A method for segmenting key objects in standard distribution station areas based on regional priors

Abstract

The present invention discloses a method for segmenting key objects in a standard distribution station area based on regional priors. In order to overcome the problem of poor recognition effect of the existing visual recognition algorithm in complex environments, the present invention includes the following steps: making a shooting guide area according to the contour shape of the object to be identified in the standard distribution station area; collecting standard station area data; obtaining a training data set, a verification data set, a test data set and the corresponding ideal standard segmentation results; constructing an attention feature area, using the attention feature area image as input and the corresponding ideal standard segmentation result image as output to train the image segmentation network model; adjusting the training parameters, repeating the learning process, and completing the training of the image segmentation network model; evaluating the trained image segmentation network model. An auxiliary shooting guide area is added, and the mask of the area image is used to construct the encoded attention area image, which has higher credibility and better image segmentation accuracy.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular to a method for segmenting key objects in a standard power distribution station area based on regional priors. Background Art

[0002] Distribution substations are among the most widespread and numerous power equipment in the transmission grid. Currently, inspection and maintenance of distribution substations are primarily performed manually. With the advancement of computer vision and deep learning technologies, vision-based object detection and image segmentation methods have matured. Vision-based identification and quantification of key objects in distribution substations can reduce the workload of manual inspections and improve efficiency. Therefore, research on object detection and quantification in distribution substations is of great significance.

[0003] Currently, research on object segmentation in distribution substations is relatively limited. Computer vision researchers typically address this problem by training end-to-end recognition and segmentation networks. However, due to the diverse environments and complex backgrounds found in distribution substations, relevant deep learning models perform poorly in complex and inclement weather, failing to meet real-world requirements. Image segmentation methods based on attention mechanisms or salient region priors often fail to improve object segmentation in complex environments due to low confidence in the prior information.

[0004] For example, a "substation fault detection method and device based on visual recognition technology" disclosed in Chinese patent literature, with publication number CN112634590A, includes: acquiring image information of a target area; detecting the image information according to a preset learning model to obtain the location and status information of the target device; determining whether the location and status information meet preset warning conditions to obtain an alarm; and, in response to the obtained alarm information, transmitting the alarm information to a terminal. This solution's visual recognition algorithm performs poorly in complex environments such as inclement weather and cannot meet actual on-site requirements. Summary of the Invention

[0005] The present invention mainly solves the problem that the existing visual recognition algorithm has poor recognition effect in complex environments; it provides a standard distribution substation key object segmentation method based on regional prior, and uses regional prior to guide the object segmentation network training of the distribution substation area, thereby improving the object detection and segmentation accuracy of the distribution substation area.

[0006] The above technical problems of the present invention are mainly solved by the following technical solutions:

[0007] A method for segmenting key objects in a standard power distribution area based on regional priors includes the following steps:

[0008] S1: Create a shooting guide area based on the outline shape of the object to be identified in the standard distribution station area;

[0009] S2: Adjust the camera posture so that the object to be identified is located in the shooting guide area and collect standard area data;

[0010] S3: Divide the collected standard area data into a training dataset S, a validation dataset Y, a test dataset C, and the corresponding ideal standard segmentation results;

[0011] S4: Construct the attention feature region, take the attention feature region image as input, and the corresponding ideal standard segmentation result image as output to train the image segmentation network model;

[0012] S5: Adjust the training parameters and repeat the learning process to complete the training of the image segmentation network model;

[0013] S6: Use the validation dataset Y and the test dataset C to evaluate the trained image segmentation network model and complete the training of the key object segmentation network in the standard distribution station area.

[0014] This solution adds a photo-assisted guidance area and uses the mask of the image in this area to construct an image encoding the attention region. Compared with other attention mechanisms, this method has higher credibility, making the network trained by this solution have better image segmentation accuracy. It can achieve good recognition results in complex environments such as inclement weather, meeting actual field requirements.

[0015] Preferably, the construction of the attention feature area comprises the following steps:

[0016] The input image captured by the camera and the auxiliary guidance mask image of the shooting guidance area are convolved and normalized respectively to extract features, and the feature extraction results of the two images are added pixel by pixel;

[0017] After processing with the ReLU activation function, convolution and normalization are performed again, and activation processing is performed using the sigmoid activation function. The processed result is element-wise multiplied with the input image to obtain the encoded attention image. Based on the shape of the object to be recognized, the encoded attention region is constructed and a regional prior is generated to improve the accuracy of object detection and segmentation.

[0018] Preferably, the atlas F obtained by adding the feature extraction results of the two images pixel by pixel is:

[0019]

[0020] I motify =Sigmoid(BatchNorm(Conv(ReLU(F))))⊙Iinput

[0021] Among them, I input is the input image;

[0022] I mask A photo-assisted guidance mask image for capturing the guidance area;

[0023] Conv(·) is the convolution function;

[0024] BatchNorm(·) is the normalization function;

[0025] I motify Encode the image for the final attention;

[0026] Sigmoid is the sigmoid activation function;

[0027] It means element-by-element addition;

[0028] ⊙ represents tuple-wise multiplication.

[0029] Construct encoded attention regions and perform regional priors to improve the accuracy of object detection and segmentation.

[0030] Preferably, the image segmentation network model includes an encoder and a decoder in sequence; the encoder is used for image feature extraction; and the decoder is used for image segmentation. A network training for segmenting key objects in a standard power distribution area is performed.

[0031] Preferably, the encoding end includes, in sequence, convolution layer 1_1, convolution layer 1_2, pooling layer 1, convolution layer 2_1, convolution layer 2_2, pooling layer 2, convolution layer 3_1, convolution layer 3_2, pooling layer 3, convolution layer 4_1, convolution layer 4_2, pooling layer 4, convolution layer 5_1, convolution layer 5_2, and pooling layer 5. The convolution layer and pooling layer implement dimensionality reduction and feature extraction.

[0032] Preferably, the decoder includes, in order, convolutional layer 6_1, convolutional layer 6_2, convolutional layer 6_3, upsampling layer 1, convolutional layer 7_1, convolutional layer 7_2, upsampling layer 2, convolutional layer 8_1, convolutional layer 8_2, upsampling layer 3, convolutional layer 9_1, convolutional layer 9_2, upsampling layer 4, convolutional layer 10_1, convolutional layer 10_2, and a softmax layer. The convolutional and pooling layers implement dimensionality reduction and feature extraction, the upsampling layer implements deconvolution and information decoding, and the softmax layer implements pixel classification to achieve the purpose of image segmentation.

[0033] The beneficial effects of the present invention are:

[0034] This solution adds a photo-assisted guidance area and uses the mask of the image in this area to construct an image encoding the attention region. Compared with other attention mechanisms, this method has higher credibility, making the network trained by this solution have better image segmentation accuracy. It can achieve good recognition results in complex environments such as inclement weather, meeting actual field requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of the method for segmenting key objects in a standard distribution station area based on regional priors of the present invention. DETAILED DESCRIPTION

[0036] The technical solution of the present invention will be further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0037] Example:

[0038] This embodiment provides a method for segmenting key objects in a standard distribution station area based on regional priors, such as Figure 1 As shown, the following steps are included:

[0039] S1: Create a shooting guide area based on the outline shape of the object to be identified in the standard distribution station area.

[0040] For example, consider power transmission equipment near a field or two utility poles with a distribution box between them. Based on the outline of the object to be identified, a shooting guide area is created. This H-shaped area completely encompasses the object to be identified.

[0041] S2: Adjust the camera posture so that the object to be identified is located within the shooting guide area and collect standard area data. Collect a large amount of standard area data for training the image segmentation network model.

[0042] S3: Divide the collected standard area data into a training dataset S, a validation dataset Y, a test dataset C, and the corresponding ideal standard segmentation results. The ground truth segmentation (ideal standard segmentation result) includes the ideal standard training data segmentation result S_g corresponding to the training dataset S, the ideal standard validation data segmentation result Y_g corresponding to the validation dataset Y, and the ideal standard test data segmentation result C_g corresponding to the test dataset C.

[0043] S4: Construct the attention feature region, take the attention feature region image as input, and the corresponding ideal standard segmentation result image as output to train the image segmentation network model.

[0044] Constructing the attention feature region includes the following steps:

[0045] The input image taken by the camera and the auxiliary guidance mask image of the shooting guidance area are convolved and normalized respectively to extract features, and the feature extraction results of the two images are added pixel by pixel.

[0046] The atlas F after adding the feature extraction results of the two images pixel by pixel is:

[0047]

[0048] After processing with the ReLU activation function, convolution and normalization are performed again, and activation processing is performed using the sigmoid activation function. The processed result is multiplied element-by-element with the input image to obtain the encoded attention image.

[0049] I motify =Sigmoid(BatchNorm(Conv(ReLU(F))))⊙I input

[0050] Among them, I input is the input image;

[0051] I mask A photo-assisted guidance mask image for capturing the guidance area;

[0052] Conv(·) is the convolution function;

[0053] BatchNorm(·) is the normalization function;

[0054] I motify Encode the image for the final attention;

[0055] Sigmoid is the sigmoid activation function;

[0056] It means element-by-element addition;

[0057] ⊙ represents tuple-wise multiplication.

[0058] S5: Adjust the training parameters and repeat the learning process to complete the training of the image segmentation network model.

[0059] The image segmentation network model consists of an encoder and a decoder in sequence; the encoder is used for image feature extraction; the decoder is used for image segmentation.

[0060] The encoding end includes convolution layer 1_1, convolution layer 1_2, pooling layer 1, convolution layer 2_1, convolution layer 2_2, pooling layer 2, convolution layer 3_1, convolution layer 3_2, pooling layer 3, convolution layer 4_1, convolution layer 4_2, pooling layer 4, convolution layer 5_1, convolution layer 5_2, and pooling layer 5 in sequence.

[0061] The decoder side includes convolution layer 6_1, convolution layer 6_2, convolution layer 6_3, upsampling layer 1, convolution layer 7_1, convolution layer 7_2, upsampling layer 2, convolution layer 8_1, convolution layer 8_2, upsampling layer 3, convolution layer 9_1, convolution layer 9_2, upsampling layer 4, convolution layer 10_1, convolution layer 10_2 and a softmax layer.

[0062] The convolution layer and pooling layer realize dimensionality reduction and feature extraction, the upsampling layer realizes deconvolution and information decoding, and the softmax layer realizes pixel classification to achieve the purpose of image segmentation.

[0063] S6: Use the validation dataset Y and the test dataset C to evaluate the trained image segmentation network model and complete the training of the key object segmentation network in the standard distribution station area.

[0064] This solution adds a photo-assisted guidance area and uses the mask of the image in this area to construct an image encoding the attention region. Compared with other attention mechanisms, this method has higher credibility, making the network trained by this solution have better image segmentation accuracy. It can achieve good recognition results in complex environments such as inclement weather, meeting actual field requirements.

[0065] It should be understood that the embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope limited by the appended claims of the application.

Claims

1. A method for segmenting key objects in a standard distribution station area based on regional priors, characterized in that: The following steps are involved: S1: Create a shooting guide area based on the outline shape of the object to be identified in the standard distribution station area; S2: Adjust the camera posture so that the object to be identified is located in the shooting guide area and collect standard area data; S3: Divide the collected standard area data into a training dataset S, a validation dataset Y, a test dataset C, and the corresponding ideal standard segmentation results; S4: Construct an attention feature region, perform convolution and normalization operations on the input image captured by the camera and the auxiliary guidance mask image of the shooting guidance region, perform feature extraction, and add the feature extraction results of the two images pixel by pixel to obtain a map F. After using the ReLU activation function on the map F, it is convolved and normalized again, and activated using the sigmoid activation function. The processed result is element-wise multiplied with the input image to obtain an encoded attention region image; use the encoded attention region image as input and the corresponding ideal standard segmentation result image as output to train the image segmentation network model; S5: Adjust the training parameters and repeat the learning process to complete the training of the image segmentation network model; S6: Use the validation dataset Y and the test dataset C to evaluate the trained image segmentation network model and complete the training of the key object segmentation network in the standard distribution station area.

2. The method for segmenting key objects in a standard distribution station area based on regional priors according to claim 1, characterized in that: The atlas F after adding the feature extraction results of the two images pixel by pixel is: I motify =Sigmoid(BatchNorm(Conv(ReLU(F))))⊙I input Among them, I input is the input image; I mask A photo-assisted guidance mask image for capturing the guidance area; Conv(·) is the convolution function; BatchNorm(·) is the normalization function; Sigmoid is the sigmoid activation function; I motify Encode the image for the final attention; It means element-by-element addition; ⊙ represents tuple-wise multiplication.

3. The method for segmenting key objects in a standard distribution station area based on regional priors according to claim 1 or 2, characterized in that: The image segmentation network model includes an encoder and a decoder in sequence; the encoder is used for image feature extraction; and the decoder is used for image segmentation.

4. The method for segmenting key objects in a standard distribution station area based on regional priors according to claim 3, characterized in that: The encoding end includes convolution layer 1_1, convolution layer 1_2, pooling layer 1, convolution layer 2_1, convolution layer 2_2, pooling layer 2, convolution layer 3_1, convolution layer 3_2, pooling layer 3, convolution layer 4_1, convolution layer 4_2, pooling layer 4, convolution layer 5_1, convolution layer 5_2 and pooling layer 5 in sequence.

5. The method for segmenting key objects in a standard distribution station area based on regional priors according to claim 3, characterized in that: The decoder end includes convolution layer 6_1, convolution layer 6_2, convolution layer 6_3, upsampling layer 1, convolution layer 7_1, convolution layer 7_2, upsampling layer 2, convolution layer 8_1, convolution layer 8_2, upsampling layer 3, convolution layer 9_1, convolution layer 9_2, upsampling layer 4, convolution layer 10_1, convolution layer 10_2 and a softmax layer in sequence.

6. The method for segmenting key objects in a standard distribution station area based on regional priors according to claim 5, characterized in that: The convolution layer and pooling layer realize dimensionality reduction and feature extraction, the upsampling layer realizes deconvolution and information decoding, and the softmax layer realizes pixel classification to achieve the purpose of image segmentation.

Images