Insulator detection method based on hierarchical rotation Gaussian attention network

Abstract

The invention discloses an insulator detection method based on a hierarchical rotation Gaussian attention network, and belongs to the technical field of image processing, and the method comprises the steps: decomposing an insulator aerial image into hierarchical feature maps of different scales through a backbone network CSPNet; noise and insulators in the hierarchical feature map are distinguished by rotating RGCA (Gaussian Convolution Attention), and an output feature map for insulator distinguishing is generated; and inputting the output feature map for insulator distinguishing into a multi-pole rotation Gaussian attention MRGA, expanding target information from hierarchical feature maps of different scales through a multi-pole rotation Gaussian attention mechanism of the multi-pole rotation Gaussian attention MRGA, and positioning insulators in different scales. The insulator detection method based on the hierarchical rotation Gaussian attention network solves the problem that an existing insulator detection method is poor in detection effect.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to an insulator detection method based on a hierarchical rotation Gaussian attention network. Background Art

[0002] As one of the key components in power transmission lines, insulators play the role of electrical insulation and mechanical fixation in power systems. However, external environments (such as extreme weather) often cause insulator damage, posing huge safety hazards to power lines. In addition, the bursting of insulators may paralyze power lines in the entire region, causing significant economic losses. Therefore, insulator inspection is crucial for the normal operation of the power system.

[0003] At present, insulator detection is mainly carried out by taking insulator images by unmanned aerial vehicles (UAVs) and then performing image analysis. Based on this, many insulator detection methods have been proposed to monitor power systems, most of which can be divided into traditional image processing methods and deep learning-based methods. Traditional insulator detection methods are usually completed by designing different target features. These methods rely on the subjective experience of operators and are therefore prone to low accuracy and false detection. For example: a detection method that eliminates background noise and locates insulators by extracting local features and spatial order, but these features are easily affected by external interference, resulting in missed detection; an insulator string detection method based on appearance detection, which uses binary feature pooling to generate target features and then uses support vector machines (SVM) to locate insulators, but it is difficult to accurately obtain insulators due to low resolution and complex background. These methods have low robustness in insulator detection due to low signal-to-noise ratio and complex background. Since deep learning-based methods perform well in target detection, they are also applied to insulator detection. For example: a cascaded convolutional neural network detects insulators in complex backgrounds by combining insulator localization and defect detection, but this method is insensitive to small insulators; a densely connected feature pyramid network (FPN) [, used to perform fault detection of small insulators, using deep semantic information and shallow location information to enhance the location information of small insulators. Although these features can be effectively fused to improve fault detection performance, there are still missed detections due to the occlusion of complex environments. In addition, a rotation region network locates insulators in complex scenes through five prior features, namely rectangular box representation, feature extraction, candidate box generation, anchor design, and feature alignment, and extracts feature maps using ResNet and FPN. Although FPN retains features of different scales, it always ignores the details between hierarchical convolutional layers. An improved Faster RCNN is used to detect insulators of different aspect ratios, but it is difficult to distinguish multi-scale insulators in complex environments. In addition, a Faster R-Transformer uses a self-attention mechanism to locate insulators from complex aerial environments. However, this method requires calculating attention scores for all channels and spatial feature maps, which often leads to redundant information.

[0004] In summary, although these convolutional neural network (CNN) methods show good results, they still face the following problems in practical applications: (1) Insulators are often obscured by complex environments, such as Figure 1 and Figure 2 Fog, forests and farmland in the environment may cause insulator inspection failure; (2) Figure 2As shown in the figure, most insulators change with the change of the tower support structure, resulting in incomplete detection of small insulators; (3) Some effective information in the convolutional neural network (CNN) is distorted in different levels of convolution, reducing the robustness of the convolutional neural network; (4) The attention-based insulator detection method needs to calculate the attention weights of all channels and spaces in the feature map, resulting in redundant information. Therefore, the existing insulator detection method still has the problem of poor detection effect. Summary of the invention

[0005] In order to overcome the defects of the prior art, the present invention provides an insulator detection method based on a hierarchical rotation Gaussian attention network to solve the above problems.

[0006] The technical solution adopted by the present invention to solve the technical problem is: an insulator detection method based on a hierarchical rotation Gaussian attention network, comprising the following steps:

[0007] S1: The insulator aerial image is decomposed into hierarchical feature maps of different scales through the backbone network CSPNet;

[0008] S2: Distinguish the noise and insulators in the hierarchical feature map by rotating Gaussian convolutional attention RGCA, and generate an output feature map for insulator differentiation;

[0009] S3: The output feature map of insulator distinction is input into the Multi-pole Rotated Gaussian Attention MRGA. Through the Multi-pole Rotated Gaussian Attention mechanism of the Multi-pole Rotated Gaussian Attention MRGA, the target information is expanded from the hierarchical feature maps of different scales to locate insulators in different scales.

[0010] It is worth noting that the step S1 includes:

[0011] S11: normalize the aerial image of the insulator to [-1, 1] and input it into the backbone network;

[0012] S12: The convolutional layer in the main pole network CSPNet performs feature decomposition on the normalized aerial image of the insulator to obtain hierarchical feature maps of different scales, where each convolutional layer corresponds to a hierarchical feature map of one scale.

[0013] Preferably, in step S2, the rotated Gaussian convolution attention RGCA includes a rotated channel Gaussian convolution attention module RCGCA and a rotated spatial Gaussian convolution attention module RSGCA; effective channel dependency is stimulated by the rotated channel Gaussian convolution attention module RCGCA; and the spatial distribution information in the Gaussian neighborhood around the insulator is learned by the rotated spatial Gaussian convolution attention module RSGCA.

[0014] Specifically, in the step S2, the formula of the rotational spatial Gaussian convolution attention module RSGCA is: RSGCA(x)=GGPA[SRC(Q), SRC(K), SRC(GN(G))]; GGPA represents Gaussian-guided probabilistic attention, SRC is spatial rotational convolution, Q, K and G are respectively the query, key and Gaussian vector obtained from the convolutional layer Conv2d layer in the main pole network CSPNet, and GN(G) is the normalized Gaussian vector.

[0015] Optionally, in step S2, the query Q is rotated by four set angles in the spatial dimension using spatial rotation convolution SRC to obtain four feature maps, and then the four feature maps corresponding to the query Q are concatenated and input into the Conv2d layer to obtain a multi-view feature map corresponding to the query Q.

[0016] It is worth noting that in step S2, the key K is rotated by four set angles in the spatial dimension using spatial rotation convolution SRC to obtain four feature maps, and then the four feature maps corresponding to the key K are spliced ​​and input into the Conv2d layer to obtain the multi-view feature map corresponding to the key K.

[0017] Specifically, in step S2, the normalized Gaussian vector G is rotated by four set angles in the spatial dimension using spatial rotation convolution SRC to obtain four feature maps, and then the four feature maps corresponding to the normalized Gaussian vector G are concatenated and input into the Conv2d layer to obtain a multi-view feature map corresponding to the normalized Gaussian vector G.

[0018] Preferably, in step S2, the multi-view feature map corresponding to the query Q, the multi-view feature map corresponding to the key K and the multi-view feature map corresponding to the normalized Gaussian vector G are used as the output of the rotational space Gaussian convolution attention module RSGCA; wherein the set angles are 0°, 90°, 180° and 270° respectively.

[0019] Optionally, in step S2, the formula of the rotation channel Gaussian convolution attention module RCGCA is: RCGCA(x)=GGPA[CRC(Q),CRC(K),CRC(GN(G))]; GGPA represents Gaussian-guided probabilistic attention, CRC is channel rotation convolution, CRC(Q) is the value of the multi-view feature map corresponding to the query Q generated by the output of the rotation space Gaussian convolution attention module RSGCA after channel rotation convolution, CRC(K) is the value of the multi-view feature map corresponding to the key K generated by the output of the rotation space Gaussian convolution attention module RSGCA after channel rotation convolution, CRC(GN(G) is the value of the multi-view feature map corresponding to the normalized Gaussian vector G generated by the output of the rotation space Gaussian convolution attention module RSGCA after channel rotation convolution.

[0020] Specifically, in step S3, the formula of the multi-pole rotation Gaussian attention MRGA is: MRGA(X)=GGPA{Norm[MRA(Q,K,V)]};

[0021] GGPA represents Gaussian-guided probabilistic attention; Norm represents normalization operation; MRA represents multipolar rotation attention module; MRA(Q,K,V) represents the output of multi-polar rotation attention; Q is the query feature map, K is the key feature map, and V is the value feature map; represents the cyclic operation of i from 1 to n, concatenating and merging the n results of MR, where each i corresponds to a rotation operation or attention head; MR represents the multi-pole rotation module, which is responsible for rotating the output of the scaled dot product attention SDPA; SDPA represents scaled dot product attention; is the weighted Cha Xun feature map, is the weighted key feature map, is the weighted value feature map; The linear transformation weight matrix of the i-th attention head Used to perform linear transformation on the query feature map Q, the linear transformation weight matrix of the i-th attention head Used to linearly transform the key feature map K, the linear transformation weight matrix of the i-th attention head Used for linear transformation of the value feature map V; d k is the dimension of the key vector, which is used to scale the dot product result; Softmax is a normalization operation that converts the similarity result of the dot product into a probability distribution.

[0022] The beneficial effects of the present invention are as follows: in the insulator detection method based on the hierarchical rotated Gaussian attention network, by setting the rotated Gaussian convolutional attention RGCA, it is possible to stimulate effective channel dependency and learn the multi-perspective spatial distribution information around the Gaussian neighborhood, so as to distinguish insulators from a complex environment; then, by setting the multi-pole rotated Gaussian attention MRGA, by capturing details of different scales and multiple perspectives, multi-scale insulators are located from the hierarchical convolutional layer, thereby improving the performance in the insulator detection process and providing assistance for power grid management. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a schematic diagram of an insulator in dense fog;

[0024] Figure 2 Schematic diagram of the multi-level changes of insulators in aerial images;

[0025] Figure 3It is a flowchart of an insulator detection method based on a hierarchical rotation Gaussian attention network in one embodiment of the present invention;

[0026] Figure 4 A schematic diagram of the structure of a hierarchical rotation Gaussian attention network for insulator detection in one embodiment of the present invention;

[0027] Figure 5 A schematic diagram of generating a hierarchical feature map by a main pole network CSPNet in one embodiment of the present invention;

[0028] Figure 6 A schematic diagram of the structure of the Rotated Gaussian Convolution Attention RGCA in one embodiment of the present invention;

[0029] Figure 7 It is a schematic diagram of the structure of the multi-pole rotation Gaussian attention MRGA in one embodiment of the present invention. DETAILED DESCRIPTION

[0030] The specific embodiments of the present invention are further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0031] like Figure 1-7 As shown, an insulator detection method based on a hierarchical rotation Gaussian attention network includes the following steps:

[0032] S1: Figure 5 As shown in the figure, the aerial image of the insulator is decomposed into hierarchical feature maps of different scales through the backbone network CSPNet;

[0033] The step S1 comprises:

[0034] S11: normalize the aerial image of the insulator to [-1, 1] and input it into the backbone network;

[0035] S12: The convolution layer in the main pole network CSPNet performs feature decomposition on the normalized aerial image of the insulator to obtain hierarchical feature maps of different scales, wherein each convolution layer corresponds to a hierarchical feature map of one scale; in this embodiment, the feature maps of the convolution layer Conv128, the convolution layer Conv256, the convolution layer Conv512, and the convolution layer Conv1024 are used as hierarchical feature maps;

[0036] Feature maps of different scales contain different information. For example, low-scale feature maps contain more noise, while high-scale feature maps retain more target information. Executing step S1 can remove noise and retain target feature details through the backbone network.

[0037] S2: Figure 4 As shown, the noise and insulators in the hierarchical feature map are distinguished by rotating Gaussian convolution attention RGCA, and an output feature map for insulator distinction is generated;

[0038] In step S2, the rotation Gaussian convolution attention RGCA includes a rotation channel Gaussian convolution attention module RCGCA and a rotation space Gaussian convolution attention module RSGCA; Figure 4 As shown, the rotation channel Gaussian convolution attention module RCGCA is used to stimulate effective channel dependency without having to model channel dependency for all feature maps; the rotation spatial Gaussian convolution attention module RSGCA is used to learn the spatial distribution information in the Gaussian neighborhood around the insulator instead of performing a global search in the entire feature map.

[0039] In step S2, the formula of the rotational spatial Gaussian convolutional attention module RSGCA is: RSGCA(x)=GGPA[SRC(Q), SRC(K), SRC(GN(G))]; GGPA represents Gaussian-guided probabilistic attention, SRC is spatial rotational convolution, Q, K and G are respectively the query, key and Gaussian vector obtained from the convolutional layer Conv2d layer in the main pole network CSPNet, and GN(G) is the normalized Gaussian vector. GN(G) is the normalized Gaussian vector in which GN represents the Gaussian vector that satisfies Gaussian normalization operation.

[0040] In the step S2, the query Q, the key K and the Gaussian vector G are obtained from the Conv2d layer; and then the Gaussian vector G is normalized;

[0041] Use spatial rotation convolution SRC to rotate the query Q by four set angles in the spatial dimension to obtain four feature maps, and then concatenate the four feature maps corresponding to the query Q and input them into the Conv2d layer to obtain the multi-view feature map corresponding to the query Q;

[0042] Use spatial rotation convolution SRC to rotate key K by four set angles in the spatial dimension to obtain four feature maps, then concatenate the four feature maps corresponding to key K and input them into the Conv2d layer to obtain the multi-view feature map corresponding to key K;

[0043] Use spatial rotation convolution SRC to rotate the normalized Gaussian vector G by four set angles in the spatial dimension to obtain four feature maps, and then concatenate the four feature maps corresponding to the normalized Gaussian vector G and input them into the Conv2d layer to obtain the multi-view feature map corresponding to the normalized Gaussian vector G;

[0044] The multi-view feature map corresponding to the query Q, the multi-view feature map corresponding to the key K and the multi-view feature map corresponding to the normalized Gaussian vector G are used as the output of the rotational spatial Gaussian convolution attention module RSGCA; the set angles are 0°, 90°, 180° and 270° respectively.

[0045] In this embodiment, the spatial rotation convolution SRC is a single-angle rotation operation. In order to achieve 0°, 90°, 180° and 270° rotations, four spatial rotation convolutions SRC need to be performed respectively for the query Q, key K or the normalized Gaussian vector G to achieve four rotation operations. First, the Gaussian vector G is generated from the Conv2d layer, and the input feature map is used to generate the query Q and key K; then, the Gaussian vector G is normalized, that is, the mean μ is subtracted and divided by the variance σ: Next, the three feature maps (Gaussian vector G, query Q, and key K) are rotated in the channel dimension, each feature map is rotated to four different angles, namely 0°, 90°, 180°, and 270°, and then the four feature maps are concatenated and input into the Conv2d layer to obtain new feature maps Q, K, and G. Finally, Q, K, and G are input into the Gaussian-guided probabilistic attention module GGPA to generate valid features.

[0046] Finally, the query Q, key K and normalized Gaussian vector G are passed to the Gaussian-guided probabilistic attention module GGPA to learn the spatial distribution information of insulators in the feature map. Through the Gaussian-guided probabilistic attention mechanism GGPA, the deterministic distribution between noise and insulators is modeled to obtain the standard Gaussian distribution GGPA(Q,K,G)=Norm(e L(Q ,K,G)-argmax[L(Q,K,G)] )G, the attention score follows the mean μ of 0 and the standard deviation δ of 1; Norm() represents the normalization operation, L(Q,K,G) represents probability optimization, α and β are learnable parameters, c is the number of channels, Q represents the query, K is the key, G is a standard Gaussian vector after Gaussian normalization, satisfying G(x)~N(μ,δ), x is the input feature, μ is the mean of the feature, δ is the standard deviation of the feature, and N() represents Gaussian distribution; in this embodiment, the standard Gaussian vector G(x) satisfies the standard Gaussian distribution with a mean of 0 and a standard deviation of 1, that is, the result of Gaussian normalization, and this normalization is achieved by the following formula: The purpose of this normalization operation is to ensure that the feature vector conforms to the standard Gaussian distribution, thereby providing input for the subsequent probabilistic attention mechanism, which can enhance the model's ability to distinguish between noise and insulator distribution and improve the robustness of the model. If it is a standard normal distribution, the mean μ = 0, the standard deviation δ = 1, which can be expressed as G(x) ~ N(0,1). In this case, the Gaussian distribution formula is

[0047] Based on this, the distribution of the Gaussian-guided probabilistic attention GGPA mechanism can be expressed as:

[0048] q j represents the jth feature unit in the query Q, each q j is a unit in the input feature map that is used in the attention mechanism to interact with the key K and value V; g(j) is an activation unit used to represent the feature unit q of the query Q j Characteristic unit k with key K j The probability of g(j) is to capture the target features through probability modeling while eliminating noise; p j (q j ,k j |x j ) represents the feature unit q of query Q j Characteristic unit k for bond K j Under the condition x j The probability distribution under the condition of ; X represents the input feature matrix or feature map, which is the high-dimensional data output by the convolution layer, which contains the features of multiple samples or channels; x represents a single feature vector or the value of a specific position in X. For example, x can be a channel, a pixel value, or a feature of a local area in the feature map X; j represents a specific feature channel index or the index of a sample. For a three-dimensional feature map X (channel, height, width), j can refer to the index of a channel; N represents the total number of feature channels, that is, the number of channels in X. For example, in a convolutional neural network, the output feature map usually has multiple channels, and N represents the total number of these channels.

[0049] Unlike the traditional attention mechanism, the value V is not calculated directly, but randomly sampled through the Gaussian vector G. By using Gaussian distribution and activation unit g(j), the attention score can better distinguish between noise and target features. This mechanism helps improve the detection ability of the model in complex backgrounds. Among them, when querying the unit feature unit q in the standard Gaussian normalization G(x) j When g(j) is the characteristic unit k of the activation key K j The probability, p j (qj ,k j |x j ) represents the probability of the query pair (q, k). Different from the traditional attention mechanism, the value V is replaced by a standard Gaussian vector G, so that the attention score can be randomly sampled in the Gaussian distribution, thereby increasing the distribution of the attention score ratio between noise and target data. In addition, the introduction of the activation unit g(j) can help the network effectively capture the target information and eliminate noise.

[0050] The output of RSGCA is the multi-view feature map corresponding to the generated query Q, the multi-view feature map corresponding to the key K, and the multi-view feature map corresponding to the normalized Gaussian vector G, which already contain the multi-view spatial features generated by spatial rotation convolution (SRC); next, these multi-view feature maps will be input into CRC for channel rotation convolution operation.

[0051] In the step S2, the formula of the rotation channel Gaussian convolution attention module RCGCA is: RCGCA(x)=GGPA[CRC(Q),CRC(K),CRC(GN(G))]; GGPA represents Gaussian-guided probabilistic attention, CRC is channel rotation convolution, which is a rotation operation on the input feature map in the channel dimension, CRC(Q) is the value of the multi-view feature map corresponding to the query Q generated by the output of the rotation space Gaussian convolution attention module RSGCA after channel rotation convolution, CRC(K) is the value of the multi-view feature map corresponding to the key K generated by the output of the rotation space Gaussian convolution attention module RSGCA after channel rotation convolution, CRC(GN(G) is the multi-view feature map corresponding to the normalized Gaussian vector G generated by the output of the rotation space Gaussian convolution attention module RSGCA after channel rotation convolution. The values ​​after the channel rotation convolution. After these feature maps are rotated in the channel dimension, they are used to capture the dependencies between channels, further improving the model's ability to extract sub-features. Since the rotational spatial Gaussian convolution attention module RSGCA generates multi-view feature maps, these multi-view feature maps are processed separately for each view through the channel rotation convolution CRC of the rotational channel Gaussian convolution attention module RCGCA, thereby capturing the multi-view dependency information between channels. Therefore, since the output of the rotational spatial Gaussian convolution attention module RSGCA contains feature maps of multiple views, each view needs to be processed separately, which means that the channel rotation convolution operation of the rotational channel Gaussian convolution attention module RCGCA needs to be performed multiple times to match the number of views of the multi-view feature map output by the rotational spatial Gaussian convolution attention module RSGCA.

[0052] In summary, introducing the rotated Gaussian convolutional attention RGCA into the network in spatial and channel dimensions can model the distribution between noise and insulators, and further expand the attention score between noise and insulators through Gaussian distribution. Specifically, this module enables the network to adaptively select target regions to detect insulators from complex environments by stimulating effective channel dependencies and learning spatial distribution information in the Gaussian neighborhood around insulators.

[0053] like Figure 6 The rotated Gaussian convolution attention RGCA shown in the figure, where the rotated spatial Gaussian convolution attention module aims to learn the spatial distribution information within the Gaussian neighborhood around the insulator, and the rotated channel Gaussian convolution attention module is developed to stimulate effective channel dependencies between feature maps. Here, Conv2d, Norm, and C represent two-dimensional convolution, normalization, and concatenation, respectively.

[0054] S3: Figure 5 As shown in the figure, the output feature map of insulator differentiation is input into the multi-pole rotation Gaussian attention MRGA. Through the multi-pole rotation Gaussian attention mechanism of the multi-pole rotation Gaussian attention MRGA, the target information is expanded from the hierarchical feature maps of different scales to locate the insulators in different scales. In fact, the multi-pole rotation Gaussian attention mechanism of the multi-pole rotation Gaussian attention MRGA can capture the detailed information of different scales and multiple perspectives from the feature map.

[0055] In order to further enhance the representation ability of insulators, the Gaussian-guided probabilistic attention GGPA is placed after the multi-pole rotated Gaussian attention MRGA to expand the feature information and retain more attention to small insulators by increasing the attention score ratio of the target area; the rotated Gaussian convolutional attention RGCA and the multi-pole rotated Gaussian attention MRGA are executed successively, namely:

[0056] First, the rotated Gaussian convolution attention RGCA is performed. The rotated Gaussian convolution attention RGCA is mainly used to capture the feature dependencies between space and channels, and distinguish insulators in complex environments through the rotated Gaussian convolution mechanism. The output result is a feature map enhanced with spatial and channel features, which has removed some noise and highlighted the significance of insulators. The output of the rotated Gaussian convolution attention RGCA is used as the input of the multi-pole rotated Gaussian attention MRGA, which further processes the output features of the rotated Gaussian convolution attention RGCA, focusing on capturing multi-scale and multi-view features.

[0057] The proposed multi-pole rotated Gaussian attention (MRGA) is used to detect multi-scale insulators from aerial images by capturing more details of different scales and expanding the insulator information from the hierarchical convolutional layers. The multi-pole rotated Gaussian attention (MRGA) uses multi-pole rotated attention to obtain detailed features of different scales and multiple perspectives, and expands the insulator information through Gaussian-guided probabilistic attention (GGPA).

[0058] In step S3, the formula of the multi-polar rotation Gaussian attention MRGA is: MRGA(X)=GGPA{Norm[MRA(Q,K,V)]}; GGPA represents Gaussian-guided probabilistic attention; Norm represents normalization operation; MRA represents multi-polar rotation attention module; MRA(Q,K,V) represents the output of multipolar rotation attention (which is the comprehensive feature fused through the multipolar rotation mechanism); Q is the query feature map, K is the key feature map, and V is the value feature map; It represents the cyclic operation of i from 1 to n, splicing and merging n results of MR, each i corresponds to a rotation operation or attention head, and the splicing result contains the feature representation of all i, which increases the modeling ability of the multi-polar rotation attention module for multi-angle features; MR represents the multi-polar rotation module, which is responsible for rotating the output of the scaled dot product attention SDPA for multi-view modeling, usually including rotating the feature map by 0°, 90°, 180°, and 270° respectively, and further processing these rotated features; SDPA represents the scaled dot product attention, which is the core part of the attention calculation; is the weighted Cha Xun feature map, is the weighted key feature map, is the weighted value feature map; The linear transformation weight matrix of the i-th attention head Used to perform linear transformation on the query feature map Q, the linear transformation weight matrix of the i-th attention head Used to linearly transform the key feature map K, the linear transformation weight matrix of the i-th attention head Used for linear transformation of the value feature map V, Used to map input features to different feature spaces, allowing the attention mechanism to focus on different feature dimensions or angles; d k is the dimension of the key vector, which is used to scale the dot product result to prevent the value from being too large; Softmax is a normalization operation that converts the similarity result of the dot product into a probability distribution to obtain an n×n attention weight matrix. The sum of the weights of each row is 1, which represents the attention distribution for each query vector and key-value pair. Q and K are matched, and the corresponding features are extracted from V. Through the attention mechanism, the most relevant features are focused on, while irrelevant features are suppressed.

[0059] The workflow of Multi-Pole Rotated Gaussian Attention MRGA is as follows:

[0060] Input linear transformation: The input Q, K, and V feature maps are transformed by linear weight matrices Perform transformation to generate different subspace features;

[0061] Scaled dot product attention calculation: for each group of transformed features Calculate dot product attention SDPA;

[0062] Multipolar Rotation Module (MR): Each SDPA result is passed through the multipolar rotation module MR to rotate the feature map at different angles (e.g., 0°, 90°, 180°, 270°) to capture multi-view feature information;

[0063] Concatenate the results: Concatenate the results of each i (corresponding to a rotation or attention head) through Concatenate and integrate all multi-view attention outputs to form the final rotational attention RA(Q,K,V).

[0064] In this embodiment, the query feature graph Q (Query) represents the features that need to be extracted from the key-value pair, the key feature graph K (Key) represents the features used to match the query, and the value feature graph V represents the actual value of the corresponding key, which is used to generate the final output.

[0065] like Figure 7 In the multi-polar rotated Gaussian attention MRGA shown in Figure 1, first, Q(Query), K(Key), and V(Value) are processed by linear and scaled dot product attention, and then these feature maps Q, K, and V are multi-polar rotated to generate multi-view attention weights. Next, these attention weights are concatenated and normalized to obtain Gaussian vectors, which will be fed into the Gaussian guided probabilistic attention (GGPA) block to generate the attention mask. Finally, the input feature map is multiplied by the attention mask to produce the output feature map.

[0066] like Figure 4 As shown, in the insulator detection method based on the hierarchical rotated Gaussian attention network, by setting the rotated Gaussian convolutional attention RGCA, it is possible to stimulate effective channel dependency and learn the multi-view spatial distribution information around the Gaussian neighborhood, so as to distinguish insulators from a complex environment; then, by setting the multi-pole rotated Gaussian attention MRGA, by capturing details of different scales and multiple perspectives, multi-scale insulators are located from the hierarchical convolutional layer, thereby improving the performance in the insulator detection process and providing assistance for power grid management.

[0067] In this scheme, the rotated Gaussian convolutional attention RGCA is used to distinguish insulators from complex environments; the multi-pole rotated Gaussian attention MRGA is used to detect multi-scale insulators from aerial images. The rotated Gaussian convolutional attention RGCA can stimulate effective channel dependencies instead of modeling channel dependencies between all feature maps, while learning multi-view spatial distribution information around the Gaussian neighborhood to distinguish insulators from complex environments. The multi-pole rotated Gaussian attention MRGA can capture details of different scales and multiple perspectives, expand the information of insulators from the hierarchical convolutional layers, and is used to locate multi-scale insulators in aerial images.

[0068] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. An insulator detection method based on hierarchical rotation Gaussian attention network, characterized in that: The following steps are involved: S1: The insulator aerial image is decomposed into hierarchical feature maps of different scales through the backbone network CSPNet; S2: Distinguish the noise and insulators in the hierarchical feature map by rotating Gaussian convolutional attention RGCA, and generate an output feature map for insulator differentiation; S3: The output feature map of insulator distinction is input into the Multi-pole Rotated Gaussian Attention MRGA. Through the Multi-pole Rotated Gaussian Attention mechanism of the Multi-pole Rotated Gaussian Attention MRGA, the target information is expanded from the hierarchical feature maps of different scales to locate insulators in different scales.

2. The insulator detection method based on hierarchical rotation Gaussian attention network according to claim 1 is characterized in that: The step S1 comprises: S11: normalize the aerial image of the insulator to [-1, 1] and input it into the backbone network; S12: The convolutional layer in the main pole network CSPNet performs feature decomposition on the normalized aerial image of the insulator to obtain hierarchical feature maps of different scales, where each convolutional layer corresponds to a hierarchical feature map of one scale.

3. The insulator detection method based on hierarchical rotation Gaussian attention network according to claim 2 is characterized in that: In step S2, the rotated Gaussian convolution attention RGCA includes a rotated channel Gaussian convolution attention module RCGCA and a rotated spatial Gaussian convolution attention module RSGCA; effective channel dependency is stimulated by the rotated channel Gaussian convolution attention module RCGCA; and spatial distribution information in the Gaussian neighborhood around the insulator is learned by the rotated spatial Gaussian convolution attention module RSGCA.

4. The insulator detection method based on hierarchical rotation Gaussian attention network according to claim 3 is characterized in that: In the step S2, the formula of the rotational spatial Gaussian convolution attention module RSGCA is: RSGCA(x)=GGPA[SRC(Q),SRC(K),SRC(GN(G))]; GGPA represents Gaussian-guided probabilistic attention, SRC is spatial rotational convolution, Q, K and G are respectively the query, key and Gaussian vector obtained from the convolutional layer Conv2d layer in the main pole network CSPNet, and GN(G) is the normalized Gaussian vector.

5. The insulator detection method based on hierarchical rotation Gaussian attention network according to claim 4 is characterized in that: In step S2, the query Q is rotated by four set angles in the spatial dimension using spatial rotation convolution SRC to obtain four feature maps, and then the four feature maps corresponding to the query Q are concatenated and input into the Conv2d layer to obtain a multi-view feature map corresponding to the query Q.

6. The insulator detection method based on hierarchical rotation Gaussian attention network according to claim 5 is characterized by: In step S2, spatial rotation convolution SRC is used to rotate key K by four set angles in the spatial dimension to obtain four feature maps, and then the four feature maps corresponding to key K are spliced ​​and input into the Conv2d layer to obtain a multi-view feature map corresponding to key K.

7. The insulator detection method based on hierarchical rotation Gaussian attention network according to claim 6 is characterized in that: In the step S2, the normalized Gaussian vector G is rotated by four set angles in the spatial dimension using the spatial rotation convolution SRC to obtain four feature maps, and then the four feature maps corresponding to the normalized Gaussian vector G are concatenated and input into the Conv2d layer to obtain a multi-view feature map corresponding to the normalized Gaussian vector G.

8. The insulator detection method based on hierarchical rotation Gaussian attention network according to claim 7 is characterized in that: In the step S2, the multi-view feature map corresponding to the query Q, the multi-view feature map corresponding to the key K and the multi-view feature map corresponding to the normalized Gaussian vector G are used as the output of the rotational space Gaussian convolution attention module RSGCA; the set angles are 0°, 90°, 180° and 270° respectively.

9. The insulator detection method based on hierarchical rotation Gaussian attention network according to claim 8, characterized in that: In the step S2, the formula of the rotation channel Gaussian convolution attention module RCGCA is: RCGCA(x)=GGPA[CRC(Q),CRC(K),CRC(GN(G))]; GGPA represents Gaussian-guided probabilistic attention, CRC is channel rotation convolution, CRC(Q) is the value of the multi-view feature map corresponding to the query Q generated by the output of the rotation space Gaussian convolution attention module RSGCA after channel rotation convolution, CRC(K) is the value of the multi-view feature map corresponding to the key K generated by the output of the rotation space Gaussian convolution attention module RSGCA after channel rotation convolution, CRC(GN(G) is the value of the multi-view feature map corresponding to the normalized Gaussian vector G generated by the output of the rotation space Gaussian convolution attention module RSGCA after channel rotation convolution.

10. The insulator detection method based on hierarchical rotation Gaussian attention network according to claim 9, characterized in that: In step S3, the formula of the multi-pole rotation Gaussian attention MRGA is: MRGA(X)=GGPA{Norm[MRA(Q,K,V)]}; GGPA represents Gaussian-guided probabilistic attention; Norm represents normalization operation; MRA represents multipolar rotation attention module; MRA(Q,K,V) represents the output of multi-polar rotation attention; Q is the query feature map, K is the key feature map, and V is the value feature map; represents the cyclic operation of i from 1 to n, concatenating and merging n results of MR, where each i corresponds to a rotation operation or attention head; MR stands for Multi-Pole Rotation Module, which is responsible for rotating the output of scaled dot product attention SDPA; SDPA stands for scaled dot product attention; is the weighted Cha Xun feature map, is the weighted key feature map, is the weighted value feature map; The linear transformation weight matrix of the i-th attention head Used to perform linear transformation on the query feature map Q, the linear transformation weight matrix of the i-th attention head Used to linearly transform the key feature map K, the linear transformation weight matrix of the i-th attention head Used for linear transformation of the value feature map V; d k is the dimension of the key vector, which is used to scale the dot product result; Softmax is a normalization operation that converts the similarity result of the dot product into a probability distribution.

Images