Insulator defect detection method based on multi-scale convolution kernel parallel learning

Abstract

The invention provides an insulator defect detection method based on unmanned aerial vehicle vision multi-scale convolution kernel parallel learning, and aims at the characteristics and problems of insulator defect detection, the structural characteristics of a convolutional neural network are deeply studied. Local detail features are extracted through small-size (such as 3 * 3) convolution kernels to retain fine-grained information of insulator defects, a plurality of deep convolution kernels of different sizes are applied to be connected in parallel to capture information of different sizes and contextual information, a receptive field of a model is expanded, the contextual information is obtained by introducing a context anchor attention mechanism (CAA), and the probability of the insulator defects is reduced. And the one-dimensional convolution kernel is used to fuse different scale feature maps with the CAA output feature map, and different scale defect targets and target context information are fully extracted, so that the feature dimensions representing the defect targets and parameters of a learning model are reduced, the extraction capability of different scale target features is enhanced, and the detection precision of insulator defects is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an insulator defect detection method based on multi-scale convolution kernel parallel learning. Background Art

[0002] Electricity is the lifeblood of modern society, and the security and stability of its supply are directly linked to the smooth operation of the economy and society. With the continuous development of the national economy, the demand for electricity across all industries is increasing, leading to the ever-increasing length of transmission lines. Transmission lines, as the cornerstone of the power grid, bear the heavy responsibility of power transmission. Their safety, stability, and reliability are crucial to ensuring a secure and stable power supply. Insulators are essential components of high-voltage transmission systems, used to electrically isolate and support wires. They are essential components of transmission lines. Their primary function is to ensure the stable suspension of wires, connect wires and cables, and provide necessary electrical isolation to protect transmission lines from current leakage. Insulators are numerous, diverse, and widely distributed, exposed to the elements for long periods of time. Insulators are subject to complex geological conditions, harsh climates, and environmental influences, leading to defects such as spontaneous string drop, burning, corrosion, and aging damage. These insulator defects can cause short circuits, electrical discharges, fires, and other serious problems that can severely impact the safe and stable operation of transmission lines. Prompt detection of insulator failures through inspections is crucial for ensuring transmission line safety.

[0003] With the advancement of computing power and continuous algorithm optimization, deep learning theory has achieved significant breakthroughs in the field of object recognition. Object detection methods based on deep learning have begun to be widely used in image recognition for power equipment. Although deep learning has achieved initial success in insulator defect detection, practical applications still face several challenges: the impact of complex and changing natural environments (such as varying lighting conditions, complex backgrounds, and inclement weather) on image quality; the scarcity of insulator defect data samples, coupled with issues such as small pixel ratios, complex backgrounds, and line occlusion, all contribute to low insulator defect detection accuracy based on deep learning methods; the small size of defect targets in drone vision, the significant variation in the sizes of different insulator detection targets, and the varying scales and resolutions of drone vision targets at different altitudes. These factors lead to limited adaptability of deep learning detection models, making detection results prone to missed and false detections. These challenges place higher demands on the robustness, accuracy, and reliability of insulator depth detection models and algorithms. New insulator defect detection technologies are urgently needed to improve power grid inspection technology and ensure the safe operation of transmission lines.

[0004] In view of this, an insulator defect detection method based on multi-scale convolution kernel parallel learning is proposed. Summary of the Invention

[0005] In view of the fact that the existing insulator defect depth detection model in the prior art does not extract sufficient information when facing multi-scale and multi-resolution drone visual target images, it cannot fully and efficiently utilize the semantic information and contextual information of the multi-scale convolution feature map in the insulator defect target detection, resulting in the problem of low accuracy of insulator defect detection and recognition in complex scenarios. The present invention provides an insulator defect detection method using multi-scale convolution kernel parallel learning, which can extract local detail features through small-size (such as 3×3) convolution kernels to retain the fine-grained information of insulator defects, and use a set of convolution kernels of different sizes in parallel to capture information of different scales, expand the receptive field of the model, obtain contextual information by introducing the context anchor attention mechanism (CAAM), and apply one-dimensional convolution kernels to fuse different scale feature maps with them, fully extracting targets of different scales and target context information, which not only reduces the feature dimension and model parameters, but also enhances the extraction capability of target features of different scales, greatly improving the detection accuracy of insulator defects. The specific technical solution is as follows:

[0006] An insulator defect detection method using multi-scale convolution kernel parallel learning includes the following steps:

[0007] Construct a typical defect sample dataset and clean and manually annotate the collected dataset, dividing it into training dataset, validation dataset, and test dataset according to training and testing requirements;

[0008] Using a lightweight VGG network as the backbone, an initial module with a small convolution kernel extracts local detail features. A set of deep convolution kernels of different sizes is used to capture contextual information at different scales. A contextual anchor attention mechanism is introduced to enhance central features. A convolution kernel is then applied to fuse multi-scale feature maps. Based on this, a multi-scale insulator defect target detection network is constructed.

[0009] Position detection is used to obtain the detection space area of ​​interest, and insulator defect prediction classification and bounding box regression are performed based on insulator feature maps of different resolutions;

[0010] Initialize the parameters of each layer of the network and use an end-to-end strategy to iteratively optimize and train the model.

[0011] Preferably, the data set includes more than one string of insulators, multiple types of insulators including porcelain insulators, glass insulators, and composite insulators, and several sub-images of insulators at multiple angles, including positive samples and negative samples, where the positive samples are insulators containing defective blocks.

[0012] Preferably, for the positive sample insulator image, its spatial position and spatial area are marked; for the defective insulator sample, the spatial position and spatial area of ​​the defective block are marked.

[0013] Preferably, the process of constructing a multi-scale insulator defect target detection network is as follows:

[0014] The model contains three 3×3 convolutional layers, and five convolutions in parallel: 5×5, 7×7, 9×9, 11×11, and 13×13. Each convolution has a different size, providing information at different scales.

[0015] Combine large convolution kernels with small convolution kernels, use small convolution kernels to capture local features, use large convolution kernels to obtain global information, and finally merge the output into a small convolution kernel operation;

[0016] The output features of convolution kernels of different sizes are fused through a Conv 1×1 convolution to obtain feature maps for characterizing insulators of different scales;

[0017] The contextual feature maps extracted by the CAAM attention mechanism are fused through a Conv 1×1 convolution to obtain a fusion feature that integrates global, multi-scale and contextual feature maps.

[0018] Preferably, the detection space region of interest is obtained by position detection as follows:

[0019] The threshold IoU of the intersection-over-union ratio between the candidate box area and the real box GT is set in advance as the distance indicator setting indicator;

[0020] Calculate the IoU between each GT and all candidate boxes. For each candidate box, find the GT with the largest IoU. According to whether the IoU value exceeds the set positive sample threshold or is lower than the negative sample threshold, the candidate box is assigned as a positive sample or a negative sample.

[0021] The K-means clustering algorithm is used to calculate K prior regions to form the target frame true value clustering result, and the optimization threshold ε is set to search for the optimal aspect ratio of the prior frame.

[0022] Preferably, the prediction classification and border regression of insulator defects based on insulator feature maps of different resolutions are specifically as follows: the feature map corresponding to each candidate region box is obtained by 1×1 convolution fusion to obtain a fused feature map, which is then input into Softmax for classification and border regression.

[0023] Preferably, during the model training process, the loss function LLOSS value is calculated using the stochastic gradient descent method, and the gradient of the network parameters of each layer is obtained by back propagation to update the parameters of the learning network of each layer.

[0024] Preferably, model validation is also included, specifically using a test dataset to validate the trained model, testing the model's recognition capabilities and the network model's generalization capabilities. Since the dataset constructed by the drone is a small dataset, the ratio of the training set, validation set, and test set is 6:2:2.

[0025] A computer-readable storage medium includes a stored program, wherein when the program is run, the device where the computer-readable storage medium is located is controlled to execute the insulator defect detection method of multi-scale convolution kernel parallel learning as described above.

[0026] A processor is used to run a program, wherein when the program is run, the insulator defect detection method using multi-scale convolution kernel parallel learning as described above is executed.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. The present invention uses convolution kernels of different sizes in parallel to provide information of different scales, making it suitable for target detection of insulator defect images of different scales in UAV inspection vision. It also combines large and small convolution kernels, using the small convolution kernel to capture local features and the large convolution kernel to obtain global information, and finally merges the output into a small convolution kernel operation. While maintaining feature extraction efficiency, it can also effectively utilize the respective advantages of large and small convolution kernels in training.

[0029] 2. The present invention adopts small convolution to fuse global, multi-scale and contextual feature maps, and introduces the K-means clustering algorithm in the non-object prior module that guides the search for the spatial position of insulators in the source image, so as to realize the fast optimization search of customized anchor selection boxes, thereby reducing false positives and improving the detection accuracy of small defective targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly describes the drawings required for the specific embodiments or the description of the prior art. Similar elements or parts are generally identified by similar reference numerals throughout the drawings. Elements or parts in the drawings are not necessarily drawn to scale.

[0031] Figure 1 A diagram of the model structure provided by an embodiment of the present invention;

[0032] Figure 2 This is a schematic diagram of insulator defect detection results proposed in an embodiment of the present invention;

[0033] Figure 3 This is a schematic diagram of the insulator defect detection results proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0035] It will be understood that when used in this specification and the appended claims, the terms “comprises” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0036] It should also be understood that the terms used in the present specification are only for the purpose of describing particular embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0037] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0038] The embodiment of the present invention uses the VGG network (Visual Geometry Group Network) as the backbone network, uses convolution kernels of different sizes to capture contextual information of the target at different scales, introduces the contextual anchor attention mechanism (CAAM) to enhance the central features, and fuses the multi-scale features to establish a feature extraction network. The K-means clustering algorithm is used to generate the target prior candidate box (Non-object prior), and the output results of adjacent levels are combined through jump-layer links to achieve feature sharing. The insulator feature maps of different resolutions are input into the Softmax classifier to realize the prediction classification and border regression of insulator defects. It is capable of extracting large-scale changes and diversified contextual information of drone visual targets under complex backgrounds, reducing the impact of different scale changes of insulators under drone vision, and enriching semantic information. The specific implementation methods are as follows:

[0039] Step 1: Dataset Construction: Deep learning-based insulator defect detection requires a large amount of labeled sample data for training, so a dataset of typical defect samples must be constructed. Considering different flight altitudes and application scenarios, drones are used to collect a large number of datasets of normal and defective insulators, ensuring data diversity and multi-scale variations. The collected datasets are then cleaned and manually annotated, and divided into training, validation, and test datasets based on training and testing requirements.

[0040] The dataset includes 1,000 images of insulators with more than one string, including various types of insulators, such as porcelain, glass, and composite insulators, and insulators at various angles. Of these, 700 are positive samples (including defective blocks) and 300 are negative samples. In addition, insulator image data was collected through online searches. For positive insulator images, their spatial location and area are annotated. For defective insulator samples, the spatial location and area of ​​the defective blocks are also annotated, expressed as x and y coordinates.

[0041] Step 2. Model construction: Build a VGG-like structure, first using a set of convolutions to obtain local image information; then immediately connect a set of parallel convolutions of different sizes to extract image target information at different scales; introduce the contextual attention mechanism CAAM module to capture the contextual information of the drone's visual target; then fuse the outputs of different scales with the CAAM output, combining local and global information to enhance feature expression.

[0042] As a specific embodiment, the present invention uses a target detection basic network based on multi-scale features and context information fusion feature maps to complete the target detection task of insulator defects, and can realize the positioning and classification of multi-scale targets in the insulator dataset according to the pre-defined default box. Using a VGG-like framework as the main skeleton model, the model contains 3 3×3 convolution layers and 5 convolutions in parallel: 5×5, 7×7, 9×9, 11×11, 13×13. Each convolution has a different size and provides information at different scales. Here, large convolution kernels and small convolution kernels are combined, and small convolution kernels are used to capture local features, and large convolution kernels are used to obtain global information, and finally merged into a small convolution kernel operation. While maintaining the efficiency of feature extraction, the respective advantages of large and small convolution kernels in training can also be effectively utilized. The structure is as follows: Figure 1 shown.

[0043] The output features of convolution kernels of different sizes are fused through a Conv 1×1 convolution to obtain feature maps for characterizing insulators of different scales, and then fused with the context feature maps extracted by the CAAM attention mechanism through a Conv 1×1 convolution to obtain fused features that integrate global, multi-scale and context feature maps.

[0044] Step 3: Target prior: Apply the non-object prior module of the partition-based K-means clustering algorithm to achieve this task. It is used to guide the algorithm to search for the spatial location of insulators in the source image to reduce false positives and improve detection accuracy.

[0045] A threshold for the Interset of Union (IOU) between the candidate box (anchor or proposal) and the ground truth box (GT) is set in advance as a distance indicator, and positive and negative samples are determined based on this threshold. The specific steps include: calculating the Intersection over Union (IoU) between each GT and all candidate boxes. For each candidate box, find the GT with the largest IoU. Based on whether the IoU value exceeds the set positive sample threshold or is lower than the negative sample threshold, the candidate box is assigned as a positive or negative sample.

[0046] The K-means clustering algorithm is used to calculate K prior regions to form the ground truth clustering result of the target box. The optimization threshold ε is set to search for the optimal aspect ratio of the prior box. The detailed algorithm is as follows:

[0047] Step 1. Initialize cluster centers

[0048] Randomly select I data points as the initial cluster center v, and the cluster center initialization matrix is ​​V (0)

[0049] V (0) =v{v1,v2,…v I}

[0050] Step 2. Assign data points to the nearest cluster center

[0051] For each data sample point (x), calculate its distance from each cluster center and assign it to the nearest cluster center. Then the distance from the kth data sample point to the i-th cluster center is:

[0052]

[0053] Here i Represents the i-th cluster center. m represents the dimension of the sample data point.

[0054] Step 3. Recalculate cluster centers

[0055] For each cluster, calculate the mean of all its data points and use the mean as the new cluster center.

[0056] Step 4. Repeat steps 2 and 3 until the cluster center no longer changes

[0057] Repeat steps 2 and 3 until the cluster center no longer changes significantly, that is, the change in the cluster center is less than the set threshold ε, where n represents the nth iteration.

[0058] ‖V (n) -V (n+1) ‖<ε

[0059] Step 5. Output the final clustering results

[0060] Output the cluster to which each data point belongs and the final cluster center.

[0061] Step 4, classification and regression: Based on the detection space region of interest obtained in step 3, the integrated feature maps of insulators of different scales are input into the Softmax classifier to realize the detection and recognition of insulator defects and the regression of the border. The feature maps corresponding to each candidate region box obtained in the above steps are fused by 1×1 convolution to obtain the fused feature map, which is then input into the Softmax for classification and border regression. The present invention uses the typical classification loss function L CLS And the typical regression loss function L REG The weighted sum of is used as the total model training loss function.

[0062] L LOSS =λ1L CLS +λ2L REG

[0063]

[0064] L cls (p i ,p i * )=-log[p i * p i +(1-p i * )(1-p i )]

[0065] L reg (v i ,v i * )=smooth L1 (v i -v i * )

[0066]

[0067] Among them, p i is the probability of Anchor prediction as the target, i is the index in the candidate box, p i *∈{0,1} is an indicator parameter. When it is 1, it means that the prior box matches the true value. i is the predicted coordinate vector, v i * is the real coordinate vector. The purpose of setting Smooth in this way is to make the loss more robust to outliers. Compared with other loss functions, it is not sensitive to outliers (points far from the center) and outliers. It can control the magnitude of the gradient and is beneficial to model training. λ1 and λ2 are hyperparameters. N cls is the normalization parameter of the classification item, N reg is the normalization parameter of the regression term.

[0068] For each predicted box, the classification loss function is used to determine whether there are insulators in the image, negative samples are filtered out, and predicted boxes that do not exceed the threshold are filtered according to the target detection confidence. Finally, the non-maximum suppression (NMS) algorithm is used to filter out predicted boxes with large overlap rates to obtain the corresponding classification confidence and box position value.

[0069] Step 5, model training: Initialize the parameters of each layer of the network and use an end-to-end strategy to iteratively optimize the model. During the model training process, calculate the loss function L LOSS The stochastic gradient descent method is used to calculate the value, and the gradient of each layer of network parameters is obtained by back propagation to update the parameters of each layer of learning network.

[0070] Stochastic gradient descent was used as the basic iterator, with momentum set to 0.9 and an initial learning rate of 0.001. To mitigate instability in the Stochastic Gradient Descent (SGD) optimizer during later training, the learning rate was appropriately reduced to 0.0001. An improved training strategy was employed to search the negative sample space. During the model training phase, multiple batches of images were randomly selected from the dataset to construct the training sample dataset, with a positive-to-negative sample ratio of 1:5. Furthermore, all insulator scales were categorized into 10 scales based on expert experience, ensuring that at least one selected region corresponded to each scale.

[0071] Step 6: Model Validation: Use the test dataset to validate the trained model and test the model's recognition and generalization capabilities. Since the dataset constructed using drones is small, the ratio of the training set, validation set, and test set is 6:2:2.

[0072] This patent uses objective and universal accuracy, recall, and mAP (high-quality model accuracy) metrics to evaluate the model's detection performance. Accuracy is the proportion of correct data predicted as true; recall represents the recall rate, or the probability of predicting the correct category box; and mAP@0.5 represents the average detection accuracy of all detection categories when the Intersection over Union (IoU) threshold is 0.5.

[0073]

[0074] AP is the precision-recall curve, which is the area enclosed by the PR curve and the coordinate axes. In the PR curve, P is the precision, or accuracy, and R is the recall, or recall. P(r) represents the smoothing of the PR curve. C is the number of detection categories.

[0075] Figure 2 、 Figure 3 The insulator defect detection results of this patented method are shown:

[0076] The white box shows the insulator string drop detection and recognition results. As can be seen from the figure, insulator defects can be detected and recognized with high accuracy despite processing image backgrounds of varying scales. Table 1 compares the detection and recognition rates of different methods.

[0077] Table 1 Comparison of detection and recognition rates of different methods for insulator string drop

[0078]

[0079] In summary, the present invention uses convolution kernels of different sizes in parallel to provide information of different scales, making it suitable for target detection of insulator defect images of different scales in drone inspection vision. It also combines large and small convolution kernels, using small convolution kernels to capture local features and large convolution kernels to obtain global information, and finally merges the output into a small convolution kernel operation. While maintaining feature extraction efficiency, it can also effectively utilize the respective advantages of large and small convolution kernels in training. In addition, while using small convolution to fuse global, multi-scale and contextual feature maps, the present invention introduces the K-means clustering algorithm in the non-object prior module that guides the search for the spatial position of insulators in the source image, realizing a customized anchor selection box for rapid optimization search, thereby reducing false positives and improving the detection accuracy of small defective targets.

[0080] Those skilled in the art will appreciate that the units of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition of each example has been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0081] In the embodiments provided by the present invention, it should be understood that the division of units is merely a logical function division, and there may be other division methods in actual implementation, for example, multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored, etc.

[0082] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0083] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-0nly Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk, etc., various media that can store program code.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A method for detecting insulator defects using multi-scale convolution kernel parallel learning, characterized in that: The following steps are involved: Construct a typical defect sample dataset and clean and manually annotate the collected dataset, dividing it into training dataset, validation dataset, and test dataset according to training and testing requirements; Using a lightweight VGG network as the backbone, an initial module with a small convolution kernel extracts local detail features. A set of deep convolution kernels of different sizes is used to capture contextual information at different scales. A contextual anchor attention mechanism is introduced to enhance central features. A convolution kernel is then applied to fuse multi-scale feature maps. Based on this, a multi-scale insulator defect target detection network is constructed. Position detection is used to obtain the detection space area of ​​interest, and insulator defect prediction classification and bounding box regression are performed based on insulator feature maps of different resolutions; Initialize the parameters of each layer of the network and use an end-to-end strategy to iteratively optimize and train the model.

2. The insulator defect detection method using multi-scale convolution kernel parallel learning according to claim 1 is characterized in that: The dataset contains multiple images of insulators of more than one string, including multiple types of insulators including porcelain insulators, glass insulators, and composite insulators, and multiple angles. It includes positive and negative samples, with positive samples being insulators with defective blocks.

3. The insulator defect detection method using multi-scale convolution kernel parallel learning according to claim 2 is characterized in that: For positive sample insulator images, their spatial position and spatial area are marked; for defective insulator samples, the spatial position and spatial area of ​​the defective block are marked.

4. The insulator defect detection method using multi-scale convolution kernel parallel learning according to claim 1 is characterized in that: The process of constructing a multi-scale insulator defect target detection network is as follows: The model contains three 3×3 convolutional layers, and five convolutions in parallel: 5×5, 7×7, 9×9, 11×11, and 13×13. Each convolution has a different size, providing information at different scales. Combine large convolution kernels with small convolution kernels, use small convolution kernels to capture local features, use large convolution kernels to obtain global information, and finally merge the output into a small convolution kernel operation; The output features of convolution kernels of different sizes are fused through a Conv 1×1 convolution to obtain feature maps for characterizing insulators of different scales; The contextual feature maps extracted by the CAAM attention mechanism are fused through a Conv 1×1 convolution to obtain a fusion feature that integrates global, multi-scale and contextual feature maps.

5. The insulator defect detection method using multi-scale convolution kernel parallel learning according to claim 1, characterized in that: The detection space area of ​​interest obtained by position detection is as follows: The threshold IoU of the intersection-over-union ratio between the candidate box area and the real box GT is set in advance as the distance indicator setting indicator; Calculate the IoU between each GT and all candidate boxes. For each candidate box, find the GT with the largest IoU. According to whether the IoU value exceeds the set positive sample threshold or is lower than the negative sample threshold, the candidate box is assigned as a positive sample or a negative sample. The K-means clustering algorithm is used to calculate K prior regions to form the target frame true value clustering result, and the optimization threshold ε is set to search for the optimal aspect ratio of the prior frame.

6. The insulator defect detection method using multi-scale convolution kernel parallel learning according to claim 1, characterized in that: The prediction classification and bounding box regression of insulator defects based on insulator feature maps of different resolutions are as follows: the feature map corresponding to each candidate region box is fused through 1×1 convolution to obtain a fused feature map, which is then input into Softmax for classification and bounding box regression.

7. The insulator defect detection method using multi-scale convolution kernel parallel learning according to claim 1, characterized in that: During model training, the loss function L is calculated LOSS The stochastic gradient descent method is used to calculate the value, and the gradient of each layer of network parameters is obtained by back propagation to update the parameters of each layer of learning network.

8. The insulator defect detection method using multi-scale convolution kernel parallel learning according to claim 1, characterized in that: This also includes model validation, specifically using a test dataset to verify the trained model and test the model's recognition capabilities and the network model's generalization capabilities. Because the dataset constructed using drones is small, the ratio of the training set, validation set, and test set is 6:2:

2.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored program, wherein when the program is run, the device where the computer-readable storage medium is located is controlled to execute the insulator defect detection method of multi-scale convolution kernel parallel learning according to any one of claims 1 to 8.

10. A processor, characterized in that: The processor is used to run a program, wherein the program, when running, executes the insulator defect detection method using multi-scale convolution kernel parallel learning as described in any one of claims 1 to 8.

Images