Electrical Equipment Classification Method and System Based on Improved Fast Region Convolutional Neural Network

Abstract

The present invention discloses a classification method and system for electrical equipment of an improved fast regional convolutional neural network, which collects samples of electrical equipment images or related data; preprocesses the electrical equipment images; inputs the preprocessed pictures into an improved Residual Network 50 network for feature extraction to obtain feature maps; selects candidate boxes in the feature maps and the original images through an RPN network to preliminarily determine the target objects in the images; uses the RoI pooling layer to localize the target objects in the images preliminarily determined, and completes the construction of the improved fast regional convolutional neural network; the present invention improves the Faster-RCNN network, optimizes the network depth, enhances the ability to retain deep semantic information, and thus improves the recognition accuracy and speed of electrical equipment under complex backgrounds.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning target detection, and in particular to a method and system for identifying and classifying electrical equipment based on an improved Faster R-CNN. Background Art

[0002] Object detection technology in the field of computer vision is a vital task, which plays a key role in many applications such as autonomous driving, video surveillance, medical image analysis, etc. Although traditional object detection methods, such as those based on manual features and sliding window methods, have achieved certain results, their performance is still limited when faced with complex scenes such as scale changes, complex backgrounds, and occlusions.

[0003] In recent years, with the rise of deep learning technology, especially the successful application of convolutional neural networks (CNN), the field of object detection has made significant breakthroughs. Deep learning methods not only improve detection performance, but also can learn more discriminative features from large-scale data and have stronger generalization capabilities. However, although deep learning has achieved remarkable results in object detection, there are still some problems that need to be solved.

[0004] One of them is that in the face of a large amount of image detection data, it is mainly dependent on professionals to identify electrical equipment images, which is prone to misjudgment and inefficient. Rapidly extracting and accurately classifying a large number of electrical equipment images to determine whether they are operating normally is the key to ensuring the normal operation of the power system.

[0005] In recent years, with the development of artificial intelligence graphics technology, deep learning has become increasingly mature in image processing. Compared with traditional machine learning, it has a faster detection speed, higher accuracy, and is not easily affected by the environment. Current deep learning object detection algorithms can be divided into anchor algorithms and non-anchor algorithms according to whether an anchor framework is set. Among them, the anchor box algorithm can be further divided into two-stage methods and single-stage methods. The two-stage algorithm first generates candidate regions and then classifies and locates the targets through a convolutional neural network. These methods often have high accuracy but low speed. Commonly used two-stage models include Region-based Convolutional Neural Network (RCNN), Fast Region-based Convolutional Neural Network (Fast-RCNN), and Faster Region-based Convolutional Neural Network (Faster-RCNN). A refinement stage is introduced in the output part of the Faster-RCNN network model to increase the classification and regression refinement of target features, achieving accurate classification and coordinate positioning. The single-stage algorithm directly predicts the position and category of the target from the image without generating candidate regions. Such algorithms often have high speed but low accuracy. Common single-stage models include Single Shot MultiBox Detector (SSD) and You Only Look Once (YOLO) object detection system models. Accurately extracting image features is an important step in improving classification accuracy. The above methods all have certain effects on feature extraction to varying degrees, but in the case of small target objects and complex backgrounds, the features of the targets are less, it is difficult to extract sufficient feature information, and the target features will be affected by irrelevant information such as the background, resulting in unclear target features and a decrease in detection accuracy. Summary of the Invention

[0006] Object of the Invention: To solve the problem that in the case of small target objects and complex backgrounds, the features of the targets are less, it is difficult to extract sufficient feature information, and the target features will be affected by irrelevant information such as the background, resulting in unclear target features and a decrease in detection accuracy, the present invention provides an improved method for classifying electrical equipment based on the Fast Region-based Convolutional Neural Network.

[0007] Technical Solution: To achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] An improved method for classifying electrical equipment based on the Fast Region-based Convolutional Neural Network, comprising the following steps

[0009] Step 1, collect samples of electrical equipment images or related data to obtain an electrical equipment data set. Classify the collected electrical equipment data set according to the training set.

[0010] Step 2, preprocess the electrical equipment image to obtain a preprocessed picture.

[0011] Step 3: Input the preprocessed image into the improved Residual Network 50 network for feature extraction to obtain a feature map.

[0012] Step 4: The RPN network is used to select candidate boxes from the feature map and the original image to preliminarily determine the target object in the image.

[0013] Step 5: Use the RoIpooing layer to localize the target object in the preliminarily determined image and complete the construction of the improved fast regional convolutional neural network.

[0014] Step 6: Train the improved fast regional convolutional neural network using the training set, and perform target detection and classification using the trained improved fast regional convolutional neural network.

[0015] Preferably: the Residual Network 50 network includes a first convolutional layer, a first pooling layer, a residual network, a first average pooling layer, and a first fully connected layer connected in sequence, and the residual network includes four residual blocks connected in sequence, a BN layer is added to the residual block, and jump links are performed between the four residual blocks. Four residual stages are designed in the Residual Network 50 network for feature extraction, the first residual stage consists of 3 residual blocks, the second residual stage consists of 4 residual blocks, the third residual stage consists of 6 residual blocks, and the fourth residual stage consists of 3 residual blocks. Adding a BN layer during the extraction process will directly add part of the original image to the next extraction process.

[0016] Preferably: the residual block includes a first convolution layer, a second convolution layer, and a third convolution layer connected in sequence, the first convolution layer increases the dimension of the input electrical equipment diagram, the second convolution layer extracts features of the input electrical equipment diagram, and the third convolution layer is used to further increase the dimension of the extracted features.

[0017] Preferred: The convolution operation after the residual network is:

[0018]

[0019] in, Indicates Layer feature map output, is the activation function, is the receptive field of the input layer, for Layer feature map output, represents the convolution operation, is the convolution kernel, For the The bias parameter of the th neuron in the layer feature map.

[0020] Preferably, the calculation formula of the BN layer is as follows:

[0021]

[0022]

[0023]

[0024]

[0025] Among them, is the normalized value, is the length of the batch, is the th sample, is the variance, is to normalize to zero mean and unit variance, is to prevent the denominator from being zero, is the linearly transformed value after normalization, is a scale-learnable parameter, is an offset-operation learnable parameter.

[0026] Preferably, the linear representation relationship in the convolutional network can be expressed as

[0027]

[0028] Among them, represents the layer's linear relationship, represents layer's weight coefficient, represents layer's input, represents the offset in the convolutional process.

[0029] The activation function is:

[0030]

[0031] Among them, represents the activation function, represents layer's output.

[0032] Introduce matrix to transform the dimension, and include in the BN layer, so the output is expressed as:

[0033]

[0034] Among them, represents the input of the layer, represents the activation function, represents the linear relationship of the layer, represents the introduced matrix.

[0035] Preferably: The method for localizing the target object in the preliminarily determined image by using the RoI pooling layer in step 5 is as follows: After generating the candidate boxes using the RPN, the candidate boxes are localized through the RoI Pooling Layer to complete object detection and classification. For each candidate box generated by the RPN, the features of the corresponding region are extracted from the feature map. Each candidate box is evenly divided into sub-regions of a fixed size. The feature values within each sub-region are pooled to obtain an output of a fixed size. These outputs are concatenated into a fixed-length vector as the localization feature of the candidate box. Each candidate box is mapped to a feature vector of a fixed size, and the mapped feature vector is input into the classifier and the bounding box regressor for object classification and position adjustment. By mapping candidate boxes of different sizes to feature vectors of the same size, object detection and classification are achieved.

[0036] Preferably: The method for classifying the collected electrical equipment dataset into a training set and a validation set in step 1 is to collect samples of various electrical equipment images or related data. The collected data is labeled, and the correct label is assigned to each sample. The entire dataset is divided into a training set and a validation set with a ratio of 9:1. The trained improved fast region convolutional neural network is verified through the validation set.

[0037] Another object of the present invention is to provide an electrical equipment classification system for an improved fast region convolutional neural network, used to implement the electrical equipment classification method of the improved fast region convolutional neural network, including a collection unit, a classification unit, a preprocessing unit, an improved fast region convolutional neural network unit, and an output unit, where:

[0038] The collection unit is used to collect samples of electrical equipment images or related data to obtain an electrical equipment dataset.

[0039] The classification unit is used to classify the collected electrical equipment dataset according to the training set.

[0040] The preprocessing unit is used to preprocess the electrical equipment image to obtain a preprocessed picture.

[0041] The improved fast regional convolutional neural network unit is used to input the preprocessed picture into the improved Residual Network 50 to extract features and obtain a feature map. The RPN network is used to select candidate boxes from the feature map and the original image to initially determine the target objects in the image. The RoI pooling layer is used to localize the target objects initially determined in the image, completing the construction of the improved fast regional convolutional neural network. The improved fast regional convolutional neural network is trained with a training set, and target detection and classification are performed using the trained improved fast regional convolutional neural network.

[0042] The output unit is used to output target detection and classification.

[0043] Another object of the present invention is to provide a computer system, including a memory and a processor, where the memory is used to store computer programs / instructions. The processor is used to execute the computer programs / instructions to implement the classification method of electrical equipment for the improved fast regional convolutional neural network.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) Adding a BN layer and performing skip connections deepen the network depth, enabling the network to learn better, recognize and classify images more accurately and quickly. An adaptive learning rate is added during the network learning process, continuously updating the network learning rate to make the network learning effect better. The residual network parameters and the parameters of the used Adam optimizer are optimized to make the learning effect better.

[0046] (2) We have successfully applied our method to the evaluation experiment of the electrical equipment dataset. The improved algorithm can improve the recognition accuracy and speed up the image processing speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is the classification of the collected dataset and the number used for training in each category.

[0048] Figure 2 It is the block diagram of the improved Faster R-CNN system.

[0049] Figure 3 It is the network diagram of the improved ResNet50 feature extraction.

[0050] Figure 4 It is the effect diagram of recognition and classification under complex backgrounds.

[0051] Figure 5 It is the curve of the network loss function during the entire training process.

[0052] Figure 6It is the recognition and classification effect diagram of small target objects contained in the figure. Detailed implementation manners

[0053] The present invention will be further clarified below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, various equivalent modifications made by those skilled in the art fall within the scope defined by the appended claims of this application.

[0054] An electrical equipment classification method based on an improved fast region convolutional neural network, as Figure 1-3 shown, includes the following steps

[0055] Step 1, collect samples of electrical equipment images or related data to obtain an electrical equipment data set. Classify the collected electrical equipment data set according to the training set.

[0056] For the method of classifying the collected electrical equipment data set according to the training set and the validation set, collect samples of various electrical equipment images or related data. As Figure 1 shown, label the collected data, assign the correct label to each sample, and divide the entire data set into a training set and a validation set with a ratio of 9:1. Verify the trained improved fast region convolutional neural network through the validation set.

[0057] Collect samples containing various electrical equipment images or related data. This may include images under different angles and lighting conditions, etc. Ensure the diversity of the data set so that the trained model has good generalization ability. Label the collected data, assign the correct label to each sample, and divide the entire data set into a training set and a validation set with a ratio of 9:1. The purpose of such division is to train the model and evaluate the model performance. Prepare for the neural network training of the electrical equipment data set so that the model can perform well in the training and validation stages.

[0058] Step 2, preprocess the electrical equipment images to obtain preprocessed pictures.

[0059] As Figure 2 shown, the specific steps for preprocessing the RGB pictures of the collected electrical equipment images and preparing them for feature extraction by inputting into the network are as follows:

[0060] In the present invention, the RGB images of the collected electrical equipment images are standardized to ensure the consistency of the image information input into the network. Standardization helps to eliminate factors such as different brightness and contrast in the images, making it easier for the network to learn and generalize. Assume the input image is After passing through a convolutional layer with a convolutional kernel of the image is processed into The image has 64 channels, making the image deeper and with more channels, which helps the network better capture the abstract features in the image and improves the classification performance of the model for electrical equipment.

[0061] After passing through the convolutional layer, it enters a max pooling layer, which reduces the spatial dimension of the input by taking the maximum value within the local area of the input. Max pooling is usually used to reduce the spatial size of the data, retain the main features, and reduce the computational complexity of the model. It prepares for further feature extraction by inputting the image into the ResNet50 network next.

[0062] Step 3: Input the preprocessed picture into the improved Residual Network 50 network for feature extraction to obtain a feature map.

[0063] As Figure 3 shown, the specific steps for inputting the preprocessed picture into the improved Residual Network 50 (ResNet50) network for feature extraction to obtain a feature map are as follows:

[0064] When using the ResNet50 residual network for feature extraction, first add a BN layer in the residual block to retain more semantic information, use residual connections to prevent gradients from vanishing during convolution, and design four stages to extract feature information to continuously deepen the network structure, thus obtaining a feature extraction network suitable for fast classification. During the feature extraction process, using the residual network RestNet50, the feature map contains more semantic information. In addition, adding a BN layer and performing skip connections deepen the network depth, making the network learn better and perform image recognition and classification more accurately and quickly. The residual stage consists of different residual blocks connected to each other. The residual block introduces skip connections, adding the input feature and the output feature after multiple transformations, which helps alleviate the gradient vanishing problem and simplifies network training.

[0065] The Residual Network 50 network includes a first convolutional layer, a first pooling layer, a residual network, a first average pooling layer, and a first fully connected layer connected in sequence. The residual network includes four residual blocks connected in sequence, with a BN layer added in the residual block, and skip connections are made between the four residual blocks. Four residual stages are designed in the Residual Network 50 network for feature extraction. The first residual stage consists of 3 residual blocks, the second residual stage consists of 4 residual blocks, the third residual stage consists of 6 residual blocks, and the fourth residual stage consists of 3 residual blocks. During the extraction process, a BN layer is added, and a part of the original image is directly added to the next extraction process.

[0066] The residual block includes a first convolutional layer, a second convolutional layer, and a third convolutional layer connected in sequence. The first convolutional layer increases the dimension of the input electrical equipment diagram. The second convolutional layer extracts features from the input electrical equipment diagram. The third convolutional layer is used to further increase the dimension of the extracted features.

[0067] The BN layer makes the training process more stable and fast by normalizing the input at each layer. At the same time, it has a certain regularization effect to prevent overfitting. It can also accelerate the training of deep neural networks and improve model performance. The calculation formula of the BN layer is as follows:

[0068]

[0069]

[0070]

[0071]

[0072] Among them, is the normalized value, is the length of the batch, is the th sample, is the variance, is to normalize to zero mean and unit variance, is to prevent the denominator from being zero, is the linearly transformed value after normalization, is a scale-learnable parameter, is an offset-operation learnable parameter.

[0073] Performs a linear transformation on the normalized value so that the model can recover the original data distribution, and are learnable parameters for scale and offset operations respectively.

[0074] The residual structure is designed with two convolutions and one convolutional layer. The residual structure first uses convolution to increase the dimension of the input electrical equipment diagram, then extracts features from it through convolution, and finally uses convolution to further increase the dimension in preparation for further feature extraction. During the extraction process, the BN layer is added to directly accumulate part of the original diagram into the next extraction process, improving the network expression ability during the entire extraction process.

[0075] In the convolutional network, the linear representation relationship can be expressed as:

[0076]

[0077] Among them, represents the linear relationship of the layer, represents the weight coefficient of the layer, represents the input of the layer, represents the offset during the convolution process.

[0078] From the above formula, it can be obtained that from the layer to the convolution linear relationship of the layer, after passing through the following activation function, obtains the output of the layer. The activation function introduces a non-linear transformation, enabling the neural network to learn and represent complex non-linear relationships. Without the activation function, a multi-layer neural network is just a combination of a series of linear transformations and cannot capture non-linear features.

[0079] The activation function is:

[0080]

[0081] Among them, represents the activation function, represents the output of the layer.

[0082] And so on, in the layer, the feature values of the input layer are introduced, and the input layer is directly added to the output of the first layer as the input of the second layer. Due to different dimensions, the matrix is introduced to transform the dimensions, and the is included in the BN layer, so the output is expressed as:

[0083]

[0084] Among them, represents the input of the layer, represents the activation function, represents the linear relationship of the layer, represents the introduced matrix.

[0085] The convolution operation after passing through the residual network is:

[0086]

[0087] Among them, represents the th feature map output of the layer, is the activation function, is the receptive field of the input layer, is the th feature map output of the represents the convolution operation, is the convolutional kernel, is the th bias parameter of the th neuron in the

[0088] The residual block structure is designed with two convolutional layers and one convolutional layer. The residual block structure first uses convolution to increase the dimension of the input electrical equipment diagram, and then uses convolution to extract its features. Finally, the dimension is further increased through convolution to prepare for further feature extraction. Each residual stage consists of several residual blocks. The first stage consists of 3 residual blocks, the second stage consists of 4 residual blocks, the third stage consists of 6 residual blocks, and the fourth stage consists of 3 residual blocks. During the extraction process, the BN layer is added, and a part of the original image is directly added to the next extraction process, improving the network expression ability in the whole extraction process.

[0089] Step 4: Through the RPN network, candidate boxes are selected from the feature map and the original image to initially determine the target objects in the image.

[0090] As Figure 4 shown, the specific steps to initially determine the target objects in the image by selecting candidate boxes through the Region Proposal Network (RPN) are as follows:

[0091] Feature maps are extracted through an improved residual network. The RPN network is applied to the feature maps to generate a series of candidate boxes on the image, which are considered likely to contain the target object. The RPN uses multiple predefined anchor boxes with different sizes and aspect ratios, covering various shapes where the target may appear in the image. These anchor boxes slide on the feature maps and generate candidate boxes at each position. For each generated candidate box, the RPN outputs the corresponding bounding box regression information to adjust the position of the anchor box to better fit the actual position of the target. The RPN also classifies each candidate box to determine whether the box contains the target object. Usually, binary classification is used, that is, the target exists or does not exist. For all candidate boxes that have undergone classification and regression, the non-maximum suppression algorithm is used to remove highly overlapping boxes and retain boxes with higher confidence. The RPN efficiently provides candidate boxes in the image that may contain the target. These candidate boxes are then passed to the subsequent object detection module for more refined object classification and position adjustment.

[0092] Step 5: Use the Region of Interest (RoI) pooling layer to localize the target object initially determined in the image, and complete the construction of the improved fast region convolutional neural network.

[0093] The method for localizing the target object in the preliminary determined image using the RoI pooling layer is as follows: After generating candidate boxes using the RPN, the next step is usually to localize these candidate boxes through the RoI Pooling Layer to complete object detection and classification. For each candidate box generated by the RPN, the features of the corresponding region are extracted from the feature map. This step is called RoI Pooling. The purpose of RoI Pooling is to map candidate boxes of different sizes into feature map regions of the same size for input into the subsequent classifier. The specific process of RoI Pooing is as follows: Each candidate box is evenly divided into sub-regions of a fixed size; the feature values within each sub-region are pooled (usually max pooling) to obtain an output of a fixed size; these outputs are concatenated into a fixed-length vector as the localized feature of the candidate box. RoI Pooling maps each candidate box into a feature vector of a fixed size, so that regardless of the original size of the candidate box, it can be input into the subsequent classifier. The mapped feature vector is input into the classifier and the bounding box regressor for object classification and position adjustment. Usually, these classifiers and regressors are connected to a shared fully connected layer to extract and learn the features of the candidate boxes. By mapping candidate boxes of different sizes into feature vectors of the same size through RoI Pooling, object detection and classification are achieved.

[0094] Step 6: Train the improved fast regional convolutional neural network using the training set, and perform object detection and classification using the trained improved fast regional convolutional neural network.

[0095] In another embodiment, an electrical equipment classification system for an improved fast regional convolutional neural network is provided, which is used to implement the electrical equipment classification method of the improved fast regional convolutional neural network, and includes an acquisition unit, a classification unit, a preprocessing unit, an improved fast regional convolutional neural network unit, and an output unit, where:

[0096] The acquisition unit is used to acquire samples of electrical equipment images or related data to obtain an electrical equipment data set.

[0097] The classification unit is used to classify the acquired electrical equipment data set according to the training set.

[0098] The preprocessing unit is used to preprocess the electrical equipment image to obtain a preprocessed picture.

[0099] The improved fast regional convolutional neural network unit is used to input the preprocessed picture into the improved Residual Network 50 network for feature extraction to obtain a feature map. The RPN network is used to select candidate boxes in the feature map and the original image to initially determine the target objects in the image. The RoI pooling layer is used to localize the target objects initially determined in the image, and the construction of the improved fast regional convolutional neural network is completed. The improved fast regional convolutional neural network is trained with a training set, and target detection and classification are performed through the trained improved fast regional convolutional neural network.

[0100] The output unit is used to output target detection and classification.

[0101] In another embodiment, a computer system is provided, including a memory and a processor. The memory is used to store computer programs / instructions. The processor is used to execute the computer programs / instructions to implement the classification method of electrical equipment for the improved fast regional convolutional neural network.

[0102] As Figure 5 shown, it is the network loss function curve during the entire training process. As Figure 6 shown, the figure contains the recognition and classification effect diagrams of small target objects. The present invention improves the Faster-RCNN network, optimizes the network depth, enhances the ability to retain deep semantic information, and thus improves the recognition accuracy and speed of electrical equipment under complex backgrounds. The present invention not only focuses on improving the detection accuracy of small targets under complex backgrounds, but also effectively optimizes the algorithm computing power, improves the recognition speed and the image processing speed.

[0103] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An improved fast regional convolutional neural network electrical equipment classification method, characterized in that: The following steps are included Step 1, collecting samples of electrical equipment images or related data to obtain an electrical equipment data set; classifying the collected electrical equipment data set according to the training set; Step 2, preprocessing the electrical equipment image to obtain a preprocessed image; Step 3, input the preprocessed image into the improved Residual Network 50 network for feature extraction to obtain a feature map; The Residual Network 50 network includes a first convolutional layer, a first pooling layer, a residual network, a first average pooling layer, and a first fully connected layer connected in sequence, and the residual network includes four residual blocks connected in sequence, a BN layer is added to the residual block, and jump links are performed between the four residual blocks; Four residual stages are designed in the Residual Network 50 network for feature extraction. The first residual stage consists of 3 residual blocks, the second residual stage consists of 4 residual blocks, the third residual stage consists of 6 residual blocks, and the fourth residual stage consists of 3 residual blocks. The BN layer is added in the extraction process to directly add part of the original image to the next extraction process. The convolution operation after the residual network is: in, Indicates Layer feature map output, is the activation function, is the receptive field of the input layer, for Layer feature map output, represents the convolution operation, is the convolution kernel, For the The feature map of the layer The bias parameters of each neuron; The calculation formula of the BN layer is as follows: in, is the standardized value, is the length of the batch, It is samples, is the variance, Yes Normalized to zero mean and unit variance, To prevent the denominator from being zero, is the standardized linear transformation value, is a scale-learnable parameter, is a learnable parameter of the offset operation; Step 4: The RPN network is used to select candidate boxes from the feature map and the original image to preliminarily determine the target object in the image; Step 5: Use the RoIpoing layer to localize the target object in the preliminarily determined image and complete the construction of the improved fast regional convolutional neural network; Step 6: Train the improved fast regional convolutional neural network using the training set, and perform target detection and classification using the trained improved fast regional convolutional neural network.

2. The electrical equipment classification method of the improved fast regional convolutional neural network according to claim 1 is characterized in that: The residual block includes a first convolution layer, a second convolution layer, and a third convolution layer connected in sequence. The first convolution layer increases the dimension of the input electrical equipment diagram, the second convolution layer extracts features of the input electrical equipment diagram, and the third convolution layer is used to further increase the dimension of the extracted features.

3. The electrical equipment classification method of the improved fast regional convolutional neural network according to claim 2 is characterized in that: The linear representation relationship in the convolutional network can be expressed as: in, Represents The linear relationship of the layers, express The weight coefficient of the layer, express The input of the layer, Represents the offset during the convolution process; The activation function is: in, represents the activation function, express The output of the layer; Introducing the Matrix Transform the dimension, Contained in the BN layer, the output is represented as: in, express The input of the layer, represents the activation function, Represents The linear relationship of the layers, Indicates the introduction of matrix.

4. The electrical equipment classification method of the improved fast regional convolutional neural network according to claim 3 is characterized in that: The method of using RoIpooing layer to localize the target object in the preliminary determined image in step 5 is as follows: after using RPN to generate candidate boxes, the candidate boxes are localized by RoI Pooling Layer to complete target detection and classification; for each candidate box generated by RPN, the features of the corresponding area are extracted from the feature map; each candidate box is evenly divided into sub-regions of fixed size; the feature values ​​in each sub-region are pooled to obtain a fixed-size output; these outputs are connected into a fixed-length vector as the localized feature of the candidate box; each candidate box is mapped to a feature vector with a fixed size, and the mapped feature vector is input into the classifier and bounding box regressor to classify and adjust the position of the target; by mapping candidate boxes of different sizes to feature vectors of the same size, the detection and classification of the target is achieved.

5. The electrical equipment classification method of the improved fast regional convolutional neural network according to claim 4 is characterized in that: Step 1 collects samples of various electrical equipment images or related data by classifying the collected electrical equipment data set according to the training set and verification set; annotates the collected data, assigns a correct label to each sample, and divides the entire data set into a training set and a verification set with a ratio of 9:1; verifies the trained improved fast regional convolutional neural network through the verification set.

6. An improved fast regional convolutional neural network electrical equipment classification system, characterized by: An electrical equipment classification method for implementing the improved fast regional convolutional neural network described in any one of claims 1 to 5, comprising a collection unit, a classification unit, a preprocessing unit, an improved fast regional convolutional neural network unit, and an output unit, wherein: The acquisition unit is used to acquire samples of electrical equipment images or related data to obtain an electrical equipment data set; The classification unit is used to classify the collected electrical equipment data set according to the training set; The preprocessing unit is used to preprocess the electrical equipment image to obtain a preprocessed image; The improved fast regional convolutional neural network unit is used to input the preprocessed image into the improved ResidualNetwork 50 network for feature extraction to obtain a feature map; select candidate boxes between the feature map and the original image through the RPN network to preliminarily determine the target object in the image; use the RoIpooing layer to localize the target object in the preliminarily determined image to complete the construction of the improved fast regional convolutional neural network; train the improved fast regional convolutional neural network through the training set, and perform target detection and classification through the trained improved fast regional convolutional neural network; The output unit is used to output target detection and classification.

7. A computer system, characterized in that: It comprises a memory and a processor, wherein the memory is used to store computer programs / instructions; and the processor is used to execute the computer programs / instructions to implement the electrical equipment classification method of the improved fast regional convolutional neural network as described in any one of claims 1-6.

Images