A method for processing current transformer wiring images
By introducing a variety of attention modules and layered convolutional structures, the problem of noise and light and shadow interference in industrial environments is solved, and the clarity and readability of the image is improved, and it is suitable for real-time industrial scenarios.
Patent Information
- Application Number
- CN202411442896.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-16
- Publication Date
- 2024-12-17
- Estimated Expiration
- 2044-10-16
AI Technical Summary
The prior art is difficult to effectively deal with noise, blur and light and shadow interference in the wiring images of current transformers in industrial environments, resulting in the inability to accurately identify wiring details and key points, affecting the normal operation and maintenance of the equipment.
Using a variety of attention modules, including adaptive channel attention modules and hybrid attention modules, we can extract image features in layers, combining image preprocessing, shallow convolution, feature map sparseness, middle-layer convolution and deep convolution modules. Generate enhanced wiring images.
It improves image clarity and readability, can enhance the image details of the current transformer wiring under different lighting conditions, while retaining global structural information, reducing computing complexity, and is suitable for real-time industrial scenarios.
Smart Images
Figure CN118967457B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of image enhancement, and in particular relates to a current transformer wiring image processing method. Background Art
[0002] Current transformers are widely used electrical equipment in power systems, mainly used to convert large currents for measurement, protection and control. During the installation and maintenance of power equipment, the clarity of the wiring image of the current transformer is particularly important in detection, monitoring and diagnosis. Existing current transformer wiring detection mainly relies on manual observation or simple image processing methods. However, since the installation location of current transformers in industrial environments is usually complex, the wiring environment is limited, and the lighting conditions are poor, it is easy to cause the captured wiring images to have noise, blur, light and shadow interference and other problems. These problems make it impossible to accurately identify the wiring details and key points when analyzing the current transformer wiring image, which in turn affects the normal operation and maintenance of the equipment.
[0003] In recent years, with the development of image processing and deep learning technology, image enhancement methods based on convolutional neural networks (CNNs) have been widely used in various fields, especially in image denoising, detail enhancement and feature extraction. However, most of the existing image enhancement methods are aimed at natural images or medical images. The current transformer wiring images in industrial scenes have complex backgrounds, many wiring details, and are extremely sensitive to light and noise. The existing technology cannot effectively solve the enhancement needs of such images. Therefore, in order to meet the enhancement needs of current transformer wiring images under different lighting and complex environments, it is of great significance to propose an image enhancement method that can improve the clarity and readability of wiring images in industrial scenes. Summary of the invention
[0004] The present invention provides a current transformer wiring image processing method, which aims to enhance the focus on key features in the current transformer wiring image by introducing different attention modules. The current transformer wiring image processing method comprises the following steps:
[0005] S1. Preparation of current transformer wiring image dataset: using an industrial camera to take pictures of current transformers, obtaining current transformer wiring images of current transformers under different lighting conditions, and obtaining a current transformer wiring image dataset;
[0006] S2, constructing an image preprocessing module IPM, including an image conversion module and a foreground extraction module;
[0007] S3, build a shallow convolution module SCM, including 3×3 convolution, adaptive channel attention module ACAM, ReLU activation function, batch normalization and maximum pooling;
[0008] S4, construct a feature map sparsification module FSM, including 1×1 convolution;
[0009] S5, construct the middle layer convolution module MCM, including 3×3 convolution, batch normalization, ReLU activation function, adaptive channel attention module ACAM, spatial attention module SAM and maximum pooling;
[0010] S6. Construct a deep convolution module DCM, including 3×3 convolution, batch normalization, ReLU activation function, adaptive channel attention module ACAM, hybrid attention module CBAM, global average pooling and fully connected layer;
[0011] S7, constructing a current transformer wiring enhanced image generation module EIGM, including three 3×3 deconvolution layers;
[0012] S8. Construct a current transformer wiring image processing model, including input, image preprocessing module, shallow convolution module, feature map sparsification module, middle convolution module, deep convolution module, current transformer wiring enhanced image generation module and output.
[0013] Preferably, in step S2, the image preprocessing module IPM includes an image conversion module and a foreground extraction module. The input of the image preprocessing module IPM is a current transformer wiring image, which is an image in an RGB color space. The image conversion module converts the current transformer wiring image in the RGB color space into a current transformer wiring image in the HSV color space. The specific method is: calculate the brightness V: , calculate the saturation S: , calculate the hue H: if it satisfies ,implement If satisfied ,implement If satisfied ,implement , where R is the value on the R channel of the current transformer wiring image, G is the value on the G channel of the current transformer wiring image, and B is the value on the B channel of the current transformer wiring image. It is the maximum value of the three channels R, G, and B of the current transformer wiring image. It is the minimum value of the three RGB channel values of the current transformer wiring image; the foreground extraction module creates a mask according to the set HSV color threshold, and uses the mask to extract the foreground in the current transformer wiring image in the RGB color space. The specific method is: create a mask according to the set HSV color threshold, the mask is a binary image, and the pixels of the image are divided into foreground and background according to the set threshold condition. Pixels that meet the threshold condition are marked as foreground and represented as white in the mask, and pixels that do not meet the condition are marked as background and represented as black in the mask. The mask is bit-operated with the current transformer wiring image in the RGB color space. For pixels that meet the threshold condition, the mask keeps them unchanged, and for pixels that do not meet the threshold condition, the mask sets them to black, extracts the current transformer wiring area, and obtains the current transformer wiring preprocessing image. , Current transformer wiring preprocessing image The image is in RGB color space.
[0014] Preferably, in step S3, the input of the shallow convolution module SCM is the preprocessed image of the current transformer wiring , where H, W and 3 represent the preprocessing images of current transformer wiring The height, width and number of channels of the current transformer wiring are first preprocessed by a 3×3 standard convolution. Perform preliminary feature extraction to obtain : ,in Represents a 3×3 convolution, and then uses the adaptive channel attention module ACAM to enhance channel attention, uses the ReLU activation function for activation, and then performs batch normalization and maximum pooling operations. The specific process is: ,in Output feature map for shallow convolution module, ,in , and C are the output feature maps of the shallow convolutional module respectively. The height, width and number of channels of represents the maximum pooling operation, BN represents batch normalization, ReLU is the activation function, and ACAM is the adaptive channel attention module.
[0015] Preferably, in steps S3, S5 and S6, the adaptive channel attention module ACAM includes global average pooling, dynamic 1×1 convolution of adaptive convolution kernel, Sigmoid activation function and channel weighting, and the input feature map ,in , and Respectively represent the height, width and number of channels of the input feature map. First, the input feature map is globally averaged pooled in the spatial dimension to obtain the global information of each channel. The specific process is as follows: , is the global average of channel c, and the output size of global average pooling is ; Calculate the dynamic convolution kernel size according to the number of channels of the input feature map. The calculation process of the dynamic convolution kernel k is: , where min() represents the minimum value operation, and max() represents the maximum value operation. is the preset maximum convolution kernel size, is the preset minimum convolution kernel size, is the number of channels of the input feature map; then the dynamic convolution kernel k is applied to perform 1×1 convolution to generate the attention weight of each channel, and the Sigmoid activation function is used to activate it to obtain the channel weight. The specific process is: ,in is the channel weight, and its size is , Sigmoid represents the Sigmoid activation function, represents 1×1 convolution, is the global average value of channel c, k is the dynamic convolution kernel; finally, the channel weight is multiplied element by element with each channel of the input feature map. The specific process is: , and get the output feature map ,in , and Respectively represent the height, width and number of channels of the output feature map.
[0016] Preferably, in step S4, the input of the feature map sparsification module FSM is the output feature map of the shallow convolution module , a 1×1 convolution is used to perform feature map sparse operation. The specific process is as follows: ,in is the output feature map of the feature map sparsification module FSM, ,in , and They are the output feature maps of the feature map sparsification module FSM. The height, width and number of channels of Represents a 1×1 convolution.
[0017] Preferably, in step S5, the input of the middle convolution module MCM is the output feature map of the feature map sparsification module FSM , ,in , and They are The height, width and number of channels of the middle convolution module MCM output for ,in represents the maximum pooling operation, SAM is the spatial attention module, ACAM is the adaptive channel attention module, ReLU is the activation function, BN represents batch normalization, Represents 3×3 convolution, the output of the middle convolution module MCM ,in , and C are The height, width, and number of channels of the
[0018] Preferably, in step S6, the input of the deep convolution module DCM is the output of the middle convolution module MCM ,in , and C are The height, width and number of channels of the deep convolution module DCM for , where FC is the fully connected layer, GAP is the global average pooling, CBAM is the hybrid attention module, ACAM is the adaptive channel attention module, ReLU is the activation function, and BN represents batch normalization. Represents 3×3 convolution, the output of the deep convolution module DCM ,in , and 2C are The height, width, and number of channels of the
[0019] Preferably, in step S7, the input of the current transformer wiring enhanced image generation module EIGM is the output of the deep convolution module DCM ,in , and 2C are The height, width and number of channels of the network are upsampled through three 3×3 deconvolution layers. The specific process is as follows: ,in The output of the 3×3 deconvolution and the current transformer wiring enhanced image generation module EIGM , where H, W and 3 represent the height, width and number of channels of the enhanced image of the current transformer wiring, respectively.
[0020] Compared with the prior art, the present invention has the following technical effects: introducing multiple attention modules, performing layer-by-layer feature extraction on images, from low-level edges and textures to high-level global information, effectively enhancing the ability to capture key features, and being able to retain global structural information while enhancing image details. It also has higher versatility and can be applied to other industrial image processing tasks, while reducing computational complexity, making the entire solution suitable for real-time industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is a flow chart of a current transformer wiring image processing method provided by the present invention.
[0022] Figure 2 This is a diagram of the current transformer wiring image processing process provided by the present invention.
[0023] Figure 3 It is a structural diagram of the adaptive channel attention module ACAM provided by the present invention.
[0024] Figure 4 This is a diagram showing the enhanced effect of the current transformer wiring image in an embodiment provided by the present invention. DETAILED DESCRIPTION
[0025] The following will be combined with the accompanying drawings of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0026] See also Figures 1 to 4 ,The present invention provides a current transformer wiring image processing method, which aims to enhance the attention to key features in the current transformer wiring image by introducing different attention modules.,A current transformer wiring image processing method includes the following steps.
[0027] S1. The current transformer wiring image dataset is prepared. An industrial camera is used to take pictures of the current transformer, and the current transformer wiring images of the current transformer under different lighting conditions are obtained. A current transformer wiring image dataset is obtained, with a total of 3500 images, which are divided into training set and validation set according to 7:3.
[0028] S2. Construct an image preprocessing module IPM, including an image conversion module and a foreground extraction module.
[0029] Further, in step S2, the image preprocessing module IPM includes an image conversion module and a foreground extraction module. The input of the image preprocessing module IPM is the current transformer wiring image, which is an image in the RGB color space. The image conversion module converts the current transformer wiring image in the RGB color space into the current transformer wiring image in the HSV color space. The specific method is: calculate the brightness V: , calculate the saturation S: , calculate the hue H: if it satisfies ,implement If satisfied ,implement If satisfied ,implement , where R is the value on the R channel of the current transformer wiring image, G is the value on the G channel of the current transformer wiring image, and B is the value on the B channel of the current transformer wiring image. It is the maximum value of the three channels R, G, and B of the current transformer wiring image. It is the minimum value of the three RGB channel values of the current transformer wiring image; the foreground extraction module creates a mask according to the set HSV color threshold, and uses the mask to extract the foreground in the current transformer wiring image in the RGB color space. The specific method is: create a mask according to the set HSV color threshold, the mask is a binary image, and the pixels of the image are divided into foreground and background according to the set threshold condition. Pixels that meet the threshold condition are marked as foreground and represented as white in the mask, and pixels that do not meet the condition are marked as background and represented as black in the mask. The mask is bit-operated with the current transformer wiring image in the RGB color space. For pixels that meet the threshold condition, the mask keeps them unchanged, and for pixels that do not meet the threshold condition, the mask sets them to black, extracts the current transformer wiring area, and obtains the current transformer wiring preprocessing image. , Current transformer wiring preprocessing image For an image in RGB color space, in this embodiment, the HSV color thresholds are set as follows: the threshold of hue H is 0 to 180, the threshold of saturation S is 0 to 50, and the threshold of brightness V is 60 to 255.
[0030] S3. Build a shallow convolution module SCM, including 3×3 convolution, adaptive channel attention module ACAM, ReLU activation function, batch normalization and maximum pooling.
[0031] Furthermore, in step S3, the input of the shallow convolution module SCM is the preprocessed image of the current transformer wiring. , where 480, 640 and 3 represent the preprocessing images of current transformer wiring respectively The height, width and number of channels of the current transformer wiring are first preprocessed by a 3×3 standard convolution. Perform preliminary feature extraction to obtain : ,in Represents a 3×3 convolution, and then uses the adaptive channel attention module ACAM to enhance channel attention, uses the ReLU activation function for activation, and then performs batch normalization and maximum pooling operations. The specific process is: ,in Output feature map for shallow convolution module, , where 240, 320 and 64 are the output feature maps of the shallow convolution module The height, width and number of channels of represents the maximum pooling operation, BN represents batch normalization, ReLU is the activation function, and ACAM is the adaptive channel attention module.
[0032] S4. Construct a feature map sparsification module FSM, including 1×1 convolution.
[0033] Furthermore, in step S4, the input of the feature map sparsification module FSM is the output feature map of the shallow convolution module , a 1×1 convolution is used to perform feature map sparse operation. The specific process is as follows: ,in is the output feature map of the feature map sparsification module FSM, , where 240, 320 and 32 are the output feature maps of the feature map sparsification module FSM The height, width and number of channels of Represents a 1×1 convolution.
[0034] S5. Construct the middle-layer convolution module MCM, including 3×3 convolution, batch normalization, ReLU activation function, adaptive channel attention module ACAM, spatial attention module SAM and maximum pooling.
[0035] Furthermore, in step S5, the input of the middle convolution module MCM is the output feature map of the feature map sparsification module FSM , , where 240, 320 and 32 are The height, width and number of channels of the middle convolution module MCM output for ,in represents the maximum pooling operation, SAM is the spatial attention module, ACAM is the adaptive channel attention module, ReLU is the activation function, BN represents batch normalization, Represents 3×3 convolution, the output of the middle convolution module MCM , where 120, 160 and 64 are The height, width, and number of channels of the
[0036] S6. Construct a deep convolution module DCM, including 3×3 convolution, batch normalization, ReLU activation function, adaptive channel attention module ACAM, hybrid attention module CBAM, global average pooling and fully connected layer.
[0037] Furthermore, in step S6, the input of the deep convolution module DCM is the output of the middle convolution module MCM , where 120, 160 and 64 are The height, width and number of channels of the deep convolution module DCM for , where FC is the fully connected layer, GAP is the global average pooling, CBAM is the hybrid attention module, ACAM is the adaptive channel attention module, ReLU is the activation function, and BN represents batch normalization. Represents 3×3 convolution, the output of the deep convolution module DCM , where 60, 80 and 128 are The height, width, and number of channels of the
[0038] Furthermore, in steps S3, S5 and S6, the adaptive channel attention module ACAM includes global average pooling, dynamic 1×1 convolution of adaptive convolution kernel, Sigmoid activation function and channel weighting, and the input feature map ,in , and Respectively represent the height, width and number of channels of the input feature map. First, the input feature map is globally averaged pooled in the spatial dimension to obtain the global information of each channel. The specific process is as follows: , is the global average of channel c, and the output size of global average pooling is ; Calculate the dynamic convolution kernel size according to the number of channels of the input feature map. The calculation process of the dynamic convolution kernel k is: , where min() represents the minimum value operation, and max() represents the maximum value operation. is the preset maximum convolution kernel size, is the preset minimum convolution kernel size, is the number of channels of the input feature map; then the dynamic convolution kernel k is applied to perform 1×1 convolution to generate the attention weight of each channel, and the Sigmoid activation function is used to activate it to obtain the channel weight. The specific process is: ,in is the channel weight, and its size is , Sigmoid represents the Sigmoid activation function, represents 1×1 convolution, is the global average value of channel c, k is the dynamic convolution kernel; finally, the channel weight is multiplied element by element with each channel of the input feature map. The specific process is: , and get the output feature map ,in , and Respectively represent the height, width and number of channels of the output feature map.
[0039] S7. Construct a current transformer wiring enhanced image generation module EIGM, including three 3×3 deconvolution layers.
[0040] Further, in step S7, the input of the current transformer wiring enhanced image generation module EIGM is the output of the deep convolution module DCM , where 60, 80 and 128 are The height, width and number of channels of the network are upsampled through three 3×3 deconvolution layers. The specific process is as follows: ,in The output of the 3×3 deconvolution and the current transformer wiring enhanced image generation module EIGM , where 480, 640 and 3 represent the height, width and number of channels of the enhanced image of the current transformer wiring, respectively.
[0041] S8. Construct a current transformer wiring image processing model, including input, image preprocessing module, shallow convolution module, feature map sparsification module, middle convolution module, deep convolution module, current transformer wiring enhanced image generation module and output.
[0042] Furthermore, the current transformer wiring image processing model was coded using the Pycharm application and Python language, using the Pytorch framework, and the model input was trained with internal defect images of the high-voltage distribution cabinet with a resolution of 480×640×3. The model was trained for 300 epochs, of which 50 epochs were frozen.
[0043] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.
Claims
1. A method for processing a current transformer wiring image, characterized in that: The following steps are involved: S1. Preparation of current transformer wiring image dataset: using an industrial camera to take pictures of current transformers, obtaining current transformer wiring images of current transformers under different lighting conditions, and obtaining a current transformer wiring image dataset; S2, constructing an image preprocessing module IPM, including an image conversion module and a foreground extraction module; S3, build a shallow convolution module SCM, including 3×3 convolution, adaptive channel attention module ACAM, ReLU activation function, batch normalization and maximum pooling; S4, construct a feature map sparsification module FSM, including 1×1 convolution; S5, construct the middle layer convolution module MCM, including 3×3 convolution, batch normalization, ReLU activation function, adaptive channel attention module ACAM, spatial attention module SAM and maximum pooling; S6. Construct a deep convolution module DCM, including 3×3 convolution, batch normalization, ReLU activation function, adaptive channel attention module ACAM, hybrid attention module CBAM, global average pooling and fully connected layer; S7, constructing a current transformer wiring enhanced image generation module EIGM, including three 3×3 deconvolution layers; S8. Construct a current transformer wiring image processing model, including input, image preprocessing module, shallow convolution module, feature map sparsification module, middle convolution module, deep convolution module, current transformer wiring enhanced image generation module and output.
2. A current transformer wiring image processing method according to claim 1, characterized in that: In step S2, the image preprocessing module IPM includes an image conversion module and a foreground extraction module. The input of the image preprocessing module IPM is a current transformer wiring image, which is an image in an RGB color space. The image conversion module converts the current transformer wiring image in the RGB color space into a current transformer wiring image in the HSV color space. The specific method is: calculate the brightness V: , calculate the saturation S: , calculate the hue H: if it satisfies ,implement If satisfied ,implement If satisfied ,implement , where R is the value on the R channel of the current transformer wiring image, G is the value on the G channel of the current transformer wiring image, and B is the value on the B channel of the current transformer wiring image. It is the maximum value of the three channels R, G, and B of the current transformer wiring image. It is the minimum value of the three RGB channels of the current transformer wiring image; The foreground extraction module creates a mask according to the set HSV color threshold, and uses the mask to extract the foreground in the current transformer wiring image in the RGB color space. The specific method is as follows: a mask is created according to the set HSV color threshold. The mask is a binary image. The pixels of the image are divided into foreground and background according to the set threshold condition. Pixels that meet the threshold condition are marked as foreground and represented as white in the mask. Pixels that do not meet the condition are marked as background and represented as black in the mask. The mask is bit-operated with the current transformer wiring image in the RGB color space. For pixels that meet the threshold condition, the mask keeps them unchanged. For pixels that do not meet the threshold condition, the mask sets them to black. The current transformer wiring area is extracted to obtain the current transformer wiring preprocessing image. , Current transformer wiring preprocessing image The image is in RGB color space.
3. The current transformer wiring image processing method according to claim 1 is further characterized in that: In step S3, the input of the shallow convolution module SCM is the preprocessed image of the current transformer wiring , where H, W and 3 represent the preprocessing images of current transformer wiring The height, width and number of channels of the current transformer wiring are first preprocessed by a 3×3 standard convolution. Perform preliminary feature extraction to obtain : ,in Represents a 3×3 convolution, and then uses the adaptive channel attention module ACAM to enhance channel attention, uses the ReLU activation function for activation, and then performs batch normalization and maximum pooling operations. The specific process is: ,in Output feature map for shallow convolution module, ,in , and C are the output feature maps of the shallow convolutional module respectively. The height, width and number of channels of represents the maximum pooling operation, BN represents batch normalization, ReLU is the activation function, and ACAM is the adaptive channel attention module.
4. The current transformer wiring image processing method according to claim 1 is further characterized in that: In steps S3, S5 and S6, the adaptive channel attention module ACAM includes global average pooling, dynamic 1×1 convolution of adaptive convolution kernel, sigmoid activation function and channel weighting, and the input feature map ,in , and Respectively represent the height, width and number of channels of the input feature map. First, the input feature map is globally averaged pooled in the spatial dimension to obtain the global information of each channel. The specific process is as follows: , is the global average of channel c, and the output size of global average pooling is ; Calculate the dynamic convolution kernel size according to the number of channels of the input feature map. The calculation process of the dynamic convolution kernel k is: , where min() represents the minimum value operation, and max() represents the maximum value operation. is the preset maximum convolution kernel size, is the preset minimum convolution kernel size, is the number of channels of the input feature map; then the dynamic convolution kernel k is applied to perform 1×1 convolution to generate the attention weight of each channel, and the Sigmoid activation function is used to activate it to obtain the channel weight. The specific process is: ,in is the channel weight, and its size is , Sigmoid represents the Sigmoid activation function, represents 1×1 convolution, is the global average value of channel c, k is the dynamic convolution kernel; finally, the channel weight is multiplied element by element with each channel of the input feature map. The specific process is: , and get the output feature map ,in , and Respectively represent the height, width and number of channels of the output feature map.
5. The current transformer wiring image processing method according to claim 1 is further characterized in that: In step S4, the input of the feature map sparsification module FSM is the output feature map of the shallow convolution module , a 1×1 convolution is used to perform feature map sparse operation. The specific process is as follows: ,in is the output feature map of the feature map sparsification module FSM, ,in , and They are the output feature maps of the feature map sparsification module FSM. The height, width and number of channels of Represents a 1×1 convolution.
6. The current transformer wiring image processing method according to claim 1, further characterized in that: In step S5, the input of the middle convolution module MCM is the output feature map of the feature map sparsification module FSM , ,in , and They are The height, width and number of channels of the middle convolution module MCM output for ,in represents the maximum pooling operation, SAM is the spatial attention module, ACAM is the adaptive channel attention module, ReLU is the activation function, BN represents batch normalization, Represents 3×3 convolution, the output of the middle convolution module MCM ,in , and C are The height, width, and number of channels of the 7. The current transformer wiring image processing method according to claim 1, further characterized in that: In step S6, the input of the deep convolution module DCM is the output of the middle convolution module MCM ,in , and C are The height, width and number of channels of the deep convolution module DCM for , where FC is the fully connected layer, GAP is the global average pooling, CBAM is the hybrid attention module, ACAM is the adaptive channel attention module, ReLU is the activation function, and BN represents batch normalization. Represents 3×3 convolution, the output of the deep convolution module DCM ,in , and 2C are The height, width, and number of channels of the 8. The current transformer wiring image processing method according to claim 1, further characterized in that: In step S7, the input of the current transformer wiring enhanced image generation module EIGM is the output of the deep convolution module DCM ,in , and 2C are The height, width and number of channels of the network are upsampled through three 3×3 deconvolution layers. The specific process is as follows: ,in The output of the 3×3 deconvolution and the current transformer wiring enhanced image generation module EIGM , where H, W and 3 represent the height, width and number of channels of the enhanced image of the current transformer wiring, respectively.