Water level gauge reading method based on deep learning

Abstract

The present invention discloses a water level gauge reading method based on deep learning, including: obtaining water level gauge height and E-character target height information; obtaining a target image, identifying the water level gauge position and corner point position to correct the target image; performing instance segmentation on the E-character target in the corrected image to obtain its recognition frame coordinates and segmentation results; sorting the E-character targets from top to bottom according to the recognition frame coordinates, and calculating the number of drastic changes in the y coordinate value of its horizontal projection in the vertical direction according to the segmentation results, and judging whether it is complete according to the number of drastic changes; if the updated recognition frame coordinates exceed the water level gauge position, the E-character target is judged to be an incomplete E-character target connected to the water surface, and its effective number of copies is calculated; according to the sum of the number of copies of all E-character targets, the water level gauge height and the E-character target height, the current water level is calculated. The present invention can avoid human errors, overcome the influence of environmental factors on water level measurement, and greatly improve the accuracy and reliability of water level detection.

Description

Technical Field

[0001] The present invention relates to the field of machine vision technology, and in particular to a water level gauge reading method based on deep learning. Background Art

[0002] Water level monitoring has important application value in water resource management, flood warning, ecological protection and other fields. Accurate water level detection can not only help prevent disasters, but also provide a scientific basis for the rational allocation of water resources and the protection of the ecological environment. At present, water level monitoring technology mainly relies on traditional methods such as manual reading of water level gauges or mechanical floats. However, these traditional methods have certain limitations and cannot meet the requirements of modern water level monitoring for efficiency, real-time and accuracy.

[0003] Traditional water level measurement methods usually require manual on-site readings. This method is time-consuming and inefficient, especially in remote areas or under extreme weather conditions, where manual operation becomes more difficult. In addition, manual readings are greatly affected by human factors and are prone to measurement errors, resulting in inaccurate data, which in turn affects the reliability of water level monitoring.

[0004] With the development of computer vision technology, water level detection methods based on image processing have gradually been applied. Traditional image processing methods rely on edge detection, image segmentation and other technologies for water level detection. However, traditional image processing methods have poor adaptability to complex backgrounds, different lighting conditions, occlusion and other problems, and are difficult to cope with complex situations in water level gauge images. For example, the water level gauge may be affected by dirt, damage or occlusion, which may affect the image quality and lead to inaccurate detection results.

[0005] With the rapid development of deep learning technology, especially the progress of target detection and image segmentation technology, water level detection methods based on deep learning have become a hot topic of research. Deep learning can automatically extract features in images through large-scale data training, and has stronger robustness and accuracy than traditional methods. Using deep learning models to perform target detection and instance segmentation on water level gauge images can automatically identify the position and corner points of the water level gauge, and perform real-time detection and calculation of the water level. This method can not only improve detection accuracy, but also operate effectively in various complex environments.

[0006] Although deep learning technology has shown great potential in water level monitoring, current related technologies still face some challenges. For example, the segmentation accuracy of the E-shaped target in the water level gauge image is not high, especially in the presence of dirt, occlusion or water level gauge damage, the E-shaped target may be misjudged or missed. Therefore, how to improve the segmentation accuracy of the water level gauge image, accurately estimate the missed E-shaped target, and quickly calculate the water level are still problems that need to be solved by current technology. Summary of the invention

[0007] In order to solve the technical problems existing in the above-mentioned prior art, the present invention provides a water level gauge recognition method based on deep learning. By utilizing deep learning technologies such as target detection, key point detection, and instance segmentation, it can accurately identify each E-shaped target of the water level gauge and its position, thereby realizing efficient and accurate water level monitoring.

[0008] In order to achieve the above-mentioned invention object, the technical solution provided by the present invention includes:

[0009] The water level gauge reading method based on deep learning includes the following steps:

[0010] S1. Get the water level gauge height and E-target height information;

[0011] S2. Acquire a target image including a water level gauge, identify the water level gauge position and the corner point position to correct the target image, and obtain a corrected image;

[0012] S3. Perform instance segmentation on the E-shaped target in the corrected image to obtain its recognition box coordinates and segmentation results;

[0013] S4. Sort the E-shaped targets from top to bottom according to the recognition frame coordinates, and calculate the number of drastic changes in the y coordinate values ​​of their horizontal projections in the vertical direction according to the segmentation results. If the number of drastic changes is 5, the E-shaped target is determined to be complete; if it is less than 5, it is determined to be incomplete, and the recognition frame coordinates are estimated and updated; each E-shaped target is divided into 5 parts according to the number of drastic changes;

[0014] S5. If the updated identification frame coordinates exceed the water level gauge position, the E-shaped target is determined to be an incomplete E-shaped target connected to the water surface, and the number of valid copies is calculated;

[0015] S6. Calculate the current water level according to the total number of all E-targets, the height of the water level gauge and the height of the E-target.

[0016] Furthermore, the method of identifying the water level gauge position and the corner point position to correct the target image includes:

[0017] Set the detection category of the yolov9-pose model to 1, the number of key points to 4, and use data enhancement training to complete the training of the yolov9-pose model;

[0018] Input the target image into the trained yolov9-pose model, and output the water level gauge recognition frame coordinates and corner point position coordinates;

[0019] The target image is cropped according to the coordinates of the water level gauge identification frame to obtain a process image containing only the water level gauge;

[0020] An affine matrix is ​​calculated according to the corner point position coordinates and an affine transformation is performed, and the process image is corrected to obtain a corrected image.

[0021] Furthermore, the method of calculating the affine matrix according to the corner point position coordinates and performing affine transformation includes:

[0022] The corner point position coordinates in the target image are translated to the process image to obtain the corner point position coordinates before correction (x1, y1), (x2, y2), (x3, y3), (x4, y4), respectively;

[0023] The corrected corner point coordinates are (x1, y1), (x2, y1), (x2, y4), (x1, y4);

[0024] An affine matrix is ​​calculated according to the position coordinates of the corner points before and after correction, and the process image is corrected according to the affine matrix.

[0025] Furthermore, the method for acquiring the horizontal projection of the E-shaped target includes:

[0026] Get the binary mask image of the E-shaped target segmentation result, perform opening and closing operations to denoise it, get the histogram of all non-zero values ​​at x in the y coordinate of the denoised image, and complete the horizontal projection.

[0027] Furthermore, the method for calculating the number of drastic changes includes:

[0028] ;

[0029] Where num is the number of times the y-coordinate value of the horizontal projection of the segmentation result of the E-shaped target in the vertical direction changes dramatically, h is the height of the E-shaped target, w is the width of the E-shaped target, hist[y] is the horizontal projection value of the current y-coordinate position, before is the y-coordinate position that met the drastic change condition last time, δ(x) is the indicator function and .

[0030] Furthermore, before calculating the current water level, it is necessary to determine whether there is an E-shaped target that has been missed. The specific methods include:

[0031] Calculate the number of missed E-targets E_loss between each two adjacent E-targets:

[0032] E_loss=int(0.5+2*(E2_y1-E1_y2) / (E2_y2-E2_y1+E1_y2-E1_y1)): where E1_y1, E1_y2, E2_y1 and E2_y2 are the y coordinates of the upper and lower edges of two adjacent E-shaped targets E1 and E2 respectively; int() is the rounding function.

[0033] Further, according to the total number of copies of all E-targets, the water level gauge height information and the E-target height information, the method for calculating the current water level includes:

[0034] Water level = h1-h2 / 5×n; where h1 is the height of the water level gauge, h2 is the height of the E-shaped target, and n is the total number of portions.

[0035] Beneficial Effects

[0036] The present invention can accurately identify the position, corner points and E-shaped targets of the water level gauge, avoid human errors, and overcome the influence of environmental factors (such as bad weather, dirt, damage, etc.) on water level measurement, greatly improving the accuracy and reliability of water level detection, and can realize fully automated monitoring and real-time calculation of water level, greatly reducing the need for manual operation. Especially in remote areas or complex environments, unattended automated monitoring can be realized, reducing labor costs and operational difficulties.

[0037] The present invention has strong robustness and can adapt to complex backgrounds, different lighting conditions, and problems such as occlusion and dirt of the water level gauge. Through efficient image processing and target detection algorithms, the system can effectively identify the water level gauge and calculate the water level even under low-quality image conditions, solving the problem that traditional methods cannot accurately detect under complex conditions.

[0038] The present invention estimates the missed E-shaped targets, thus avoiding the missed detection problem caused by partial E-shaped target occlusion or dirt. By calculating the number of major changes of the E-shaped targets and the number of missed detections of adjacent E-shaped frames, the missed detection situation can be effectively corrected, and the integrity and accuracy of the detection can be improved, especially when the water level gauge is damaged, dirty or blocked, it can still maintain a high stability of water level calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic flow chart of a method for reading a water level gauge based on deep learning provided in a preferred embodiment of the present invention;

[0040] Figure 2 A schematic diagram of a target image including a water level gauge obtained according to another preferred embodiment of the present invention;

[0041] Figure 3 A schematic diagram of a process image of obtaining only a water level gauge after cropping a target image provided by another preferred embodiment of the present invention;

[0042] Figure 4 A schematic diagram of correcting a process image to obtain a corrected image provided by another preferred embodiment of the present invention;

[0043] Figure 5A schematic diagram of performing instance segmentation of an E-shaped target in a rectified image provided by another preferred embodiment of the present invention;

[0044] Figure 6 A schematic diagram of a position where a complete E-shaped target undergoes a dramatic change, provided in another preferred embodiment of the present invention; DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described below in conjunction with the accompanying drawings. In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside", "outside" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the present invention.

[0046] Embodiment 1

[0047] like Figure 1 As shown, the present invention discloses a water level gauge reading method based on deep learning, comprising the steps of:

[0048] S1. Obtaining the water level gauge height and E-target height information. It should be understood that the total height information of the water level gauge to be measured and the height information of a single E-target can be obtained manually or automatically, and this information can be measured and recorded when designing the water level gauge, or through pre-acquired water level gauge dictionary information.

[0049] S2. Obtain a target image containing a water gauge, identify the position of the water gauge and the corner point position, correct the target image, and obtain a corrected image. The target image refers to an image containing the water gauge and its related background, which can be obtained by an image acquisition device (such as a camera or a surveillance camera). Figure 2 As shown, the water level gauge has multiple water level marks, i.e., E-shaped targets. In different water level gauge instances, the arrangement and orientation of the E-shaped targets are also different. The E-shaped targets represent different water level heights, and the current water level can be calculated based on the recognition of these marks. It should be understood that the background environment that may be contained in the target image, such as the water surface, surrounding objects, weather conditions, etc., may affect the visibility of the water level gauge. In addition, the dirt, obstructions, etc. of the water level gauge may affect the image quality and recognition accuracy. Furthermore, the water level gauge may produce certain distortion or deformation due to the shooting angle or environmental factors, so image correction is required. Those skilled in the art can use the image correction method commonly used in the field. In some preferred embodiments, a preferred correction method is also provided, specifically including:

[0050] S201. Set the detection category of the yolov9-pose model to 1, indicating that only the water level gauge is detected; the number of key points is 4, corresponding to the four corner points of the water level gauge; use the data enhancement training method to complete the training of the yolov9-pose model.

[0051] S202. Input the target image into the trained yolov9-pose model, and output the water level gauge recognition frame coordinates and corner point position coordinates.

[0052] S203. The target image is cropped according to the water gauge identification frame coordinates to obtain a process image containing only the water gauge, such as Figure 3 As shown. The process image only contains the water level gauge and its background information, and other irrelevant areas are removed. The size and content of the process image can also be further adjusted by those skilled in the art according to actual needs. Furthermore, the corner point position coordinates are the position coordinates of the corner point in the target image, and translation is required in the cropped process image to obtain the corrected corner point position coordinates.

[0053] S204. Calculate the affine matrix according to the corner point position coordinates and perform affine transformation, correct the process image to obtain a corrected image, such as Figure 4 As shown. The corrected image is an image that is consistent with the actual angle of the water level gauge under the standard viewing angle, in which the water level gauge has been restored to the forward viewing angle, and the four corner points of the water level gauge have been aligned with the standard coordinate system. The affine transformation is a basic transformation method in computer image processing. It is used to perform geometric transformations on images while keeping the straight lines and proportional relationships in the image unchanged. The key step is to determine the radial matrix. In some preferred embodiments, a method is provided to calculate the affine transformation matrix based on the positional relationship of the corner point coordinates before and after correction, so as to transform the tilted or distorted water level gauge image into a standard forward view. Specifically including:

[0054] S2041. Translate the corner point position coordinates in the target image to the process image to obtain the corner point position coordinates before correction (x1, y1), (x2, y2), (x3, y3), (x4, y4) respectively.

[0055] S2042. Obtain the corrected corner point position coordinates as (x1, y1), (x2, y1), (x2, y4), (x1, y4).

[0056] S2043. Calculate an affine matrix based on the corner point position coordinates before and after correction, and correct the process image based on the affine matrix.

[0057] By calculating the affine matrix, the tilted and deformed image can be corrected to a standard viewing angle, so as to adapt to the complex scenes in practical applications where the water level gauge may be in different viewing angles or environments.

[0058] S3. Perform instance segmentation on the E-shaped target in the corrected image to obtain its recognition frame coordinates and segmentation results. The segmentation results are as follows: Figure 5 In some preferred embodiments, the RTMDet model can be used to perform instance segmentation on the E-shaped target in the rectified image. The RTMDet model is trained by a data enhancement training method.

[0059] S4. Sort the E-shaped targets from top to bottom according to the recognition frame coordinates, and calculate the number of drastic changes in the y coordinate values ​​of their horizontal projections in the vertical direction according to the segmentation results. If the number of drastic changes is 5, the E-shaped target is determined to be complete; if it is less than 5, it is determined to be incomplete, and the recognition frame coordinates are estimated and updated; Figure 6 As shown, for a complete E-shaped target, it should contain 5 major changes in the y-coordinate value, and each E-shaped target is divided into 5 parts accordingly.

[0060] The horizontal projection refers to projecting the E-shaped target along the horizontal direction (i.e., the x-axis direction) to obtain the statistical characteristics or projection results of the E-shaped target in the horizontal direction. In this application, the horizontal projection is mainly used to process the segmentation results of the E-shaped target, and the integrity of the E-shaped target and its share are determined by counting the distribution of the E-shaped target in the horizontal direction. In some preferred embodiments, a method for obtaining the horizontal projection is provided, which specifically includes:

[0061] S401. Perform instance segmentation on the letter E in the water gauge image using a deep learning model (such as the RTMDet model) to obtain a binary mask image of each letter E target. In the binary mask image, the area of ​​the letter E target is 1, and the background is 0.

[0062] S402. Perform an opening operation (first erosion and then expansion, kernel size=(3,3)) on the mask image to remove noise points in the mask, and then perform a closing operation (first expansion and then erosion, kernel size=(3,3)) on the mask image to remove holes in the mask.

[0063] S403. For each binary mask image of the letter E, count the number of non-zero values ​​in each column (i.e., the x-axis direction) (i.e., the number of "pixels" occupied by the letter E in the column). This statistical process is called horizontal projection. In short, it "compresses" the distribution of the letter E in the horizontal direction and generates a projection image about the y coordinate.

[0064] S404. By counting the number of non-zero values ​​in each column, a horizontal projection diagram, i.e., histogram statistics, is obtained. The projection value corresponding to each y coordinate represents the "density" or "coverage" of the letter E at the y position.

[0065] It should be understood that the drastic change refers to a significant change in the projection value of a certain position during the horizontal projection process, which is usually manifested as a significant change in the pixel value in the horizontal direction from dense (such as 1 or more pixels) to sparse (such as 0) or vice versa. In this application, the drastic change usually occurs in the contour or edge area of ​​the E-shaped target, especially when part of the image is blocked, dirty or affected by other factors, resulting in an incomplete E-shaped target in the image. The technical thinking is as follows:

[0066] The distribution of the E-shaped target in the water gauge image is usually regular and continuous, and each "segment" of each E-shaped target presents a relatively uniform pixel distribution in the horizontal direction. If some parts of the E-shaped target are blocked or incomplete due to occlusion, dirt or damage to the water gauge, there will be a huge change in the horizontal projection image at these locations. The huge change reflects that the pixel distribution of the E-shaped target has undergone a sudden change in the horizontal direction, which means that part of the E-shaped target in the image is missing or incomplete. In the horizontal projection image, the x value corresponding to each y coordinate changes smoothly, which is manifested as a gradual increase or decrease in pixels in the horizontal direction without obvious jumps. Due to factors such as occlusion and dirt, some E-shaped targets may not be fully displayed, which is manifested in the projection as a huge change in the projection value of some y coordinates, such as a sudden change from a full column to zero, indicating that the E-shaped target in this part is missing or blocked. Furthermore, each E-shaped target on the water gauge is usually designed as equally spaced segments. A complete E-shaped target can be divided into 5 segments (5 parts), which correspond to the horizontal part of the E-shaped target (horizontal stroke) and the gap in the middle. At this time, the projection value changes significantly (big change) at a certain position, indicating that this may be the boundary of a segment of the E-shaped target. A complete E-shaped target usually produces 5 big changes in horizontal projection, which correspond to the boundaries between E-shaped target segments. Based on this, the proportion of complete and incomplete E-shaped targets can be calculated.

[0067] Furthermore, in some preferred embodiments, a specific method for calculating the number of major changes is provided, including:

[0068] The calculation method of the number of drastic changes includes:

[0069] ;

[0070] Where num is the number of times the y-coordinate value of the horizontal projection of the segmentation result of the E-shaped target in the vertical direction changes dramatically, h is the height of the E-shaped target, w is the width of the E-shaped target, hist[y] is the horizontal projection value of the current y-coordinate position, before is the y-coordinate position that met the drastic change condition last time, δ(x) is the indicator function and Among them, y=h / 10 is an initial threshold, which is set based on the consideration that since the E-shaped detection frame may have a certain deviation, setting h / 10 as the starting threshold can avoid detecting erroneous changes in the image boundary or other irrelevant areas. The above formula can be expressed in pseudo code as follows:

[0071]

[0072] S5. If the updated identification frame coordinates exceed the water level gauge position, the E-shaped target is determined to be an incomplete E-shaped target connected to the water surface, and its valid number is calculated.

[0073] If the updated recognition frame coordinates exceed the range of the water level gauge, it means that part or all of the E-shaped target may be blocked or located at the bottom of the water level gauge. At this time, it can be inferred that the E-shaped target is in contact with the water surface. Therefore, it may be an incomplete E-shaped target. At this time, the water level cannot be calculated directly according to the rules of the complete E-shaped target (i.e. divided into 5 parts). At this time, it is necessary to estimate the "valid number of parts" of the E-shaped target according to certain rules. Specifically, the valid number of parts can be obtained by calculating the number of major changes of the complete E-shaped target.

[0074] S6. Calculate the current water level according to the total number of all E-targets, the height of the water level gauge and the height of the E-target.

[0075] In some preferred embodiments, some E-shaped targets may not be successfully detected because the water level gauge may be dirty, blocked or damaged. In order to compensate for this missed detection, it is necessary to estimate the number of missed E-shaped targets by the distance and position relationship between adjacent E-shaped targets. The specific method includes:

[0076] Calculate the number of missed E-targets E_loss between each two adjacent E-targets:

[0077] E_loss=int(0.5+2*(E2_y1-E1_y2) / (E2_y2-E2_y1+E1_y2-E1_y1)): E1_y1, E1_y2, E2_y1 and E2_y2 are the y coordinates of the upper and lower edges of two adjacent E-shaped targets E1 and E2 respectively; int() is a rounding function, such as 0.9 is rounded to 1, and 1.2 is rounded to 1. The constant 0.5 is used to ensure that at least one E-shaped frame can be estimated to be missed even in the smallest case, and to balance and adjust the accuracy of missed detection estimation. The specific design ideas are as follows:

[0078] When the int() function processes floating-point numbers, it only retains the integer part and directly discards the decimal part. For example: the result of int(0.9) is 0; the result of int(1.4) is 1. This means that even if the floating-point number is very close to the next integer (such as 0.9), the int() function will not round it to a larger integer. To solve the above problem, by adding 0.5 to the floating-point number, the value close to the next integer can be "pushed" to the upper limit of the integer. For example: int(0.5 + 0.9) = int(1.4) = 1. int(0.5 + 0.1) = int(0.6) = 0. This can achieve an effect close to rounding: when the floating-point number is 0.5 or above, it is rounded up to the next integer; when the floating-point number is lower than 0.5, it remains at the current integer.

[0079] Once the total number of all E-targets (including the number of missed E-targets) is obtained, the current water level can be calculated using the following formula, combining the total height of the water level gauge and the actual height information of each E-target:

[0080] Water level = h1-h2 / 5×n; where h1 is the height of the water level gauge, h2 is the height of the E-shaped target, and n is the total number of portions.

[0081] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A water level gauge reading method based on deep learning, characterized in that: Includes steps: S1. Get the water level gauge height and E-target height information; S2. Acquire a target image including a water level gauge, identify the water level gauge position and the corner point position to correct the target image, and obtain a corrected image; S3. Perform instance segmentation on the E-shaped target in the corrected image to obtain its recognition box coordinates and segmentation results; S4. Sort the E-shaped targets from top to bottom according to the recognition frame coordinates, and calculate the number of drastic changes in the y coordinate values ​​of their horizontal projections in the vertical direction according to the segmentation results. If the number of drastic changes is 5, the E-shaped target is determined to be complete; if it is less than 5, it is determined to be incomplete, and the recognition frame coordinates are estimated and updated; each E-shaped target is divided into 5 parts according to the number of drastic changes; S5. If the updated identification frame coordinates exceed the water level gauge position, the E-shaped target is determined to be an incomplete E-shaped target connected to the water surface, and the number of valid copies is calculated; S6. Calculate the current water level according to the total number of all E-targets, the height of the water level gauge and the height of the E-target.

2. The water level gauge reading method based on deep learning as claimed in claim 1, characterized in that: The method for correcting the target image by identifying the water gauge position and the corner point position comprises: Set the detection category of the yolov9-pose model to 1, the number of key points to 4, and use data enhancement training to complete the training of the yolov9-pose model; Input the target image into the trained yolov9-pose model, and output the water level gauge recognition frame coordinates and corner point position coordinates; The target image is cropped according to the coordinates of the water level gauge identification frame to obtain a process image containing only the water level gauge; An affine matrix is ​​calculated according to the corner point position coordinates and an affine transformation is performed, and the process image is corrected to obtain a corrected image.

3. The water level gauge reading method based on deep learning as claimed in claim 2 is characterized in that: The method of calculating the affine matrix and performing affine transformation according to the corner point position coordinates includes: The corner point position coordinates in the target image are translated to the process image to obtain the corner point position coordinates before correction (x1, y1), (x2, y2), (x3, y3), (x4, y4), respectively; The corrected corner point coordinates are (x1, y1), (x2, y1), (x2, y4), (x1, y4); An affine matrix is ​​calculated according to the position coordinates of the corner points before and after correction, and the process image is corrected according to the affine matrix.

4. The water level gauge reading method based on deep learning as claimed in claim 1, characterized in that: The method for acquiring the horizontal projection of the E-shaped target includes: Get the binary mask image of the E-shaped target segmentation result, perform opening and closing operations to denoise it, get the histogram of all non-zero values ​​at x in the y coordinate of the denoised image, and complete the horizontal projection.

5. The water level gauge reading method based on deep learning as claimed in claim 1, characterized in that: The calculation method of the number of drastic changes includes: ； Where num is the number of times the y-coordinate value of the horizontal projection of the segmentation result of the E-shaped target in the vertical direction changes dramatically, h is the height of the E-shaped target, w is the width of the E-shaped target, hist[y] is the horizontal projection value of the current y-coordinate position, before is the y-coordinate position that met the drastic change condition last time, δ(x) is the indicator function and .

6. The water level gauge reading method based on deep learning as claimed in claim 1, characterized in that: Before calculating the current water level, it is necessary to determine whether there is an E-shaped target that has been missed. The specific methods include: Calculate the number of missed E-targets E_loss between each two adjacent E-targets: E_loss=int(0.5+2*(E2_y1-E1_y2) / (E2_y2-E2_y1+E1_y2-E1_y1)): where E1_y1, E1_y2, E2_y1 and E2_y2 are the y coordinates of the upper and lower edges of two adjacent E-shaped targets E1 and E2 respectively; int() is the rounding function.

7. The water level gauge reading method based on deep learning as claimed in claim 1, characterized in that: The method for calculating the current water level according to the total number of copies of all E-targets, the water level gauge height information and the E-target height information includes: Water level = h1-h2 / 5×n; where h1 is the height of the water level gauge, h2 is the height of the E-shaped target, and n is the total number of portions.

Images