Hydropower plant instrument panel pointer reading detection method based on deep learning vision algorithm

Abstract

The invention relates to application of a deep learning image recognition technology in the field of industrial automation, and discloses a method for detecting pointer reading of an instrument panel of a hydraulic power plant by using a deep learning vision algorithm, which comprises the following steps: S1, constructing a platform and acquiring basic data information; s2, a deep convolution step; s3, carrying out fusion re-parameter convolution; s4, a channel attention mechanism is established; s5, carrying out image segmentation; and S6, instrument positioning and data reading are carried out. The method is particularly suitable for being used in a high-risk, complex and changeable industrial field environment, accurate positioning and data reading of the pointer type instrument are achieved, and the inspection efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the application of deep learning image recognition technology in the field of industrial automation, and in particular to a method for detecting pointer readings of a hydropower plant instrument panel using a deep learning visual algorithm. Background Art

[0002] In the operation monitoring of hydropower plants, the accuracy of instrument panel pointer readings is crucial to ensuring equipment safety and optimizing production processes. Traditionally, this task mainly relies on manual inspections, but this method has several limitations. First, manual readings are inefficient and cannot meet the needs of modern industry for real-time monitoring. Second, manual operations are susceptible to factors such as fatigue and distraction, which lead to reading errors and data inconsistencies, increasing production risks. In addition, manual inspections are difficult to achieve all-weather continuous monitoring, especially in conditions such as bad weather or at night, and the reliability and safety of readings are difficult to guarantee.

[0003] With the development of image recognition technology, automated instrument panel pointer reading detection has become possible. Early image recognition technology was mainly based on traditional computer vision methods, such as edge detection, feature matching, and template matching. These methods can achieve certain results under ideal conditions, but in practical applications, due to the complexity of the hydropower plant environment, such as light changes, physical damage to the instrument panel, occlusion and reflection of the pointer, the limitations of these traditional methods become particularly obvious. For example, edge detection methods are sensitive to lighting conditions, feature matching methods are prone to errors when facing pointer occlusion or similar patterns, and template matching methods require a large amount of template data and are difficult to adapt to different instrument panel types and states.

[0004] In addition, with the emergence of deep learning technology, especially the application of convolutional neural networks (CNN) in the field of image recognition, although the accuracy of readings has been improved to a certain extent, the existing algorithms are still insufficient in terms of real-time performance and accuracy. On the one hand, deep learning models usually require a lot of computing resources, which may cause the processing speed to fail to keep up with actual needs for real-time monitoring systems. On the other hand, deep learning models still have the problem of insufficient generalization ability when facing complex backgrounds and diverse dashboard designs, resulting in limited accuracy in practical applications.

[0005] In the field of target detection, although the currently popular single-step target detection algorithms have been significantly improved compared to traditional methods, these algorithms still face challenges in the specific application of pointer readings on hydropower plant dashboards. For example, these algorithms may perform poorly when dealing with small targets or targets in complex backgrounds, or have low computational efficiency when dealing with high-resolution images. In addition, these algorithms require a large amount of labeled data during the training process, which is often time-consuming and costly to obtain in actual industrial applications. Although image recognition technology has made certain progress in the field of pointer reading detection on hydropower plant dashboards, existing methods still have limitations in terms of accuracy, real-time performance, robustness, and adaptability. These limitations limit the application effect of automated reading systems in actual industrial environments and also affect the improvement of the intelligence level of hydropower plants. Summary of the invention

[0006] The technical problem to be solved by the present invention is to propose a computer vision-based intelligent inspection robot instrument positioning and data reading method to overcome the shortcomings of the existing technology, especially in high-risk and complex and changeable industrial field environments, to achieve precise positioning and data reading of pointer instruments and improve inspection efficiency and accuracy.

[0007] In order to solve the above technical problems, the present invention proposes a computer vision-based intelligent inspection robot instrument positioning and data reading method, which is characterized by comprising the following steps:

[0008] Step S1: Platform construction and basic data information collection;

[0009] S2 step: depth convolution step;

[0010] Step S3: Fusion of heavy parameter convolution;

[0011] S4 step: channel attention mechanism;

[0012] Step S5: image segmentation;

[0013] Step S6: Instrument positioning and data reading.

[0014] More specifically, the present invention proposes a computer vision-based intelligent inspection robot instrument positioning and data reading method, comprising the following steps:

[0015] S11: Hardware equipment construction platform. In the intelligent inspection robot system, the hardware platform is first built, including a four-axis wheeled explosion-proof inspection robot, an edge computing terminal, a high-definition camera, and a level to adapt to the complex environment on site and ensure the accuracy of image acquisition.

[0016] S12: Image data acquisition.

[0017] The built platform is used to collect high-definition images of pointer instruments on site. These images cover instruments of different types and states, providing basic data for subsequent algorithm training and testing.

[0018] S13: Data Annotation

[0019] The captured images are carefully annotated, including key information such as instrument location, pointers and scales, in order to train and optimize the target detection and image segmentation algorithms to ensure that they can accurately identify and learn instrument features.

[0020] S14: Establish a data set. Based on the annotation results, construct a pointer instrument positioning and segmentation data set for algorithm training and performance evaluation to ensure high-precision instrument positioning and data reading in actual industrial scenarios.

[0021] S21: The spatial-to-depth convolution module consists of spatial-to-depth layers and non-stride convolution layers, which are used to replace the stride convolution and pooling layers in traditional aggregation neural networks. First, the spatial dimension of the feature map is converted to the depth dimension through spatial-to-depth conversion, which is achieved by halving the height and width of the feature map and doubling the number of channels. Then, non-stride convolution layers are applied to keep the spatial resolution unchanged and reduce the number of channels.

[0022]

[0023]

[0024] S22: The spatial-to-depth layer is part of the spatial-to-depth convolution module. Its core is to rearrange the spatial information of the feature map to the depth channel. Specifically, the conversion of spatial information to depth information is achieved by downsampling the input feature map, that is, halving the height and width of the feature map and doubling the number of channels.

[0025] S23: Implementation steps of non-strided convolutional layer. After the SPD layer, a non-strided convolutional layer, i.e., a convolution with a stride of 1, is applied to reduce the channel dimension while keeping the spatial resolution unchanged. The purpose of this step is to further extract features and reduce the dimensionality of the feature map without losing spatial information.

[0026] S24: The improvement of the popular single-step object detection algorithm is mainly reflected in the introduction of the SPD-Conv module to replace the original stride convolution and pooling layers. In the network structure of YOLOv8, this is the most significant change by adding SPD operations after multiple layers. SPD operations are usually used to increase the number of channels while reducing the spatial size of feature maps.

[0027] S31: Fusion of re-parameterized convolution structure based on channel shuffle First, it is necessary to introduce re-parameterized convolution based on channel shuffle (re-parameterized convolution) and one-time aggregation module into the network. The channel shuffle module learns rich feature information through a multi-branch structure during the training phase, and simplifies it to a single-branch structure during the inference phase to reduce memory consumption. The one-time aggregation module reduces the network computing burden and improves computing efficiency by aggregating multiple feature cascades at one time.

[0028]

[0029]

[0030]

[0031] S32: The re-parameterized convolutional module for channel shuffling uses multiple branches, including 1x1 and 3x3 convolutions, and a direct connection during the training phase to learn rich feature representations. During the inference phase, these structures are re-parameterized into a single 3x3 convolution to reduce computational complexity and memory consumption while maintaining the feature expression capabilities learned during the training phase.

[0032] S33: One-shot aggregation structure reduces the network computation burden and improves computational efficiency by aggregating multiple feature cascades at one time. In the channel-shuffled re-parameterized convolution, the one-shot aggregation module is combined with the channel-shuffled re-parameterized convolution to form a channel-shuffled re-parameterized convolution module, which not only maintains low-cost memory consumption, but also realizes the effective extraction of semantic information.

[0033]

[0034]

[0035]

[0036] S34: Feature cascade technology is implemented by maintaining a limited number of feature cascades on the one-time aggregation path of the network. In the channel-shuffled re-parameterized convolution, especially in the channel-shuffled re-parameterized convolution module, only three feature cascades are retained, which helps to reduce the network computational burden and memory usage. This method effectively aggregates features at different levels and improves the semantic information extraction ability of the model, while avoiding the inefficiency and high resource consumption caused by over-complex network structure.

[0037] S41: The channel-spatial attention module enhances the model's attention to the pointer-type instrument images by integrating channel and spatial attention mechanisms in a feed-forward convolutional neural network. First, the channel-spatial attention module is embedded in a key position of the network to introduce the attention mechanism during feature extraction. This integration helps the model focus more on the instrument pointer during detection while improving resistance to interference.

[0038] S42: The channel attention module extracts global features by performing global maximum pooling and global average pooling on the feature map of each channel. These features are fed into a dimensionally reduced MLP to output the attention weights of each channel. These weights are obtained through the Sigmoid activation function and are used to weight the channels of the original feature map to strengthen the features of important channels.

[0039]

[0040] S43: The spatial attention module implements spatial attention by calculating the importance of each pixel. It first performs maximum pooling and average pooling on the feature map after channel aggregation, then fuses these features through a convolutional layer, and finally outputs the spatial attention weight of each pixel through the Sigmoid activation function to emphasize the key areas in the image.

[0041]

[0042] S44: The fusion of channel and spatial attention weights is achieved by multiplying the outputs of the two element-wise to obtain the final attention-enhanced feature map. This enhanced feature map will be used as the input of subsequent network layers to retain key information and suppress irrelevant noise. This attention mechanism improves the model's recognition ability for pointer instrument images, especially in the presence of complex backgrounds or interference.

[0043] S51: U2-Net-DF is an improved image segmentation network that achieves high-resolution image segmentation through a deeply nested U-shaped structure. This network design allows the depth and accuracy of the model to be improved without adding too much computational burden. In the detection of instrument panel pointer readings in hydropower plants, U2-Net-DF is used to process the located instrument images to achieve accurate image segmentation.

[0044] S52: After locating the instrument needle, the next step is to perform image segmentation on the instrument's region of interest. The U2-Net-DF network receives these RoI images as input and uses its multi-scale feature extraction capabilities to analyze the image layer by layer to distinguish the instrument needle from the background. This step is key to achieving accurate segmentation because it involves separating the instrument needle from the complex background.

[0045] S53: The U2-Net-DF network effectively extracts target pixels from the background through its deep structure and feature fusion mechanism. The network's upsampling process and edge detail recovery capabilities ensure accurate segmentation of the pointer and scale. This step is crucial for subsequent reading calculations because it provides accurate position information of the pointer relative to the scale.

[0046]

[0047] S54: After the segmentation is completed, the output of U2-Net-DF is the pixel-level data of the pointer and scale. This data includes the precise position and shape of the pointer, as well as the distribution of the scale. This information provides a basis for the automatic reading of instrument values, allowing values ​​to be directly extracted from the image without manual intervention, improving the automation and accuracy of the instrument panel reading in the hydropower plant.

[0048] S61: Before polar coordinate transformation, the hydropower plant instrument panel image is preprocessed, including denoising, contrast enhancement and image distortion correction. This step ensures that the pointer and scale lines in the image are clearly visible, laying the foundation for subsequent polar coordinate transformation and data reading. Preprocessing is a key step to improve the accuracy of image analysis and help reduce the impact of environmental factors on readings.

[0049] S62: Use the polar coordinate transformation function in the image processing library to transform the preprocessed image from the rectangular coordinate system to the polar coordinate system. This step involves setting the size of the output image, the transformation center point, and the radius of the boundary circle. The polar coordinate transformation can map the circular instrument panel surface to a rectangular area, simplifying the process of extracting the position information of the pointer and scale.

[0050] S63: In the process of polar coordinate transformation, it is necessary to select appropriate interpolation methods and polar coordinate mapping methods. For example, linear interpolation and polar coordinate forward transformation are used. The selection of these parameters will affect the effect and performance of the transformation. Correct parameter settings can ensure that the transformed image maintains high quality and provide an accurate basis for subsequent data extraction.

[0051]

[0052] S64: After the transformation, the position information of the pointer and scale is converted into radial distance and angle from the center point in the polar coordinate system. This step simplifies the data reading process because the pixel-level data of the pointer and scale are easier to extract, which facilitates the accurate reading of the instrument value.

[0053] S65: In the polar coordinate system, by measuring the radial distance and angle from the end point of the pointer to the center point, the value of the instrument can be accurately read. This image dimensionality reduction method based on polar coordinate transformation can effectively extract the target pixel from the background and obtain the pixel-level data of the segmented pointer and scale, realizing accurate and effective data reading of pointer-type instruments.

[0054] The present invention is particularly suitable for use in highly dangerous and complex and changeable industrial field environments, realizing accurate positioning and data reading of pointer-type instruments, and improving inspection efficiency and accuracy. DETAILED DESCRIPTION

[0055] The following deficiencies exist in the existing hydropower plant instrument panel pointer reading detection method:

[0056] 1. Traditional methods rely on manual measurement or simple image processing technology, which has low accuracy and is easily affected by ambient light and dashboard wear, resulting in large reading errors.

[0057] 2. Existing machine vision systems often require complex preprocessing and feature extraction steps. They are not flexible enough to recognize dynamically changing pointer positions and scales, and are difficult to adapt to dashboards of different types and states.

[0058] 3. In complex industrial environments, existing technologies make it difficult to achieve real-time monitoring and accurate reading of instrument panel pointers, especially under working conditions of high humidity, high temperature and strong vibration.

[0059] 4. The reading detection system based on traditional deep learning algorithms has high training and deployment costs, poor adaptability to small samples and specific scenarios, and is difficult to generalize to different hydropower plant environments.

[0060] In order to overcome the defects or shortcomings of the above-mentioned prior art, the present invention proposes a method for detecting the pointer reading of a hydropower plant instrument panel based on a deep learning visual algorithm, which includes the following main steps:

[0061] S11: Hardware equipment construction platform. In the intelligent inspection robot system, the hardware platform is first built, including a four-axis wheeled explosion-proof inspection robot, an edge computing terminal, a high-definition camera, and a level to adapt to the complex environment on site and ensure the accuracy of image acquisition.

[0062] S12: Image data acquisition.

[0063] The built platform is used to collect high-definition images of pointer instruments on site. These images cover instruments of different types and states, providing basic data for subsequent algorithm training and testing.

[0064] S13: Data Annotation

[0065] The captured images are carefully annotated, including key information such as instrument location, pointers and scales, in order to train and optimize the target detection and image segmentation algorithms to ensure that they can accurately identify and learn instrument features.

[0066] S14: Establish a data set. Based on the annotation results, construct a pointer instrument positioning and segmentation data set for algorithm training and performance evaluation to ensure high-precision instrument positioning and data reading in actual industrial scenarios.

[0067] S21: The spatial-to-depth convolution module consists of spatial-to-depth layers and non-stride convolution layers, which are used to replace the stride convolution and pooling layers in traditional aggregation neural networks. First, the spatial dimension of the feature map is converted to the depth dimension through spatial-to-depth conversion, which is achieved by halving the height and width of the feature map and doubling the number of channels. Then, non-stride convolution layers are applied to keep the spatial resolution unchanged and reduce the number of channels.

[0068]

[0069]

[0070] S22: The spatial-to-depth layer is part of the spatial-to-depth convolution module. Its core is to rearrange the spatial information of the feature map to the depth channel. Specifically, the conversion of spatial information to depth information is achieved by downsampling the input feature map, that is, halving the height and width of the feature map and doubling the number of channels.

[0071] S23: Implementation steps of non-strided convolutional layer. After the SPD layer, a non-strided convolutional layer, i.e., a convolution with a stride of 1, is applied to reduce the channel dimension while keeping the spatial resolution unchanged. The purpose of this step is to further extract features and reduce the dimensionality of the feature map without losing spatial information.

[0072] S24: The improvement of the popular single-step object detection algorithm is mainly reflected in the introduction of the SPD-Conv module to replace the original stride convolution and pooling layers. In the network structure of YOLOv8, this is the most significant change by adding SPD operations after multiple layers. SPD operations are usually used to increase the number of channels while reducing the spatial size of feature maps.

[0073] S31: Fusion of re-parameterized convolution structure based on channel shuffle First, it is necessary to introduce re-parameterized convolution based on channel shuffle (re-parameterized convolution) and one-time aggregation module into the network. The channel shuffle module learns rich feature information through a multi-branch structure during the training phase, and simplifies it to a single-branch structure during the inference phase to reduce memory consumption. The one-time aggregation module reduces the network computing burden and improves computing efficiency by aggregating multiple feature cascades at one time.

[0074]

[0075]

[0076]

[0077] S32: The re-parameterized convolutional module for channel shuffling uses multiple branches, including 1x1 and 3x3 convolutions, and a direct connection during the training phase to learn rich feature representations. During the inference phase, these structures are re-parameterized into a single 3x3 convolution to reduce computational complexity and memory consumption while maintaining the feature expression capabilities learned during the training phase.

[0078] S33: One-shot aggregation structure reduces the network computation burden and improves computational efficiency by aggregating multiple feature cascades at one time. In the channel-shuffled re-parameterized convolution, the one-shot aggregation module is combined with the channel-shuffled re-parameterized convolution to form a channel-shuffled re-parameterized convolution module, which not only maintains low-cost memory consumption, but also realizes the effective extraction of semantic information.

[0079]

[0080]

[0081]

[0082] S34: Feature cascade technology is implemented by maintaining a limited number of feature cascades on the one-time aggregation path of the network. In the channel-shuffled re-parameterized convolution, especially in the channel-shuffled re-parameterized convolution module, only three feature cascades are retained, which helps to reduce the network computational burden and memory usage. This method effectively aggregates features at different levels and improves the semantic information extraction ability of the model, while avoiding the inefficiency and high resource consumption caused by over-complex network structure.

[0083] S41: The channel-spatial attention module enhances the model's attention to the pointer-type instrument images by integrating channel and spatial attention mechanisms in a feed-forward convolutional neural network. First, the channel-spatial attention module is embedded in a key position of the network to introduce the attention mechanism during feature extraction. This integration helps the model focus more on the instrument pointer during detection while improving resistance to interference.

[0084] S42: The channel attention module extracts global features by performing global maximum pooling and global average pooling on the feature map of each channel. These features are fed into a dimensionally reduced MLP to output the attention weights of each channel. These weights are obtained through the Sigmoid activation function and are used to weight the channels of the original feature map to strengthen the features of important channels.

[0085]

[0086] S43: The spatial attention module implements spatial attention by calculating the importance of each pixel. It first performs maximum pooling and average pooling on the feature map after channel aggregation, then fuses these features through a convolutional layer, and finally outputs the spatial attention weight of each pixel through the Sigmoid activation function to emphasize the key areas in the image.

[0087]

[0088] S44: The fusion of channel and spatial attention weights is achieved by multiplying the outputs of the two element-wise to obtain the final attention-enhanced feature map. This enhanced feature map will be used as the input of subsequent network layers to retain key information and suppress irrelevant noise. This attention mechanism improves the model's recognition ability for pointer instrument images, especially in the presence of complex backgrounds or interference.

[0089] S51: U2-Net-DF is an improved image segmentation network that achieves high-resolution image segmentation through a deeply nested U-shaped structure. This network design allows the depth and accuracy of the model to be improved without adding too much computational burden. In the detection of instrument panel pointer readings in hydropower plants, U2-Net-DF is used to process the located instrument images to achieve accurate image segmentation.

[0090] S52: After locating the instrument needle, the next step is to perform image segmentation on the instrument's region of interest. The U2-Net-DF network receives these RoI images as input and uses its multi-scale feature extraction capabilities to analyze the image layer by layer to distinguish the instrument needle from the background. This step is key to achieving accurate segmentation because it involves separating the instrument needle from the complex background.

[0091] S53: The U2-Net-DF network effectively extracts target pixels from the background through its deep structure and feature fusion mechanism. The network's upsampling process and edge detail recovery capabilities ensure accurate segmentation of the pointer and scale. This step is crucial for subsequent reading calculations because it provides accurate position information of the pointer relative to the scale.

[0092]

[0093] S54: After the segmentation is completed, the output of U2-Net-DF is the pixel-level data of the pointer and scale. This data includes the precise position and shape of the pointer, as well as the distribution of the scale. This information provides a basis for the automatic reading of instrument values, allowing values ​​to be directly extracted from the image without manual intervention, improving the automation and accuracy of the instrument panel reading in the hydropower plant.

[0094] S61: Before polar coordinate transformation, the hydropower plant instrument panel image is preprocessed, including denoising, contrast enhancement and image distortion correction. This step ensures that the pointer and scale lines in the image are clearly visible, laying the foundation for subsequent polar coordinate transformation and data reading. Preprocessing is a key step to improve the accuracy of image analysis and help reduce the impact of environmental factors on readings.

[0095] S62: Use the polar coordinate transformation function in the image processing library to transform the preprocessed image from the rectangular coordinate system to the polar coordinate system. This step involves setting the size of the output image, the transformation center point, and the radius of the boundary circle. The polar coordinate transformation can map the circular instrument panel surface to a rectangular area, simplifying the process of extracting the position information of the pointer and scale.

[0096] S63: In the process of polar coordinate transformation, it is necessary to select appropriate interpolation methods and polar coordinate mapping methods. For example, linear interpolation and polar coordinate forward transformation are used. The selection of these parameters will affect the effect and performance of the transformation. Correct parameter settings can ensure that the transformed image maintains high quality and provide an accurate basis for subsequent data extraction.

[0097]

[0098] S64: After the transformation, the position information of the pointer and scale is converted into radial distance and angle from the center point in the polar coordinate system. This step simplifies the data reading process because the pixel-level data of the pointer and scale are easier to extract, which facilitates the accurate reading of the instrument value.

[0099] S65: In the polar coordinate system, by measuring the radial distance and angle from the end point of the pointer to the center point, the value of the instrument can be accurately read. This image dimensionality reduction method based on polar coordinate transformation can effectively extract the target pixel from the background and obtain the pixel-level data of the segmented pointer and scale, realizing accurate and effective data reading of pointer-type instruments.

[0100] Finally, it should be noted that the above specific implementation methods are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A method for positioning and reading data of an intelligent inspection robot instrument based on computer vision, characterized in that: The following steps are involved: Step S1: Platform construction and basic data information collection; S2 step: depth convolution step; Step S3: Fusion of heavy parameter convolution; S4 step: channel attention mechanism; Step S5: image segmentation; Step S6: Instrument positioning and data reading.

2. According to the computer vision-based intelligent inspection robot instrument positioning and data reading method of claim 1, it is characterized by: The S1 step includes the following sub-steps: S11: Hardware equipment construction platform. In the intelligent inspection robot system, the hardware platform is first built, including a four-axis wheeled explosion-proof inspection robot, an edge computing terminal, a high-definition camera, and a level to adapt to the complex environment on site and ensure the accuracy of image acquisition; S12: Image data acquisition: Use the built platform to collect high-definition images of pointer instruments on site. These images cover instruments of different types and states, providing basic data for subsequent algorithm training and testing; S13: Data annotation: Detailed annotation of the collected images, including key information such as instrument location, pointer and scale, in order to train and optimize the object detection and image segmentation algorithms to ensure that they can accurately identify and learn instrument features; S14: Establish a data set. Based on the annotation results, construct a pointer instrument positioning and segmentation data set for algorithm training and performance evaluation to ensure high-precision instrument positioning and data reading in actual industrial scenarios.

3. The method for positioning and reading data of an intelligent inspection robot based on computer vision according to claim 2 is characterized in that: The S2 step includes the following sub-steps: S21: The spatial-to-depth convolution module consists of a spatial-to-depth layer and a non-strided convolution layer, which is used to replace the strided convolution and pooling layers in traditional aggregation neural networks. First, the spatial dimension of the feature map is converted to the depth dimension through spatial-to-depth conversion. This step is achieved by halving the height and width of the feature map and doubling the number of channels. Then, a non-strided convolution layer is applied to keep the spatial resolution unchanged and reduce the number of channels: ； ； S22: The spatial-to-depth layer is part of the spatial-to-depth convolution module. Its core is to rearrange the spatial information of the feature map to the depth channel. Specifically, the input feature map is downsampled, that is, the height and width of the feature map are halved, and the number of channels is doubled, so as to realize the conversion of spatial information to depth information. S23: Implementation step of non-strided convolutional layer. After the SPD layer, a non-strided convolutional layer, i.e., a convolution with a stride of 1, is applied to reduce the channel dimension while keeping the spatial resolution unchanged. The purpose of this step is to further extract features and reduce the dimension of the feature map without losing spatial information. S24: The improvement of the popular single-step object detection algorithm is mainly reflected in the introduction of the SPD-Conv module to replace the original stride convolution and pooling layers. In the network structure of YOLOv8, the most significant change is to add SPD operations after multiple layers. SPD operations are usually used to increase the number of channels while reducing the spatial size of the feature map.

4. The method for positioning and reading data of an intelligent inspection robot based on computer vision according to claim 3 is characterized in that: The S3 step includes the following sub-steps: S31: Fusion of heavy parameter convolution structure based on channel shuffle First, it is necessary to introduce heavy parameter convolution based on channel shuffle and one-time aggregation module in the network. The channel shuffle module learns rich feature information through a multi-branch structure during the training phase, and simplifies it to a single branch structure during the inference phase to reduce memory consumption. The one-time aggregation module reduces the network calculation burden and improves the calculation efficiency by aggregating multiple feature cascades at one time: ； ； ； S32: The channel-shuffled re-parameterized convolution module uses multiple branches during training, including 1x1 and 3x3 convolutions, and a direct connection to learn rich feature representations. During inference, these structures are re-parameterized into a single 3x3 convolution to reduce computational complexity and memory consumption while maintaining the feature expression capabilities learned during training. S33: The one-time aggregation structure reduces the network computing burden and improves computing efficiency by aggregating multiple feature cascades at one time. In the channel-shuffled heavy parameter convolution, the one-time aggregation module is combined with the channel-shuffled heavy parameter convolution to form a channel-shuffled heavy parameter convolution module, which not only maintains low-cost memory consumption, but also realizes the effective extraction of semantic information: ； ； ； S34: Feature cascade technology is implemented by maintaining a limited number of feature cascades on the one-time aggregation path of the network. In the channel-shuffled re-parameterized convolution, especially in the channel-shuffled re-parameterized convolution module, only three feature cascades are retained, which helps to reduce the network computational burden and memory usage. This method effectively aggregates features at different levels and improves the model's ability to extract semantic information, while avoiding the inefficiency and high resource consumption caused by over-complex network structure.

5. The method for positioning and reading data of an intelligent inspection robot based on computer vision according to claim 4 is characterized in that: The S4 step includes the following sub-steps: S41: The channel-spatial attention module enhances the model's attention to the pointer-type instrument images by integrating channel and spatial attention mechanisms in a feed-forward convolutional neural network. First, the channel-spatial attention module is embedded in a key position of the network to introduce the attention mechanism during feature extraction. This integration helps the model focus more on the instrument pointer during detection and improves resistance to interference. S42: The channel attention module extracts global features by performing global maximum pooling and global average pooling on the feature map of each channel. These features are fed into a dimensionally reduced MLP and output the attention weights of each channel. These weights are obtained through the Sigmoid activation function and used to weight the channels of the original feature map to strengthen the features of important channels: ； S43: The spatial attention module implements spatial attention by calculating the importance of each pixel. It first performs maximum pooling and average pooling on the feature map after channel aggregation, then fuses these features through a convolutional layer, and finally outputs the spatial attention weight of each pixel through the Sigmoid activation function to emphasize the key areas in the image: ； S44: The fusion of channel and spatial attention weights is achieved by multiplying their outputs element-wise to obtain the final attention-enhanced feature map. This enhanced feature map will be used as the input of subsequent network layers to retain key information and suppress irrelevant noise. This attention mechanism improves the model's recognition ability for pointer instrument images, especially in the presence of complex backgrounds or interference.

6. The method for positioning and reading data of an intelligent inspection robot based on computer vision according to claim 5 is characterized in that: Step S5 includes the following sub-steps: S51: U2-Net-DF is an improved image segmentation network that achieves high-resolution image segmentation through a deeply nested U-shaped structure. This network design allows the depth and accuracy of the model to be improved without increasing excessive computational burden. In the detection of instrument panel pointer readings in hydropower plants, U2-Net-DF is used to process the located instrument images to achieve accurate image segmentation. S52: After locating the instrument needle, the next step is to segment the instrument's region of interest. The U2-Net-DF network receives these RoI images as input and uses its multi-scale feature extraction capability to analyze the image layer by layer to distinguish the instrument needle from the background. This step is key to achieving accurate segmentation because it involves separating the instrument needle from the complex background. S53: The U2-Net-DF network effectively extracts target pixels from the background through its deep structure and feature fusion mechanism. The network's upsampling process and edge detail recovery capabilities ensure accurate segmentation of the pointer and scale. This step is crucial for subsequent reading calculations because it provides accurate position information of the pointer relative to the scale: ； S54: After the segmentation is completed, the output of U2-Net-DF is the pixel-level data of the pointer and scale. This data includes the precise position and shape of the pointer, as well as the distribution of the scale. This information provides a basis for the automatic reading of instrument values, allowing values ​​to be directly extracted from the image without manual intervention, improving the automation and accuracy of the instrument panel reading in the hydropower plant.

7. The method for positioning and reading data of an intelligent inspection robot based on computer vision according to claim 6 is characterized in that: Step S6 includes the following sub-steps: S61: Before polar coordinate transformation, the hydropower plant instrument panel image is preprocessed, including denoising, contrast enhancement and image distortion correction. This step ensures that the pointer and scale lines in the image are clearly visible, laying the foundation for subsequent polar coordinate transformation and data reading. Preprocessing is a key step to improve the accuracy of image analysis and helps reduce the impact of environmental factors on readings. S62: Use the polar coordinate transformation function in the image processing library to transform the preprocessed image from the rectangular coordinate system to the polar coordinate system. This step involves setting the size of the output image, the transformation center point, and the radius of the boundary circle. The polar coordinate transformation can map the circular instrument panel surface to a rectangular area, simplifying the process of extracting the position information of the pointer and scale. S63: In the process of polar coordinate transformation, it is necessary to select appropriate interpolation methods and polar coordinate mapping methods, use linear interpolation and polar coordinate forward transformation, the selection of these parameters will affect the effect and performance of the transformation, the correct parameter setting can ensure that the transformed image maintains high quality and provide an accurate basis for subsequent data extraction: ； S64: After the transformation, the position information of the pointer and scale is converted into radial distance and angle from the center point in the polar coordinate system. This step simplifies the data reading process because the pixel-level data of the pointer and scale are easier to extract, which facilitates the accurate reading of the instrument value. S65: In the polar coordinate system, the value of the instrument can be accurately read by measuring the radial distance and angle from the end point of the pointer to the center point. This image dimensionality reduction method based on polar coordinate transformation can effectively extract target pixels from the background and obtain pixel-level data of the segmented pointer and scale, thereby realizing accurate and effective data reading of pointer instruments.

Images