Water gauge identification method and system based on three-channel comparative learning
By combining three-channel contrastive learning and sliding window technology with frequency domain and reflection features, the problem of low water surface segmentation accuracy in water level gauge identification was solved, and high-precision water level detection was achieved in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-24
- Publication Date
- 2026-04-14
AI Technical Summary
Existing water level gauge identification methods are not very accurate due to factors such as lighting, noise, aquatic plants, and floating objects. Furthermore, data annotation is resource-intensive and requires high accuracy, making it difficult to reliably identify water surface boundaries.
A three-channel contrastive learning approach is adopted, which combines deep learning features and sliding windows with frequency domain features and reflection features. The model is trained using multi-task learning to identify the features of the water surface and non-water surface, calculate the feature distance between the upper and lower window regions, and determine the water surface boundary.
It improves the accuracy of water surface segmentation, has strong environmental adaptability, reduces resource consumption, and can stably detect water level lines under different lighting and transparency conditions.
Smart Images
Figure CN121861633A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image processing, specifically relating to a water level gauge recognition method and system based on three-channel contrast learning. Background Technology
[0002] Current image recognition-based water level gauge identification methods mainly include the following: Gradient feature-based approach: This approach extracts water level gradient features and obtains the water surface boundary line by extracting straight lines or lateral projections.
[0003] Segmentation-based approach: This approach trains a water area segmentation model and segments images containing water level gauges. The intersection of the water area and the water level gauge is the water surface boundary line.
[0004] Based on a detection and classification method: The letter "E" on the water gauge is a scale mark used to represent a specific unit of length, and it is divided into 5 equal parts vertically. This method divides the water gauge into different equal parts based on the "E" and its reflection, assigning different categories. Then, it detects the category of the lowest target from top to bottom, thereby inferring the water surface boundary line. Figure 2 The blue "E" and red [marker] on the water level gauge in the example image on the left .
[0005] The above-mentioned image recognition-based water level gauge identification method has the following problems: 1. Gradient feature-based approach: Illumination, noise, aquatic plants, floating objects, waves, etc. can all cause significant interference to the gradient extraction of water surface segmentation lines. Moreover, it is difficult to control the parameters manually, resulting in low accuracy and unstable performance.
[0006] 2. Segmentation-based approach: This method requires a large amount of labeled data to train the water segmentation model. Different lighting conditions, water bodies, and transparency all pose significant challenges to the model, and the accuracy of the annotation directly affects the algorithm's performance. It involves a large workload and requires high annotation accuracy.
[0007] 3. Based on detection and classification: After the letter "E" is cut, there is only a horizontal line or a blank space, which contains too little information and is easily affected by floating objects. An "E" can be cut a maximum of 5 times, with a maximum accuracy of 0.2E. When the water is clear or has high transparency, the reflection of the "E" below or above the water surface can also cause the algorithm to produce false detections. Summary of the Invention
[0008] To address the difficulty of labeling segmented water surface data, water level gauge recognition requires precise marking of the water-shore boundary, which demands significant manpower and resources. This invention provides a water level gauge recognition method based on three-channel contrastive learning. Through contrastive learning, it learns the feature representations of water and non-water surfaces without requiring precise marking of the water surface boundary. Regarding the poor accuracy of water surface segmentation, current methods mainly use gradient projection or water area segmentation to obtain the water surface dividing line, which is highly susceptible to environmental factors and dataset quality. This invention achieves pixel-level segmentation accuracy through deep learning features and a sliding window approach. Specifically, this invention includes an offline feature extraction network training stage and an online water level gauge recognition stage. The training stage includes: The system collects monitoring images containing water surfaces, uses SAM for automatic segmentation, identifies all instance masks, repairs erroneous masks, and marks the water and non-water parts within the masks. It then obtains a set of water surface images and a set of non-water surface images as sample pairs. The original image features, frequency feature maps, and reflectance feature maps of the sample pairs are input into a feature extraction network, which is trained using a multi-task learning approach to obtain a trained model. The identification phase includes: acquiring images containing water level gauges, marking the four vertices of the water level gauges, and calibrating the water level gauges; identifying all numbers within the water level gauge area, finding the smallest number by sorting, recording the water level value and data frame information of the smallest number, and calculating the range of the water surface boundary area; taking image blocks within the water surface boundary area, with the center line of the image block sliding pixel by pixel from top to bottom within the water surface boundary area, calculating the three-channel input features of the image respectively, using the model to extract the feature vectors of the upper and lower halves of the image block, calculating the distance between them, and counting the maximum distance as the position of the water surface boundary, which is used as the water level value at this time.
[0009] Furthermore, the formula for calculating the frequency characteristic map is: Where M and N are the width and height of the image, respectively; x and y are the coordinates of a pixel; f(x,y) represents the gray value of the original image at coordinates (x,y); and u and v are the frequency coordinates, respectively. For frequency domain representation, the extracted frequency domain amplitude spectrum features are used as the frequency domain feature map, and the formula is: .
[0010] Furthermore, the formula for calculating reflection characteristics is as follows: ,in, , where L is obtained by convolving the original image I with the Gaussian filter G, x, y are the image coordinates, I is the original image, G is the Gaussian kernel function, L is the estimated illumination component, and R is the reflection feature map.
[0011] Furthermore, training using a multi-task learning approach includes: The feature extraction network comprises three branches, each employing either a CNN or a Transformer with pooling operations. The feature vectors from the three branches are concatenated to form a final concatenated feature vector, which is then compressed in dimension using a fully connected layer. For a given sample pair, the original image features, frequency domain feature map, and reflectance feature map are processed by the feature extraction network to extract the feature vectors from each branch and the compressed concatenated feature vector. The feature vectors from each branch are connected to a binary classification head, and a binary classification loss function is used to calculate the loss. The compressed concatenated feature vector is then subjected to a contrastive loss function for loss calculation. Backpropagation optimizes the feature extraction network. Sample labels are defined as follows: 0 if both samples are water surfaces or both are non-water surfaces, and 1 otherwise. The total loss for each branch consists of two parts: the binary classification loss for each branch and the contrast loss after splicing.
[0012] Furthermore, the binary classification loss function is the cross-entropy loss function, and the formula is: Where y is the sample label, The model predicts the category; The formula for the contrastive loss function is: , where y represents the value of the label, d represents the distance between the compressed concatenated feature vectors, M represents the maximum margin, and B represents the number of samples in each iteration.
[0013] Furthermore, the formula for calculating the distance between feature vectors is: , Let X and Y represent the values of the i-th dimension components of two feature vectors X and Y, respectively, where n represents the dimension of the feature vector and dist represents the distance between the two feature vectors.
[0014] Furthermore, the four vertices of the water gauge are manually marked as the upper left, upper right, lower left, and lower right corners, denoted as: ; The water gauge was calibrated, and the rotation angles of the two perpendicular sides were calculated separately. and The calculation formula is: The water gauge rotated at an angle of 100 degrees. The center coordinates are: The image is rotated and corrected by adjusting the rotation angle and center coordinates to obtain the corrected water level gauge image; the rotation formula is: Where x and y are the coordinates of the original image, respectively. These are the center coordinates, For rotation angle, These are the coordinates of the rotated image.
[0015] Furthermore, the calculated range of the water surface boundary area is as follows: The formula is: ; These are the coordinates of the four vertices of the number box.
[0016] Furthermore, the formula for calculating the water level at this point is: ,in, This indicates that the smallest number on the water gauge, which is closest to the water surface, has been detected. Let represent the upper and lower pixel coordinates of the bounding rectangle of the closest number to the water surface, respectively; let E represent the actual height of the water gauge, expressed in meters; the height of each pixel is: The initial value of the water surface boundary area is: ; count represents the number of times the distance to the maximum feature vector is traversed from top to bottom within the water surface boundary, and value represents the water level value at this time.
[0017] The present invention also provides a water level gauge recognition system based on three-channel contrastive learning, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute a water level gauge recognition method based on three-channel contrastive learning as described in the above technical solution.
[0018] This invention provides a water level indicator recognition method based on three-channel contrastive learning. By combining the frequency domain features and reflection features of an image, it fuses three-channel features to improve the model's feature extraction performance. The contrastive learning approach enhances the distinguishability between features in water surface areas and non-water surface areas. It exhibits strong robustness to environmental factors such as water transparency, reflections, water color, and floating debris. It has low resource consumption, a simple structure, and is highly suitable for widespread application. Furthermore, this invention obtains the water level line by using a sliding window and detecting the maximum feature distance between adjacent areas. The overall features of the upper and lower window areas are used as the comparison object, and the relative segmentation method forcibly classifies each pixel, resulting in better adaptability and effectively improving the detection accuracy of the water level line. It also demonstrates good adaptability to problems such as water surface fluctuations, blurred boundaries, transparent water, and interference from floating debris. Attached Figure Description
[0019] Figure 1 A flowchart for the overall process of water level gauge identification.
[0020] Figure 2 This is a schematic diagram of instance segmentation.
[0021] Figure 3 This is a schematic diagram of water surface and non-water surface samples. The left side shows the water surface sample, and the right side shows the non-water surface sample.
[0022] Figure 4 This is a schematic diagram of the three-channel feature extraction network model.
[0023] Figure 5 This is a diagram of the training framework for the feature extraction network.
[0024] Figure 6 This is a schematic diagram of the detection range for the water surface dividing line.
[0025] Figure 7 This is a schematic diagram of the water surface dividing line detection process.
[0026] Figure 8 This is a schematic diagram of the detection results for the water surface dividing line.
[0027] Figure 9 This is a comparison chart showing the effects of the method of this invention and the SAM segmentation method.
[0028] Figure 10 The figures represent experimental data from the method of this invention. Detailed Implementation
[0029] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings.
[0030] This invention provides a water level detection method based on three-channel contrastive learning. Combining frequency domain features and reflection features, the method integrates three-channel features to enrich the model's feature representation capabilities. Contrastive learning enables the model to distinguish between water surfaces and non-water surfaces, exhibiting strong robustness to environmental factors such as water transparency, reflections, water color, and floating objects. A sliding window extracts the overall features within the upper and lower window regions, using the maximum distance between these features as the water surface boundary. A relative segmentation method forcibly classifies each pixel, effectively improving the detection accuracy of the water level line and providing better adaptability. It also exhibits good adaptability to problems such as water surface fluctuations and blurred boundaries. The overall process is as follows: Figure 1 As shown, this method is divided into an offline feature extraction network model training stage and an online water level gauge recognition stage.
[0031] I. Offline Feature Extraction Network Model Training Phase: 1. Collect a large number of monitoring images containing water surfaces.
[0032] 2. Use SAM for automatic segmentation to find all instance masks, such as Figure 2 As shown.
[0033] 3. Manually repair erroneous masks and mark the water and non-water portions of the masks.
[0034] 4. Sample Acquisition: Following steps 2 and 3, a large number of images and corresponding water surface and non-water surface sample sets are obtained. Based on a mask, a large number of water surface and non-water surface image patches are randomly collected, forming water surface image sets and non-water surface image sets. Taking water surface image patches as an example, a square patch is randomly cropped from the water surface mask image. If the entire area is water, the image patch is valid; otherwise, it is discarded. Figure 3 As shown.
[0035] 5. A feature extraction network was built. Through analysis of a large amount of data, it was found that water surfaces and non-water surfaces have high distinguishability in terms of frequency features and reflection features. This invention adopts a three-channel feature fusion method, which can effectively improve the model's ability to distinguish between water surfaces and non-water surfaces. The three channels are the original image features, frequency features, and reflection features.
[0036] The formula for calculating frequency characteristics is: Where M and N are the width and height of the image, respectively; x and y are the coordinates of a pixel; f(x,y) represents the gray value of the original image at coordinates (x,y); and u and v are the frequency coordinates, respectively. For frequency domain representation, this invention extracts the frequency domain amplitude spectrum features as the frequency domain feature map, as shown in the formula: .
[0037] The formula for calculating reflection characteristics is: ,in, , where L is obtained by convolving the original image I with the Gaussian filter G, x, y are the image coordinates, I is the original image, G is the Gaussian kernel function, L is the estimated illumination component, and R is the reflection feature map.
[0038] This invention uses the original image, frequency feature map, and reflectance feature map as inputs to the feature extraction network, such as... Figure 4 As shown, this feature extraction network consists of three branches, each of which is a CNN or Transformer with pooling operations. The input sizes are 64x64x3, 64x64x1, and 64x64x1, respectively. After feature extraction, three 64-dimensional feature vectors are obtained. These feature vectors are then concatenated to form a 196-dimensional feature vector. Finally, the concatenated 196-dimensional feature vector is compressed to 64 dimensions by a fully connected layer.
[0039] 6. Model training framework, such as Figure 5As shown, a contrastive learning training framework is constructed, employing multi-task learning. The input is a sample pair, and frequency domain features and reflection features are extracted separately. The original image features, frequency domain feature map, and reflection feature map are processed by a feature extraction network to extract feature vectors for each branch and a compressed concatenated vector. Each channel branch vector is connected to a binary classification head, and the loss is calculated using a binary classification loss function. The loss of the compressed concatenated vector is calculated using a contrastive loss function. The feature extraction model is optimized through backpropagation. Sample labels are defined as follows: 0 if both samples are water surfaces or both are non-water surfaces, and 1 otherwise.
[0040] 7. Loss Function: This invention employs a multi-task learning approach. The loss function comprises two parts: a binary classification loss function for each branch and a concatenated contrastive loss function. Each branch is optimized independently. The loss for each branch is: ,in This represents the contrast loss of the splicing features. This represents the binary classification loss of this branch. i Indicates the branch number.
[0041] The binary classification loss function is the cross-entropy loss function, and the formula is: Where y is the sample label, The category predicted by the model.
[0042] The formula for the contrastive loss function is: Where y represents the label value, d represents the distance between the compressed concatenated feature vectors, M represents the maximum margin (default is 2.0), and B represents the number of samples in each iteration. When y=0, two samples are in the same state, and the loss is the distance between these two samples. The larger the distance, the greater the penalty, which is as expected, and vice versa. The benefit of this step is to calculate the training sample loss, maximize the spatial distance between samples in different states, and compress the spatial distance between samples in the same state, thus providing impetus for optimizing the network model parameters through loss calculation. The formula for calculating the distance between feature vectors is: , Let X and Y represent the values of the i-th dimension components of two feature vectors X and Y, respectively, where n represents the dimension of the feature vector and dist represents the distance between the two feature vectors.
[0043] 8. Model Training: This invention uses the method described in step 4 to create training and testing sample sets with a ratio of 8:2. The network model training framework is constructed using step 6, with a learning rate of 0.01. Training is performed for 60 epochs (all samples are iterated once) using gradient descent. The learning rate is multiplied by 0.1 at the 10th, 20th, and 40th epochs respectively. For each batch iteration, the loss of each branch is calculated, and all parameters of that branch are optimized based on the loss. The model's accuracy is verified on the test set after each epoch of training. The benefit of this step is training the feature extraction network model and optimizing network parameters through gradient descent and backpropagation.
[0044] II. Online water level gauge identification stage: 1. Acquire images containing the water level gauge captured by the camera.
[0045] 2. The four vertices of the water gauge are manually marked as the upper left, upper right, lower left, and lower right corners, denoted as: .
[0046] 3. Adjust the water gauge and calculate the rotation angle of the two perpendicular sides respectively. and The calculation formula is: The water gauge rotated at an angle of 100 degrees. The center coordinates are: The image is rotated and corrected by adjusting the rotation angle and center coordinates to obtain the corrected water level gauge image. The rotation formula is: Where x and y are the coordinates of the original image, respectively. These are the center coordinates, For rotation angle, These are the coordinates of the rotated image.
[0047] 4. Use Paddle-OCR to identify all digits within the watermark area, find the smallest digit by sorting them, and record the detection information for that digit, including: digit... (This indicates the water level value corresponding to the lowest number detected on the water gauge, for example...) Figure 6 middle, This indicates 48 meters), the number frame (outer rectangle). ).
[0048] 5. Based on the data frame in step 4, calculate the range of the water surface boundary area as follows: The formula is: .like Figure 6 As shown. The water surface boundary area extends to a height of 3E, and the height of each pixel is: The initial value of the water surface boundary area is: .
[0049] 6. Within the water surface boundary obtained in step 5, take an image patch. The centerline of the image patch slides pixel by pixel from top to bottom within the water surface boundary. Calculate the three-channel input features of the image, and use the original features, frequency features, and reflection features as the three-channel input. Use a feature extraction network to extract the feature vectors of the upper and lower halves of the image patch. (These are all the final concatenated feature vectors), and the distance between the feature vectors of the upper and lower halves is calculated using the following formula: , Let represent the values of the i-th dimension components of two feature vectors X and Y, respectively, where n represents the dimension of the feature vector. For example... Figure 7 As shown.
[0050] 7. Based on the characteristic distances obtained from top to bottom within the water surface boundary range in step 6, the maximum distance is the location of the water surface boundary, which is: count. This count represents the position with the largest distance from top to bottom within the water surface boundary range. Figure 8 As shown.
[0051] 8. Based on the results of steps 4 and 7, the water level can be calculated. Assuming E is 0.5 meters, the formula for calculating the water level is: ,in, This indicates that the OCR detected the number on the water gauge closest to the water surface. represents the upper and lower pixel coordinates of the bounding rectangle of the closest number to the water surface, E represents the actual height of the water gauge (E represents the height of the water gauge in meters), count represents the number of times the maximum feature vector is traversed from top to bottom within the water surface boundary, and value represents the water level value at this time.
[0052] Final effect comparison image as follows Figure 9 As shown, the experimental group is as follows Figure 10 As shown in the comparison chart, it can be seen that segmentation-based methods fail to segment blurry or transparent water bodies, but this method can reliably find accurate boundary lines. The experimental group shows that this method can accurately detect water level lines in various environments, including lighting, transparency, and blurriness, demonstrating good adaptability.
[0053] On the other hand, embodiments of the present invention also provide a water level gauge recognition system based on three-channel contrastive learning, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute a water level gauge recognition method based on three-channel contrastive learning as described in the above technical solution.
[0054] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A water level gauge recognition method based on three-channel contrastive learning, characterized in that: It includes an offline feature extraction network training phase and an online water level gauge recognition phase, wherein the training phase includes: The system collects monitoring images containing water surfaces, uses SAM for automatic segmentation, identifies all instance masks, repairs erroneous masks, and marks the water and non-water parts within the masks. It then obtains a set of water surface images and a set of non-water surface images as sample pairs. The original image features, frequency feature maps, and reflectance feature maps of the sample pairs are input into a feature extraction network, which is trained using a multi-task learning approach to obtain a trained model. The identification phase includes: acquiring images containing water level gauges, marking the four vertices of the water level gauges, and calibrating the water level gauges; identifying all numbers within the water level gauge area, finding the smallest number by sorting, recording the water level value and data frame information of the smallest number, and calculating the range of the water surface boundary area; taking image blocks within the water surface boundary area, with the center line of the image block sliding pixel by pixel from top to bottom within the water surface boundary area, calculating the three-channel input features of the image respectively, using the model to extract the feature vectors of the upper and lower halves of the image block, calculating the distance between them, and counting the maximum distance as the position of the water surface boundary, which is used as the water level value at this time.
2. The water level gauge recognition method based on three-channel contrastive learning as described in claim 1, characterized in that: The formula for calculating the frequency characteristic map is: Where M and N are the width and height of the image, respectively; x and y are the coordinates of a pixel; f(x,y) represents the gray value of the original image at coordinates (x,y); and u and v are the frequency coordinates, respectively. For frequency domain representation, the extracted frequency domain amplitude spectrum features are used as the frequency domain feature map, and the formula is: .
3. The water level gauge recognition method based on three-channel contrastive learning as described in claim 1, characterized in that: The formula for calculating reflection characteristics is: ,in, , where L is obtained by convolving the original image I with the Gaussian filter G, x, y are the image coordinates, I is the original image, G is the Gaussian kernel function, L is the estimated illumination component, and R is the reflection feature map.
4. The water level gauge recognition method based on three-channel contrastive learning as described in claim 1, characterized in that: Training using a multi-task learning approach includes: The feature extraction network comprises three branches, each employing either a CNN or a Transformer with pooling operations. The feature vectors from the three branches are concatenated to form a final concatenated feature vector, which is then compressed in dimension using a fully connected layer. For a given sample pair, the original image features, frequency domain feature map, and reflectance feature map are processed by the feature extraction network to extract the feature vectors from each branch and the compressed concatenated feature vector. The feature vectors from each branch are connected to a binary classification head, and a binary classification loss function is used to calculate the loss. The compressed concatenated feature vector is then subjected to a contrastive loss function for loss calculation. Backpropagation optimizes the feature extraction network. Sample labels are defined as follows: 0 if both samples are water surfaces or both are non-water surfaces, and 1 otherwise. The total loss for each branch consists of two parts: the binary classification loss for each branch and the contrast loss after splicing.
5. The water level gauge recognition method based on three-channel contrastive learning as described in claim 4, characterized in that: The binary classification loss function is the cross-entropy loss function, and the formula is: Where y is the sample label, The model predicts the category; The formula for the contrastive loss function is: , where y represents the value of the label, d represents the distance between the compressed concatenated feature vectors, M represents the maximum margin, and B represents the number of samples in each iteration.
6. The water level gauge recognition method based on three-channel contrastive learning as described in claim 5, characterized in that: The formula for calculating the distance between feature vectors is: , Let X and Y represent the values of the i-th dimension components of two feature vectors X and Y, respectively, where n represents the dimension of the feature vector and dist represents the distance between the two feature vectors.
7. The water level gauge recognition method based on three-channel contrastive learning as described in claim 1, characterized in that: The four vertices of the manually calibrated water gauge are designated as the upper left, upper right, lower left, and lower right corners, and are denoted as: ; The water gauge was calibrated, and the rotation angles of the two perpendicular sides were calculated separately. and The calculation formula is: The water gauge rotated at an angle of 100 degrees. The center coordinates are: The image is rotated and corrected by adjusting the rotation angle and center coordinates to obtain the corrected water level gauge image; the rotation formula is: Where x and y are the coordinates of the original image, respectively. These are the center coordinates, For rotation angle, These are the coordinates of the rotated image.
8. The water level gauge recognition method based on three-channel contrastive learning as described in claim 1, characterized in that: The calculated area of the water surface boundary is: The formula is: ; These are the coordinates of the four vertices of the number box.
9. The water level gauge recognition method based on three-channel contrastive learning as described in claim 1, characterized in that: The formula for calculating the water level at this time is: ,in, This indicates that the smallest number on the water gauge, which is closest to the water surface, has been detected. Let represent the upper and lower pixel coordinates of the bounding rectangle of the closest number to the water surface, respectively; let E represent the actual height of the water gauge, expressed in meters; the height of each pixel is: The initial value of the water surface boundary area is: ; count represents the number of times the distance to the maximum feature vector is traversed from top to bottom within the water surface boundary, and value represents the water level value at this time.
10. A water level gauge recognition system based on three-channel contrastive learning, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute the water level gauge recognition method based on three-channel contrast learning as described in any one of claims 1-9.