Electric power fitting detection method, system, equipment and medium

Abstract

The invention discloses an electric power fitting detection method, system and device and a medium, and belongs to the technical field of image detection. According to the method, the cross-layer connection fusion network and the decoupling adaptive feature fusion structures are constructed, the decoupling adaptive feature fusion network is constructed based on the cross-layer connection fusion network and the decoupling adaptive feature fusion structures, then the deep learning model is updated, and the electric power fitting detection model is obtained. The electric power fitting detection of the to-be-detected power transmission line inspection image based on the electric power fitting detection model is realized. According to the invention, by constructing the decoupling adaptive feature fusion network and integrating cross-layer connection and an adaptive feature weighting mechanism, the recognition capability of the electric power fitting detection model on fine features and complex scenes of the electric power fitting is enhanced, so that the defect of low accuracy of a traditional deep learning model in the detection of the electric power fitting can be overcome, and the detection accuracy of the electric power fitting is improved. The detection accuracy of the electric power fitting under a complex background can be remarkably improved, and the technical problem of low electric power detection accuracy in the prior art can be solved.

Description

Technical Field

[0001] The present invention belongs to the field of image detection technology, and in particular relates to a method, system, equipment and medium for detecting electric hardware. Background Art

[0002] With the accelerated development of smart grids, the current mode of transmission line inspection has shifted from traditional manual inspections to an intelligent, collaborative operation using drones, robots, and helicopters. For example, workers operate drones or autonomously patrol transmission lines, while the drones' cameras capture massive amounts of high-resolution visible light images, including those of electrical hardware. Electrical hardware is a metal accessory that connects and combines various devices in the power system, playing a critical role in the transmission and distribution system. These images may reveal defects such as corrosion, deformation, or missing bolts, requiring workers to promptly identify and replace them as needed. Therefore, inspecting electrical hardware from these massive images is crucial to maintaining the stable operation of transmission lines.

[0003] Currently, the detection of electrical fittings in images mainly relies on manual inspection and methods based on traditional machine learning. Manual inspection is prone to missed detections and false detections when faced with massive amounts of images; while methods based on traditional machine learning have unsatisfactory detection accuracy due to factors such as the highly complex background in the image. Currently, detection methods based on deep learning have gradually been applied to the detection of electrical fittings. However, due to the complex background of the image and the small proportion of electrical fittings, the dense connections and mutual occlusion between electrical fittings, and the similarity between different models of the same electrical fitting, the current deep learning-based detection methods are unable to effectively detect electrical fittings and have low detection accuracy. Therefore, there is an urgent need for an electrical fitting detection method, system, equipment, and medium to address the shortcomings of the existing technology. Summary of the Invention

[0004] The present invention aims to provide a method, system, equipment and medium for detecting electric fittings to solve the technical problem of low accuracy of electric power detection in the existing technology. By constructing a decoupled adaptive feature fusion network and updating the deep learning model, the electric fitting detection model is used to realize the detection of electric fittings in transmission line inspection images, thereby improving the accuracy of electric fitting detection.

[0005] In order to solve the above technical problems, an embodiment of the present invention provides a method for detecting electric hardware, comprising:

[0006] Obtain inspection images and deep learning models of the transmission lines to be inspected;

[0007] Constructing a cross-layer connection fusion network and several decoupling adaptive feature fusion structures, and determining a decoupling adaptive feature fusion network based on the cross-layer connection fusion network and the several decoupling adaptive feature fusion structures;

[0008] The deep learning model is updated based on the decoupled adaptive feature fusion network to obtain a power fitting detection model;

[0009] The inspection image of the transmission line to be inspected is input into the power fitting inspection model to obtain the power fitting inspection result.

[0010] It can be understood that, compared with the prior art, the present invention constructs a cross-layer connection fusion network and several decoupled adaptive feature fusion structures, and constructs a decoupled adaptive feature fusion network based on the cross-layer connection fusion network and several decoupled adaptive feature fusion structures, thereby updating the deep learning model to obtain an electric fitting detection model, thereby realizing the detection of electric fittings based on the electric fitting detection model for the inspection image of the transmission line to be inspected. The present invention enhances the recognition ability of the electric fitting detection model for subtle features and complex scenes of electric fittings by constructing a decoupled adaptive feature fusion network, integrating cross-layer connection and adaptive feature weighting mechanism, thereby making up for the low accuracy of traditional deep learning models in the detection of electric fittings, significantly improving the detection accuracy of electric fittings under complex backgrounds, effectively avoiding the problems of missed detection and false detection of electric fittings, and effectively improving the detection accuracy of similar small-target electric fittings in the inspection image of the transmission line.

[0011] As a preferred solution, the deep learning model includes: a feature extraction network; the construction of a cross-layer connection fusion network and several decoupled adaptive feature fusion structures, and the construction of a decoupled adaptive feature fusion network based on the cross-layer connection fusion network and the several decoupled adaptive feature fusion structures, including:

[0012] Constructing a top-down path and a bottom-up path according to the number of feature map outputs of the feature extraction network, and constructing a cross-layer connection fusion network based on the top-down path and the bottom-up path;

[0013] Constructing a plurality of decoupling adaptive feature fusion structures according to the number of feature graphs output by the feature extraction network, each of the decoupling adaptive feature fusion structures being used to perform feature decoupling and feature fusion on the first power fitting feature graph output by the cross-layer connection fusion network;

[0014] A decoupled adaptive feature fusion network is constructed based on the cross-layer connection fusion network and several decoupled adaptive feature fusion structures.

[0015] This preferred solution constructs a cross-layer connection fusion network of top-down paths and bottom-up paths, and constructs several decoupled adaptive feature fusion structures for feature decoupling and feature fusion. Through the top-down and bottom-up path construction of the cross-layer connection fusion network, it can fully integrate feature information at different levels and enhance the expressive ability of features; and the decoupled adaptive feature fusion structure can specifically decouple and analyze the features of electrical fittings and perform fusion processing, so as to better extract the features of electrical fittings, improve the performance of subsequent electrical fitting detection models, and thereby improve the accuracy of electrical fitting detection.

[0016] As a preferred solution, the deep learning model further includes: an FPN structure and a detection module; the updating of the deep learning model based on the decoupled adaptive feature fusion network to obtain a power fitting detection model includes:

[0017] The decoupled adaptive feature fusion network replaces the FPN structure, and an initial power fitting detection model is constructed based on the feature extraction network, the cross-layer connection fusion network, and several decoupled adaptive feature fusion structures and detection modules;

[0018] The initial electric hardware detection model is trained based on a preset deep learning model training method to obtain an electric hardware detection model.

[0019] This preferred solution replaces the FPN structure with a decoupled adaptive feature fusion network and then performs model training to obtain an electric fitting detection model. This can give full play to the advantages of the decoupled adaptive feature fusion network and improve the electric fitting detection model's ability to process electric fitting features. Training is based on a preset deep learning model training method to ensure that the electric fitting detection model can learn effective features and converge, thereby obtaining an electric fitting detection model with better performance. This not only improves the detection accuracy of the electric fitting detection model, but also enhances the adaptability and robustness of the electric fitting detection model, making it more suitable for actual transmission line inspection image detection scenarios and improving the accuracy of electric fitting detection.

[0020] As a preferred solution, the initial power fitting detection model is trained based on a preset deep learning model training method to obtain the power fitting detection model, including:

[0021] Acquire a number of historical transmission line inspection hardware images, and construct a training set, a validation set, and a test set based on the number of historical transmission line inspection hardware images;

[0022] Training the initial electric fitting detection model based on the training set and the validation set until the initial electric fitting detection model converges to obtain a first electric fitting detection model;

[0023] Testing the first power fitting detection model based on the test set to evaluate the missed detection rate and false detection rate of the first power fitting detection model;

[0024] When the missed detection rate and the false detection rate of the first power fitting detection model meet the preset detection standard, the power fitting detection model is obtained.

[0025] This preferred solution constructs training sets, validation sets, and test sets through historical transmission line inspection hardware images, and evaluates the missed detection rate and false detection rate of the power hardware detection model through model training and model testing, and finally obtains a power hardware detection model that meets the preset detection standards. It can fully tap the potential of the model, enable the power hardware detection model to fully learn the diversity and complexity of power hardware, effectively reduce the missed detection rate and false detection rate in practical applications, and thus improve the accuracy of power hardware detection.

[0026] As a preferred solution, inputting the inspection image of the transmission line to be inspected into the power fitting detection model to obtain the power fitting detection result includes:

[0027] Inputting the inspection image of the transmission line to be detected into the power fitting detection model, so that the feature extraction network extracts features from the inspection image of the transmission line to be detected to obtain several layers of initial power fitting feature maps;

[0028] Inputting the initial power fitting feature graphs of the plurality of layers into the cross-layer connection fusion network, so that the cross-layer connection fusion network performs feature fusion on the initial power fitting feature graphs of the plurality of layers to obtain the first power fitting feature graphs of the plurality of layers;

[0029] Inputting the plurality of layers of first electric hardware feature maps into the plurality of decoupling adaptive feature fusion structures, so that the decoupling adaptive feature fusion structures perform feature decoupling and feature fusion on the plurality of layers of first electric hardware feature maps, to obtain a plurality of layers of second electric hardware feature maps;

[0030] The characteristic graphs of the plurality of layers of second electric fittings are input into the detection module, so that the detection module identifies the characteristic graphs of the plurality of layers of second electric fittings and obtains the detection result of the electric fittings.

[0031] This preferred solution uses a feature extraction network to extract features from the inspection image of the transmission line to be inspected, and then uses a cross-layer connection fusion network to fuse the initial power fitting feature map, integrating features at different levels and enhancing the feature expression capability. The decoupling adaptive feature fusion structure is then used to decouple and fuse the first power fitting feature map, further improving the distinguishability and representativeness of the power fitting features. The detection module then identifies the second power fitting feature map and outputs the power fitting detection results. This fully utilizes the decoupling adaptive feature fusion network to enable the power fitting detection model to accurately detect the power fittings, thereby improving the accuracy of power fitting detection.

[0032] As a preferred solution, the inputting of the several layers of initial electric hardware feature maps into the cross-layer connection fusion network so that the cross-layer connection fusion network performs feature fusion on the several layers of initial electric hardware feature maps to obtain several layers of first electric hardware feature maps, including:

[0033] Inputting the several layers of initial electric hardware feature maps into the cross-layer connection fusion network, performing a nearest neighbor interpolation upsampling operation and a convolution operation on each layer of the initial electric hardware feature map based on the top-down path, to obtain several layers of third electric hardware feature maps;

[0034] Based on the bottom-up path, a nearest neighbor interpolation downsampling operation and a convolution operation are performed on the plurality of layers of third electric fitting feature maps to obtain a plurality of layers of first electric fitting feature maps.

[0035] This preferred solution obtains the first electric fitting feature map through the nearest neighbor interpolation upsampling operation and convolution operation of the top-down path, and the nearest neighbor interpolation downsampling operation and convolution operation of the bottom-up path. It not only realizes the multi-scale processing of the initial electric fitting feature map and can capture semantic information and detail features at different levels, but also retains the integrity of the features through the nearest neighbor interpolation upsampling operation and the nearest neighbor interpolation upsampling operation, reduces the loss of feature information, and extracts the key information in the features through the convolution operation, so that the cross-layer connection fusion network can output a more representative and discriminative feature map, providing a more accurate data basis for the subsequent processing of the decoupled adaptive feature fusion structure, further improving the performance of the electric fitting detection model, and improving the accuracy of electric fitting detection.

[0036] As a preferred solution, the inputting of the several layers of first electric hardware feature maps into the several decoupling adaptive feature fusion structures so that the decoupling adaptive feature fusion structures perform feature decoupling and feature fusion on the several layers of first electric hardware feature maps to obtain several layers of second electric hardware feature maps, including:

[0037] Inputting each three layers of the first power fitting feature graph into the decoupled adaptive feature fusion structure in sequence;

[0038] Based on the decoupled adaptive feature fusion structure, a sampling operation is performed on every three layers of the first power fitting feature map to obtain a plurality of semantic context encoding feature maps;

[0039] Based on the decoupled adaptive feature fusion structure, a sampling operation is performed on every three layers of the first power fitting feature map to obtain a plurality of detail coding feature maps;

[0040] Based on the decoupled adaptive feature fusion structure, the semantic context encoding feature map and the detail encoding feature map are weightedly fused to obtain several layers of second power fitting feature maps.

[0041] This preferred solution obtains a semantic context coding feature map and a detail coding feature map through sampling operations, realizes feature decoupling of the first electric hardware feature map, and then performs weighted fusion to realize feature fusion; the overall semantic information and contextual relationship of the electric hardware are captured through the semantic context coding feature map, and the local detail features of the electric hardware are captured through the detail coding feature map; through weighted fusion, the advantages of the two features are combined, so that the final second electric hardware feature map contains rich semantic information and retains key detail features, thereby further improving the feature expression ability of the electric hardware and improving the accuracy of electric hardware detection.

[0042] Accordingly, an embodiment of the present invention provides an electric fitting detection system, comprising: a data acquisition module, a decoupling adaptive feature fusion network construction module, an electric fitting detection model construction module, and an electric fitting detection module;

[0043] The data acquisition module is used to obtain inspection images and deep learning models of the power transmission line to be inspected;

[0044] The decoupling adaptive feature fusion network construction module is used to construct a cross-layer connection fusion network and a plurality of decoupling adaptive feature fusion structures, and determine a decoupling adaptive feature fusion network based on the cross-layer connection fusion network and the plurality of decoupling adaptive feature fusion structures;

[0045] The power fitting detection model building module is used to update the deep learning model based on the decoupled adaptive feature fusion network to obtain a power fitting detection model;

[0046] The power fitting detection module is used to input the inspection image of the transmission line to be detected into the power fitting detection model to obtain the power fitting detection result.

[0047] Accordingly, an embodiment of the present invention provides a terminal device, including:

[0048] one or more processors;

[0049] a memory, coupled to the processor, for storing one or more programs;

[0050] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method for detecting electrical fittings.

[0051] Accordingly, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the above-mentioned method for detecting electrical fittings. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flowchart of a method for detecting electrical fittings provided by an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of an electric fitting detection model provided by an embodiment of the present invention;

[0054] Figure 3 This is a diagram showing the model impact of different adaptive parameters provided by an embodiment of the present invention;

[0055] Figure 4 A structural diagram of a DAFF provided in an embodiment of the present invention;

[0056] Figure 5 This is a structural diagram of an electric fitting detection system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0057] 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 the embodiments. 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.

[0058] Example 1

[0059] In order to solve the problem of low accuracy of power detection in the existing technology, please refer to Figure 1 , Figure 1 A flowchart of a method for detecting electrical fittings provided by an embodiment of the present invention includes steps S101 to S104.

[0060] Step S101: Obtain inspection images and deep learning models of the power transmission lines to be inspected.

[0061] In an optional embodiment, a staff member operates a drone or the drone autonomously cruises the transmission line, and the drone's camera captures an inspection image of the transmission line to be inspected; the deep learning model described in this optional embodiment can be a deep learning model such as a YOLOv5 model or a Faster RCNN model.

[0062] It should be noted that the YOLOv5 model (You Only Look Once version 5) is a single-stage object detection model. Its core concept is to treat the object detection task as an end-to-end regression problem. It directly predicts the bounding box position and category probability of the target through a single forward propagation, without generating candidate regions. It includes: a backbone network (Backbone), a neck network (Neck), and a detection head (Head). Faster RCNN (Faster Region-based Convolutional Network) is a deep learning model for object detection. By introducing the Region Proposal Network (RPN), it unifies candidate box generation and object detection into an end-to-end framework. Its core process includes: feature extraction, candidate box generation (RPN), and object detection.

[0063] Step S102: constructing a cross-layer connection fusion network and a plurality of decoupled adaptive feature fusion structures, and determining a decoupled adaptive feature fusion network based on the cross-layer connection fusion network and the plurality of decoupled adaptive feature fusion structures.

[0064] In this embodiment, the deep learning model includes: a feature extraction network; the construction of a cross-layer connection fusion network and a plurality of decoupled adaptive feature fusion structures, and determining a decoupled adaptive feature fusion network based on the cross-layer connection fusion network and the plurality of decoupled adaptive feature fusion structures, includes:

[0065] Constructing a top-down path and a bottom-up path according to the number of feature map outputs of the feature extraction network, and constructing a cross-layer connection fusion network based on the top-down path and the bottom-up path;

[0066] Constructing a plurality of decoupling adaptive feature fusion structures according to the number of feature graphs output by the feature extraction network, each of the decoupling adaptive feature fusion structures being used to perform feature decoupling and feature fusion on the first power fitting feature graph output by the cross-layer connection fusion network;

[0067] A decoupled adaptive feature fusion network is constructed based on the cross-layer connection fusion network and several decoupled adaptive feature fusion structures.

[0068] This embodiment constructs a cross-layer connection fusion network of top-down paths and bottom-up paths, and constructs several decoupled adaptive feature fusion structures for feature decoupling and feature fusion. Through the top-down and bottom-up path construction of the cross-layer connection fusion network, it is possible to fully integrate feature information at different levels and enhance the expressiveness of features. In addition, the decoupled adaptive feature fusion structure can be used to specifically decouple and analyze the features of electrical fittings and perform fusion processing, thereby better extracting the features of electrical fittings, improving the performance of subsequent electrical fitting detection models, and thereby improving the accuracy of electrical fitting detection.

[0069] In an optional embodiment, a cross-layer connection fusion network (Bi-Path Cross-Layer Integration Network, BCL-Net) consists of a top-down path and a bottom-up path, which is used to fuse feature maps of different resolution sizes; the cross-layer connection fusion network can effectively retain the feature information of small-target electrical fittings while effectively fusing multi-scale feature information; the decoupled adaptive feature fusion structure (Dynamic Adaptive Feature Fusion, DAFF) is used to perform feature decoupling and feature fusion on the first electrical fitting feature map output by the cross-layer connection fusion network, which can focus on two highly related but contradictory tasks in target detection: classification and positioning, and give different weights to their importance, so that DAFF can effectively distinguish densely connected similar electrical fittings; a decoupled adaptive feature fusion network (Decoupled Adaptive Feature Fusion Network, DAFFN) is constructed based on the cross-layer connection fusion network and several decoupled adaptive feature fusion structures.

[0070] Step S103: updating the deep learning model based on the decoupled adaptive feature fusion network to obtain a power fitting detection model.

[0071] In this embodiment, the deep learning model further includes: an FPN structure and a detection module; the updating of the deep learning model based on the decoupled adaptive feature fusion network to obtain a power fitting detection model includes:

[0072] The decoupled adaptive feature fusion network replaces the FPN structure, and an initial power fitting detection model is constructed based on the feature extraction network, the cross-layer connection fusion network, and several decoupled adaptive feature fusion structures and detection modules;

[0073] The initial electric hardware detection model is trained based on a preset deep learning model training method to obtain an electric hardware detection model.

[0074] This embodiment replaces the FPN structure with a decoupled adaptive feature fusion network and then performs model training to obtain an electric fitting detection model. This can give full play to the advantages of the decoupled adaptive feature fusion network and improve the electric fitting detection model's ability to process electric fitting features. Training is performed based on a preset deep learning model training method to ensure that the electric fitting detection model can learn effective features and converge, thereby obtaining an electric fitting detection model with better performance. This not only improves the detection accuracy of the electric fitting detection model, but also enhances the adaptability and robustness of the electric fitting detection model, making it more suitable for actual transmission line inspection image detection scenarios and improving the accuracy of electric fitting detection.

[0075] In this embodiment, the initial power fitting detection model is trained based on a preset deep learning model training method to obtain the power fitting detection model, including:

[0076] Acquire a number of historical transmission line inspection hardware images, and construct a training set, a validation set, and a test set based on the number of historical transmission line inspection hardware images;

[0077] Training the initial electric fitting detection model based on the training set and the validation set until the initial electric fitting detection model converges to obtain a first electric fitting detection model;

[0078] Testing the first power fitting detection model based on the test set to evaluate the missed detection rate and false detection rate of the first power fitting detection model;

[0079] When the missed detection rate and the false detection rate of the first power fitting detection model meet the preset detection standard, the power fitting detection model is obtained.

[0080] This embodiment constructs a training set, a validation set, and a test set through historical transmission line inspection fitting images, and evaluates the missed detection rate and false detection rate of the power fitting detection model through model training and model testing. Ultimately, a power fitting detection model that meets the preset detection standards is obtained. This can fully tap the potential of the model, enabling the power fitting detection model to fully learn the diversity and complexity of power fittings, effectively reducing the missed detection rate and false detection rate in practical applications, thereby improving the accuracy of power fitting detection.

[0081] In an alternative embodiment, see Figure 2 , Figure 2A schematic diagram of an electric fitting detection model provided by an embodiment of the present invention, which includes a feature extraction network, a cross-layer connection fusion network (BCL-Net), several decoupled adaptive feature fusion structures (DAFF) and a detection module; this embodiment uses the YOLOv5 model or Faster The FPN structure in the RCNN model is replaced by a decoupled adaptive feature fusion network (DAFFN), that is, it is replaced by a cross-layer connection fusion network (BCL-Net) and several decoupled adaptive feature fusion structures (DAFF), thereby obtaining an initial power fitting detection model; then, several historical transmission line inspection fitting images are obtained, and a training set, a validation set, and a test set are constructed based on several historical transmission line inspection fitting images; then, based on Pytorch or a detection library based on the Pytorch framework (such as MMdetection), the initial power fitting detection model is supervised trained using the training set and the validation set. During the training process, pre-trained parameters can be used to accelerate the training until the initial power fitting detection model converges (that is, the verification loss and accuracy of the initial power fitting detection model no longer change significantly), and the first power fitting detection model is obtained; then, the first power fitting detection model is tested based on the test set, and AP or AP is selected. 50 The performance of the first power fitting detection model is evaluated using general indicators such as [number of indicators missing] and [number of indicators false positive], and when its missed detection rate and false detection rate meet the preset detection standards, the power fitting detection model is obtained;

[0082] Furthermore, when the performance evaluation results of the first power fitting detection model show that the missed detection rate and the false detection rate do not meet the preset detection standards, the initial power fitting detection model is re-trained based on the training set and the validation set, and the adaptive parameters are adjusted during the supervised training process. After that, the first power fitting detection model is re-obtained and re-tested based on the test set. AP or AP 50 The performance of the first power fittings detection model is evaluated using general indicators such as the detection rate, false detection rate, and so on, until its missed detection rate and false detection rate meet the preset detection standards.

[0083] It should be noted that the FPN structure (Feature Pyramid Network) is a deep learning architecture used to solve the multi-scale problem in target detection. Its core idea is to fuse feature maps of different levels; it consists of three parts: Bottom-Up Pathway, Top-Down Pathway and Lateral Connections.

[0084] It should be noted that the Pytorch framework is an open source deep learning framework based on Python. It takes dynamic computation graph as its core feature and combines ease of use and high performance. It has become a widely used tool in scientific research and industry. 50 Average Precision (mAAP) is a commonly used indicator for evaluating the detection performance of a model in target detection. It is a comprehensive reflection of the model's missed detection rate and false detection rate of the target (missed detection rate and false detection rate are often weighed against each other; the larger the value, the less missed detection or false detection the model can achieve, which means that the model can more accurately identify and locate the target it wants to detect. 50 All single-category APs 50 The mean of the total number of categories after summation reflects the overall detection ability of the model.

[0085] It should be noted that in target detection, the indicators for evaluating detection performance are usually IoU (Intersection over Union), Precision, Recall, AP (Average Precision) and mAP (mean Average Precision); among them, IoU (Intersection over Union) refers to the degree of overlap between the predicted box and the true box, which is used to determine whether the detection is correct; Precision and Recall are used to measure the quality of the detection results; AP (Average Precision) measures the performance of the model at different Recall levels by calculating the area under the Precision-Recall curve; mAP (mean Average Precision) is the average of APs of multiple categories; IoU (Intersection over Union) is defined as Where B p is the prediction box, B gt is the real frame; AP 50 The Average Precision is calculated under the premise of IoU>50. The calculation steps are as follows: for each predicted box, determine whether it is True Positive (TP) or False Positive (FP), that is, if the IoU between the predicted box and a real box is ≥0.5 and the real box has not been matched, it is TP; otherwise it is FP (including insufficient IoU or repeated detection of already matched targets); then set TP i is the cumulative number of TPs when the i-th prediction box is TP, FP iis the cumulative number of FPs when the i-th predicted box is FP; at each step i, the precision (used to reflect the false positives) is calculated: Calculate the recall rate (Recall, used to reflect missed detection): Finally, the P obtained in each step i and R i As a point; after getting all (P i ,R i ) point, draw the PR curve (Precision-Recall), and then use the interpolation method or the area method of all points to get the AP. This optional embodiment uses Where n is the total number of prediction boxes; then calculate mAP, N is the total number of detection categories, which is set to 6 in this embodiment. In addition, AP 0.5:0.95 、AP 75 However, due to the low positioning accuracy of the power fittings detection in the actual transmission line, and the higher requirements for the overall detection accuracy, this embodiment uses AP 50 .

[0086] In an optional embodiment, please refer to Table 1, which is a comparative test table of the power fitting detection model and the baseline model provided by the embodiment of the present invention; Figure 3 The model impact effect diagram of different adaptive parameters provided in the embodiment of the present invention; the present invention sets a data set containing six types of electrical hardware, wherein the six types of electrical hardware are U-shaped hanging rings (U), right-angle hanging plates (Z), parallel hanging plates (PD), connecting plates (L), equalizing rings (FJH), and U-shaped hanging plates (UB); then the MMdetection detection framework is used as the basis, and the model formed by the faster-rcnn_r50_fpn.py configuration file is used as the baseline model to train the data set once. The baseline model uses ResNet50 as the backbone network to extract image features, and uses the FPN structure as Neck to realize multi-scale feature fusion. This embodiment writes the implementation file of DAFN, and replaces the FPN structure, adjusts different adaptive parameters, and trains the model multiple times. It is assumed that the adaptive parameters α and β are both 0.5, and the following is obtained: Figure 3 As shown in the comparative test diagram, it can be seen from Table 1 that the power fitting detection model (i.e. the model after adding DAFFN in Table 1) and the baseline model have the following characteristics: 50 , and mAAP 50, the power fittings detection model (i.e., the model after adding DAFFN in Table 1) is greater than the baseline model, which means that the detection accuracy of the power fittings detection model (i.e., the model after adding DAFFN in Table 1) is greater than the baseline model; Figure 3 As shown, the z-axis represents mAAP 50 The higher the value, the higher the accuracy. When the adaptive parameters α and β are both 0.5, the detection accuracy is the highest and the detection effect is the best.

[0087] Table 1

[0088]

[0089] Step S104: inputting the inspection image of the power transmission line to be inspected into the power fitting inspection model to obtain the power fitting inspection result.

[0090] In this embodiment, inputting the inspection image of the transmission line to be inspected into the power fitting detection model to obtain the power fitting detection result includes:

[0091] Inputting the inspection image of the transmission line to be detected into the power fitting detection model, so that the feature extraction network extracts features from the inspection image of the transmission line to be detected to obtain several layers of initial power fitting feature maps;

[0092] Inputting the initial power fitting feature graphs of the plurality of layers into the cross-layer connection fusion network, so that the cross-layer connection fusion network performs feature fusion on the initial power fitting feature graphs of the plurality of layers to obtain the first power fitting feature graphs of the plurality of layers;

[0093] Inputting the plurality of layers of first electric hardware feature maps into the plurality of decoupling adaptive feature fusion structures, so that the decoupling adaptive feature fusion structures perform feature decoupling and feature fusion on the plurality of layers of first electric hardware feature maps, to obtain a plurality of layers of second electric hardware feature maps;

[0094] The characteristic graphs of the plurality of layers of second electric fittings are input into the detection module, so that the detection module identifies the characteristic graphs of the plurality of layers of second electric fittings and obtains the detection result of the electric fittings.

[0095] This embodiment uses a feature extraction network to extract features from the inspection image of the transmission line to be inspected, and then uses a cross-layer connection fusion network to fuse the initial power fitting feature map, integrating features at different levels and enhancing the feature expression capability. Then, a decoupled adaptive feature fusion structure is used to decouple and fuse the first power fitting feature map, further improving the distinguishability and representativeness of the power fitting features. Then, the second power fitting feature map is identified through the detection module, and the power fitting detection result is output. The decoupled adaptive feature fusion network is fully utilized to enable the power fitting detection model to accurately detect the power fittings, thereby improving the accuracy of power fitting detection.

[0096] In this embodiment, the inputting of the several layers of initial electric fitting feature graphs into the cross-layer connection fusion network so that the cross-layer connection fusion network performs feature fusion on the several layers of initial electric fitting feature graphs to obtain several layers of first electric fitting feature graphs includes:

[0097] Inputting the several layers of initial electric hardware feature maps into the cross-layer connection fusion network, performing a nearest neighbor interpolation upsampling operation and a convolution operation on each layer of the initial electric hardware feature map based on the top-down path, to obtain several layers of third electric hardware feature maps;

[0098] Based on the bottom-up path, a nearest neighbor interpolation downsampling operation and a convolution operation are performed on the plurality of layers of third electric fitting feature maps to obtain a plurality of layers of first electric fitting feature maps.

[0099] This embodiment obtains a first electric fitting feature map through the nearest neighbor interpolation upsampling operation and convolution operation of the top-down path, and the nearest neighbor interpolation downsampling operation and convolution operation of the bottom-up path. It not only realizes multi-scale processing of the initial electric fitting feature map and can capture semantic information and detail features at different levels, but also retains the integrity of the features through the nearest neighbor interpolation upsampling operation and the nearest neighbor interpolation upsampling operation, reduces the loss of feature information, and extracts key information from the features through the convolution operation, so that the cross-layer connection fusion network can output a more representative and discriminative feature map, providing a more accurate data basis for the subsequent processing of the decoupled adaptive feature fusion structure, further improving the performance of the electric fitting detection model, and improving the accuracy of electric fitting detection.

[0100] In this embodiment, the inputting of the plurality of layers of first electric hardware feature maps into the plurality of decoupling adaptive feature fusion structures so that the decoupling adaptive feature fusion structures perform feature decoupling and feature fusion on the plurality of layers of first electric hardware feature maps to obtain plurality of layers of second electric hardware feature maps includes:

[0101] Inputting each three layers of the first power fitting feature graph into the decoupled adaptive feature fusion structure in sequence;

[0102] Based on the decoupled adaptive feature fusion structure, a sampling operation is performed on every three layers of the first power fitting feature map to obtain a plurality of semantic context encoding feature maps;

[0103] Based on the decoupled adaptive feature fusion structure, a sampling operation is performed on every three layers of the first power fitting feature map to obtain a plurality of detail coding feature maps;

[0104] Based on the decoupled adaptive feature fusion structure, the semantic context encoding feature map and the detail encoding feature map are weightedly fused to obtain several layers of second power fitting feature maps.

[0105] This embodiment obtains a semantic context coding feature map and a detail coding feature map through sampling operations, realizes feature decoupling of the first electric hardware feature map, and then performs weighted fusion to realize feature fusion; the overall semantic information and contextual relationship of the electric hardware are captured through the semantic context coding feature map, and the local detail features of the electric hardware are captured through the detail coding feature map; through weighted fusion, the advantages of the two features are combined, so that the final second electric hardware feature map contains rich semantic information and retains key detail features, thereby further improving the feature expression ability of the electric hardware and improving the accuracy of electric hardware detection.

[0106] In an alternative embodiment, see Figure 4 , Figure 4 A structural diagram of a DAFF provided in an embodiment of the present invention is shown as follows: Figure 2 and Figure 4 As shown, firstly, based on the feature extraction network, feature extraction is performed on the inspection image of the transmission line to be detected, and four layers of initial power fitting feature maps with different resolution sizes are obtained (four layers are set in this embodiment, and the specific number of layers can be determined according to actual needs), namely {C2, C3, C4, C5}, C i (i=2,3,4,5) means that only the inspection images of the transmission lines to be detected are Feature map of

[0107] Then the four layers of initial power fitting feature maps with different resolutions are input into the cross-layer connection fusion network (BCL-Net). Then each layer of initial power fitting feature map C n Through the nearest neighbor interpolation upsampling operation and convolution operation, with C n-1 Achieve aggregation and obtain the characteristic graph of the third power fittings on the fourth layer ({C 2m ,C 3m ,C 4m ,C 5m}) to preliminarily fuse the target feature information in different initial electrical fitting feature maps, thereby realizing the flow of semantic information from the low-resolution feature map to the high-resolution feature map. The specific implementation process is shown in the following formula:

[0108]

[0109] Among them, up() represents the nearest neighbor interpolation upsampling operation, Conv n×n represents n×n convolution, + represents element addition operation; C n Represents the initial electrical hardware feature map of the nth layer; Conv 1×1 (C n ) indicates that C n Perform 1×1 convolution operation; Indicates C n The feature map obtained after the nearest neighbor interpolation upsampling operation or convolution operation; C nm Represents the characteristic diagram of the third electrical fitting on the nth layer;

[0110] Then, the multi-scale features are aggregated through a bottom-up path. Specifically, the multi-scale features are downsampled and then fused, and the feature maps with richer semantic information but poorer detail perception are downsampled to generate higher-resolution feature maps. Specifically, the four-layer third power fitting feature maps are subjected to nearest neighbor interpolation downsampling and convolution operations, and then residual connections are performed to obtain the four-layer first power fitting feature maps (i.e., P2, P3, P4, and P5). The specific implementation process is shown in the following formula:

[0111]

[0112] Among them, down() represents the nearest neighbor interpolation downsampling operation; Conv n×n represents n×n convolution, + represents element addition operation; C nm Represents the characteristic graph of the third electrical fittings in the nth layer; Conv 1×1 (C n ) indicates that C n Perform 1×1 convolution operation; represents the feature map obtained by performing the nearest neighbor interpolation downsampling operation or convolution operation on the third power fitting feature map of the nth layer, P n Represents the characteristic diagram of the first electrical fitting on the nth layer;

[0113] However, since the input of the decoupled adaptive feature fusion structure (DAFF) requires three layers of adjacent feature maps (because the features of the current level are highly correlated with the features of the two adjacent layers), and outputs a single integrated feature map, that is, three layers of adjacent first power fitting feature maps are required; in order to ensure that the decoupled adaptive feature fusion structure can output four layers of second power fitting feature maps, it is necessary to perform two nearest neighbor interpolation downsampling operations on the four layers of first power fitting feature maps, that is, obtain P6 and P7 through P5; among them, P n =down(P n-1 ),n=6,7, a total of six layers of first electrical hardware characteristic diagrams.

[0114] In an optional embodiment, the design concept of the decoupled adaptive feature fusion structure (DAFF) is derived from the decoupling head; the decoupling head refers to the use of different decoupled feature maps in the detection head to perform classification and regression operations respectively. However, such a design will increase the complexity of the model. Therefore, this embodiment implements the decoupling of feature maps in the feature fusion network and realizes fusion in a weighted manner. Specifically, this embodiment implements the semantic context encoding feature map G for classification. cls and the detail encoding feature map G for positioning loc The adaptive fusion of the two decoupled feature maps is achieved through the weighted coefficients α and β, which improves the model's ability to detect similar electrical fittings. Therefore, the overall formula of DAFF can be expressed as:

[0115]

[0116] Further, if Figure 5 As shown, this embodiment uses P n 、P n+1 and P n+2 For example, the first power fitting feature map of each three layers is input into a decoupled adaptive feature fusion structure, and the P n+2 Perform the nearest neighbor interpolation upsampling twice, and n Perform the nearest neighbor interpolation downsampling twice, and then perform the residual connection to obtain the semantic context encoding feature map G cls ; Its specific expression is: G cls =P n+1 +up(P n+2 )+down(P n+1 ), n=2,3,4,5; and obtain the detail coding feature map G loc The design idea is to transform the shallow P n+1 and P n After fusion, it flows back to the next layer of feature map. The specific process is: n+2 Perform the nearest neighbor interpolation upsampling twice, and n+1After performing the nearest neighbor interpolation upsampling twice, the n Perform residual connection, then perform the nearest neighbor interpolation downsampling twice, and finally perform residual connection to obtain the detail encoding feature map G loc , detail encoding feature map G loc The specific expression is: G loc =P n+1 +up(P n+2 )+down(up(P n+1 )+P n ), n = 2, 3, 4, 5; then, based on the adaptive parameters (α, β, in this embodiment, α = 0.5, β = 0.5), the semantic context coding feature map and the detail coding feature map are weightedly fused to obtain the four-layer second power fitting feature map. The expression of the second power fitting feature map is specifically: Where α+β=1;

[0117] Then, the characteristic graph of the second electric fittings on the fourth layer is input into the detection module to obtain the detection result of the electric fittings.

[0118] This embodiment constructs a cross-layer connection fusion network and several decoupled adaptive feature fusion structures, and constructs a decoupled adaptive feature fusion network based on the cross-layer connection fusion network and the several decoupled adaptive feature fusion structures, thereby updating the deep learning model to obtain an electric fitting detection model, thereby realizing the detection of electric fittings based on the electric fitting detection model for the inspection image of the transmission line to be inspected. The present invention enhances the recognition ability of the electric fitting detection model for subtle features and complex scenes of electric fittings by constructing a decoupled adaptive feature fusion network, integrating cross-layer connection and adaptive feature weighting mechanism, thereby making up for the low accuracy of traditional deep learning models in electric fitting detection, significantly improving the detection accuracy of electric fittings under complex backgrounds, effectively avoiding the problems of missed detection and false detection of electric fittings, and effectively improving the detection accuracy of similar small-target electric fittings in the inspection image of the transmission line.

[0119] Example 2

[0120] Please refer to Figure 5 , Figure 5 A schematic structural diagram of an electric fitting detection system provided in an embodiment of the present invention includes: a data acquisition module 201 , a decoupling adaptive feature fusion network construction module 202 , an electric fitting detection model construction module 203 and an electric fitting detection module 204 .

[0121] The data acquisition module 201 is used to obtain inspection images and deep learning models of the power transmission lines to be inspected.

[0122] The decoupled adaptive feature fusion network construction module 202 is used to construct a cross-layer connection fusion network and several decoupled adaptive feature fusion structures, and determine a decoupled adaptive feature fusion network based on the cross-layer connection fusion network and the several decoupled adaptive feature fusion structures.

[0123] In this embodiment, the decoupling adaptive feature fusion network construction module 202 includes: a decoupling adaptive feature fusion network construction unit;

[0124] In the decoupled adaptive feature fusion network construction unit, the deep learning model includes: a feature extraction network;

[0125] The decoupled adaptive feature fusion network construction unit is used to construct a top-down path and a bottom-up path according to the number of feature map outputs of the feature extraction network, and to construct a cross-layer connection fusion network based on the top-down path and the bottom-up path;

[0126] Constructing a plurality of decoupling adaptive feature fusion structures according to the number of feature graphs output by the feature extraction network, each of the decoupling adaptive feature fusion structures being used to perform feature decoupling and feature fusion on the first power fitting feature graph output by the cross-layer connection fusion network;

[0127] A decoupled adaptive feature fusion network is constructed based on the cross-layer connection fusion network and several decoupled adaptive feature fusion structures.

[0128] The power fitting detection model construction module 203 is used to update the deep learning model based on the decoupled adaptive feature fusion network to obtain a power fitting detection model.

[0129] In this embodiment, the electric fitting detection model construction module 203 includes: an electric fitting detection model construction unit;

[0130] In the power fitting detection model construction unit, the deep learning model further includes: an FPN structure and a detection module;

[0131] The power fitting detection model construction unit is used to replace the FPN structure with the decoupled adaptive feature fusion network, and to construct an initial power fitting detection model based on the feature extraction network, the cross-layer connection fusion network, and the plurality of decoupled adaptive feature fusion structures and detection modules;

[0132] The initial electric hardware detection model is trained based on a preset deep learning model training method to obtain an electric hardware detection model.

[0133] In this embodiment, the electric fitting detection model building unit includes: a model training subunit;

[0134] The model training subunit is used to obtain a number of historical transmission line inspection hardware images, and construct a training set, a verification set and a test set based on the number of historical transmission line inspection hardware images;

[0135] Training the initial electric fitting detection model based on the training set and the validation set until the initial electric fitting detection model converges to obtain a first electric fitting detection model;

[0136] Testing the first power fitting detection model based on the test set to evaluate the missed detection rate and false detection rate of the first power fitting detection model;

[0137] When the missed detection rate and the false detection rate of the first power fitting detection model meet the preset detection standard, the power fitting detection model is obtained.

[0138] The power fitting detection module 204 is configured to input the inspection image of the power transmission line to be detected into the power fitting detection model to obtain the power fitting detection result.

[0139] In this embodiment, the power fitting detection module 204 includes: a power fitting detection unit;

[0140] The power fitting detection unit is used to input the inspection image of the transmission line to be detected into the power fitting detection model, so that the feature extraction network extracts features from the inspection image of the transmission line to be detected to obtain several layers of initial power fitting feature maps;

[0141] Inputting the initial power fitting feature graphs of the plurality of layers into the cross-layer connection fusion network, so that the cross-layer connection fusion network performs feature fusion on the initial power fitting feature graphs of the plurality of layers to obtain the first power fitting feature graphs of the plurality of layers;

[0142] Inputting the plurality of layers of first electric hardware feature maps into the plurality of decoupling adaptive feature fusion structures, so that the decoupling adaptive feature fusion structures perform feature decoupling and feature fusion on the plurality of layers of first electric hardware feature maps, to obtain a plurality of layers of second electric hardware feature maps;

[0143] The characteristic graphs of the plurality of layers of second electric fittings are input into the detection module, so that the detection module identifies the characteristic graphs of the plurality of layers of second electric fittings and obtains the detection result of the electric fittings.

[0144] In this embodiment, the power fitting detection unit includes: a cross-layer connection fusion network processing subunit;

[0145] The cross-layer connection fusion network processing subunit is used to input the several layers of initial power fitting feature maps into the cross-layer connection fusion network, and perform a nearest neighbor interpolation upsampling operation and a convolution operation on each layer of the initial power fitting feature map based on the top-down path to obtain several layers of third power fitting feature maps;

[0146] Based on the bottom-up path, a nearest neighbor interpolation downsampling operation and a convolution operation are performed on the plurality of layers of third electric fitting feature maps to obtain a plurality of layers of first electric fitting feature maps.

[0147] In this embodiment, the power fitting detection unit includes: a decoupling adaptive feature fusion structure processing subunit;

[0148] The decoupling adaptive feature fusion structure processing subunit is used to sequentially input every three layers of the first power fitting feature graphs into a decoupling adaptive feature fusion structure;

[0149] Based on the decoupled adaptive feature fusion structure, a sampling operation is performed on every three layers of the first power fitting feature map to obtain a plurality of semantic context encoding feature maps;

[0150] Based on the decoupled adaptive feature fusion structure, a sampling operation is performed on every three layers of the first power fitting feature map to obtain a plurality of detail coding feature maps;

[0151] Based on the decoupled adaptive feature fusion structure, the semantic context encoding feature map and the detail encoding feature map are weightedly fused to obtain several layers of second power fitting feature maps.

[0152] This embodiment constructs a cross-layer connection fusion network and several decoupled adaptive feature fusion structures, and constructs a decoupled adaptive feature fusion network based on the cross-layer connection fusion network and the several decoupled adaptive feature fusion structures, thereby updating the deep learning model to obtain an electric fitting detection model, thereby realizing the detection of electric fittings based on the electric fitting detection model for the inspection image of the transmission line to be inspected. The present invention enhances the recognition ability of the electric fitting detection model for subtle features and complex scenes of electric fittings by constructing a decoupled adaptive feature fusion network, integrating cross-layer connection and adaptive feature weighting mechanism, thereby making up for the low accuracy of traditional deep learning models in electric fitting detection, significantly improving the detection accuracy of electric fittings under complex backgrounds, effectively avoiding the problems of missed detection and false detection of electric fittings, and effectively improving the detection accuracy of similar small-target electric fittings in the inspection image of the transmission line.

[0153] Example 3

[0154] Based on the above-mentioned embodiment of a method for detecting electrical fittings, embodiment three of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, an embodiment of a method for detecting electrical fittings of the present invention is implemented.

[0155] For example, in this embodiment, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to implement the present invention. The one or more module elements may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the terminal device.

[0156] The terminal device may be a computing device such as a desktop computer, a notebook computer, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0157] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, connecting various parts of the entire terminal device using various interfaces and lines.

[0158] Based on the above method embodiments, an embodiment of the present invention provides a computer-readable storage medium, including a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute an electrical hardware detection method described in any one of the above method embodiments of the present invention.

[0159] Wherein, the module / unit integrated in the device / terminal equipment, if implemented in the form of a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by the processor, it can implement the steps of the above-mentioned various method embodiments. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0160] In summary, the embodiment of the present invention constructs a cross-layer connection fusion network and several decoupled adaptive feature fusion structures, and constructs a decoupled adaptive feature fusion network based on the cross-layer connection fusion network and the several decoupled adaptive feature fusion structures, thereby updating the deep learning model to obtain an electric fitting detection model, thereby realizing the detection of electric fittings based on the electric fitting detection model for the inspection image of the transmission line to be inspected. The present invention enhances the recognition ability of the electric fitting detection model for subtle features and complex scenes of electric fittings by constructing a decoupled adaptive feature fusion network, integrating cross-layer connection and adaptive feature weighting mechanism, thereby making up for the low accuracy of traditional deep learning models in the detection of electric fittings, significantly improving the detection accuracy of electric fittings under complex backgrounds, effectively avoiding the problems of missed detection and false detection of electric fittings, and effectively improving the detection accuracy of similar small-target electric fittings in the inspection image of the transmission line.

[0161] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for detecting electric hardware, characterized in that: include: Obtain inspection images and deep learning models of the transmission lines to be inspected; Constructing a cross-layer connection fusion network and several decoupling adaptive feature fusion structures, and constructing a decoupling adaptive feature fusion network based on the cross-layer connection fusion network and the several decoupling adaptive feature fusion structures; The deep learning model is updated based on the decoupled adaptive feature fusion network to obtain a power fitting detection model; The inspection image of the transmission line to be inspected is input into the power fitting inspection model to obtain the power fitting inspection result.

2. The method for detecting electric power fittings according to claim 1, wherein: The deep learning model includes: a feature extraction network; The method of constructing a cross-layer connection fusion network and a plurality of decoupled adaptive feature fusion structures, and constructing a decoupled adaptive feature fusion network based on the cross-layer connection fusion network and the plurality of decoupled adaptive feature fusion structures, includes: Constructing a top-down path and a bottom-up path according to the number of feature map outputs of the feature extraction network, and constructing a cross-layer connection fusion network based on the top-down path and the bottom-up path; Constructing a plurality of decoupling adaptive feature fusion structures according to the number of feature graphs output by the feature extraction network, each of the decoupling adaptive feature fusion structures being used to perform feature decoupling and feature fusion on the first power fitting feature graph output by the cross-layer connection fusion network; A decoupled adaptive feature fusion network is constructed based on the cross-layer connection fusion network and several decoupled adaptive feature fusion structures.

3. The method for detecting electric power fittings according to claim 2, wherein: The deep learning model further includes: an FPN structure and a detection module; the updating of the deep learning model based on the decoupled adaptive feature fusion network to obtain a power fitting detection model includes: The decoupled adaptive feature fusion network replaces the FPN structure, and an initial power fitting detection model is constructed based on the feature extraction network, the cross-layer connection fusion network, and several decoupled adaptive feature fusion structures and detection modules; The initial electric hardware detection model is trained based on a preset deep learning model training method to obtain an electric hardware detection model.

4. The method for detecting electric power fittings according to claim 3, wherein: The initial electric fitting detection model is trained based on a preset deep learning model training method to obtain the electric fitting detection model, including: Acquire a number of historical transmission line inspection hardware images, and construct a training set, a validation set, and a test set based on the number of historical transmission line inspection hardware images; Training the initial electric fitting detection model based on the training set and the validation set until the initial electric fitting detection model converges to obtain a first electric fitting detection model; Testing the first power fitting detection model based on the test set to evaluate the missed detection rate and false detection rate of the first power fitting detection model; When the missed detection rate and the false detection rate of the first power fitting detection model meet the preset detection standard, the power fitting detection model is obtained.

5. The method for detecting electric power fittings according to claim 3, wherein: The step of inputting the inspection image of the transmission line to be inspected into the power fitting inspection model to obtain the power fitting inspection result includes: Inputting the inspection image of the transmission line to be detected into the power fitting detection model, so that the feature extraction network extracts features from the inspection image of the transmission line to be detected to obtain several layers of initial power fitting feature maps; Inputting the initial power fitting feature graphs of the plurality of layers into the cross-layer connection fusion network, so that the cross-layer connection fusion network performs feature fusion on the initial power fitting feature graphs of the plurality of layers to obtain the first power fitting feature graphs of the plurality of layers; Inputting the plurality of layers of first electric hardware feature maps into the plurality of decoupling adaptive feature fusion structures, so that the decoupling adaptive feature fusion structures perform feature decoupling and feature fusion on the plurality of layers of first electric hardware feature maps, to obtain a plurality of layers of second electric hardware feature maps; The characteristic graphs of the plurality of layers of second electric fittings are input into the detection module, so that the detection module identifies the characteristic graphs of the plurality of layers of second electric fittings and obtains the detection result of the electric fittings.

6. The method for detecting electric power fittings according to claim 5, wherein: The step of inputting the initial power fitting feature graphs of the plurality of layers into the cross-layer connection fusion network so that the cross-layer connection fusion network performs feature fusion on the initial power fitting feature graphs of the plurality of layers to obtain the first power fitting feature graphs of the plurality of layers includes: Inputting the several layers of initial electric hardware feature maps into the cross-layer connection fusion network, performing a nearest neighbor interpolation upsampling operation and a convolution operation on each layer of the initial electric hardware feature map based on the top-down path, to obtain several layers of third electric hardware feature maps; Based on the bottom-up path, a nearest neighbor interpolation downsampling operation and a convolution operation are performed on the plurality of layers of third electric fitting feature maps to obtain a plurality of layers of first electric fitting feature maps.

7. The method for detecting electric power fittings according to claim 6, wherein: The step of inputting the plurality of layers of first electric hardware feature maps into the plurality of decoupling adaptive feature fusion structures so that the decoupling adaptive feature fusion structures perform feature decoupling and feature fusion on the plurality of layers of first electric hardware feature maps to obtain plurality of layers of second electric hardware feature maps includes: Inputting each three layers of the first power fitting feature graph into the decoupled adaptive feature fusion structure in sequence; Based on the decoupled adaptive feature fusion structure, a sampling operation is performed on every three layers of the first power fitting feature map to obtain a plurality of semantic context encoding feature maps; Based on the decoupled adaptive feature fusion structure, a sampling operation is performed on every three layers of the first power fitting feature map to obtain a plurality of detail coding feature maps; Based on the decoupled adaptive feature fusion structure, the semantic context encoding feature map and the detail encoding feature map are weightedly fused to obtain several layers of second power fitting feature maps.

8. An electric power fitting detection system, characterized in that: include: Data acquisition module, decoupling adaptive feature fusion network construction module, power fitting detection model construction module and power fitting detection module; The data acquisition module is used to obtain inspection images and deep learning models of the power transmission line to be inspected; The decoupling adaptive feature fusion network construction module is used to construct a cross-layer connection fusion network and a plurality of decoupling adaptive feature fusion structures, and determine a decoupling adaptive feature fusion network based on the cross-layer connection fusion network and the plurality of decoupling adaptive feature fusion structures; The power fitting detection model building module is used to update the deep learning model based on the decoupled adaptive feature fusion network to obtain a power fitting detection model; The power fitting detection module is used to input the inspection image of the transmission line to be detected into the power fitting detection model to obtain the power fitting detection result.

9. A terminal device, characterized in that: include: one or more processors; a memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the power fitting detection method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is executed by a processor to implement the power fitting detection method according to any one of claims 1 to 7.

Images