water level image dataset in step S4 is specifically as follows: Orthophoto water level images under different weather, lighting and water flow conditions in the dry season and the flood season are collected, and the images are flipped, cropped, scaled, and color transformed for data enhancement to construct a water level image dataset.

5. The scaleless water level monitoring method based on improved bilateral network according to claim 1 is characterized in that: The calculation method for converting the water level coordinate position l into the actual elevation h in step S8 is: Using the corresponding relationship between pixel coordinates p and actual elevation h, h = p / Δd, the water level coordinate position l is converted to the actual elevation: h=l / Δd Where: Δd is the physical resolution of the template water level image, and the corresponding unit is pixel / mm.

Images