Power grid infrastructure target detection method and device based on MAMBA-YOLO and medium

Abstract

The invention discloses a power grid infrastructure target detection method and device based on MAMBA-YOLO, and a medium, combines a Mama architecture and a YOLO target detection technology, and solves the bottleneck problems of target detection precision and calculation efficiency in a power grid infrastructure image by introducing an SSM to optimize a YOLO model. Firstly, feature maps of different levels are extracted from input data through a feature extraction module, and then feature fusion is performed by using a neck path aggregation network. And the fused feature map is sent to a target positioning and classification module for object classification and bounding box regression. The method can efficiently identify targets such as equipment, facilities and lines in power grid infrastructure. Through a multi-direction scanning and feature fusion strategy, targets of different scales can be accurately detected under a complex background, and the method has significant advantages especially for small target detection. Compared with a traditional YOLO model, the Mamb-YOLO model has the advantages that the calculation complexity is remarkably reduced while the detection precision is guaranteed, and the requirement of power grid infrastructure target detection for real-time performance is met.

Description

Technical Field

[0001] The present invention relates to the technical field of object detection, and particularly relates to a method, device, and storage medium for power grid infrastructure object detection based on MAMBA-YOLO. Background Art

[0002] With the continuous advancement of the construction of smart power grids, the scale of power infrastructure has expanded rapidly, and the quantity and complexity of infrastructure facilities such as transmission lines and substation equipment have increased year by year. However, the operating environment of power grid infrastructure is complex and changeable, often facing harsh conditions such as high altitude, remote areas, and large-span lines, which pose great challenges to daily inspections and fault troubleshooting. The traditional manual inspection method is not only costly and inefficient but may also have potential safety hazards in complex or dangerous environments. In addition, with the expansion of the scale of power grid operation, the amount of data has also increased exponentially, and traditional monitoring methods can no longer meet the requirements of efficient processing.

[0003] In recent years, object detection technologies based on computer vision have gradually been applied to the field of power grid infrastructure, achieving partial automated inspections. However, existing object detection algorithms still have significant deficiencies in dealing with the complexity of power grid scenarios. For example, power grid infrastructure scenarios often involve problems such as significant differences in object sizes, dense small objects, and complex backgrounds, which pose higher requirements for the robustness of detection algorithms. Object detection methods represented by the YOLO series have received extensive attention in engineering applications due to their real-time performance and high efficiency. However, traditional YOLO models have a large computational amount when processing high-resolution images, and their detection effects on small objects and low-contrast objects in complex scenarios are not ideal. In addition, with the introduction of Transformer technology, although the global modeling ability has been improved, the increase in computational complexity limits its practical application in scenarios with limited resources.

[0004] In this context, new vision models based on State Space Models (SSM) have gradually received attention. SSM shows good application prospects in the fields of image classification and object detection with its linear computational complexity and long-sequence modeling ability. However, how to combine SSM technology with existing object detection frameworks to fully utilize its modeling ability for complex scenarios while retaining the characteristics of efficient real-time detection is still a key difficulty and research hotspot in the current technological development. Therefore, there is an urgent need for an innovative method to organically combine SSM and the YOLO model to address the challenges of object detection in power grid infrastructure scenarios. Summary of the Invention

[0005] The main objective of the present invention is to provide a power grid infrastructure target detection method based on MAMBA - YOLO to solve the technical problems presented in the background art. The present invention utilizes State Space Models (SSM) technology to optimize the YOLO model, combining the advantages of the Mamba architecture, which not only maintains high - efficiency computing power but also improves detection accuracy, especially in target detection tasks in complex power grid scenarios.

[0006] The present invention adopts the following technical solutions to solve the above - mentioned technical problems: A power grid infrastructure target detection method based on MAMBA - YOLO, which is executed by a computer device through the following steps: Input the pre - processed power grid infrastructure image into a pre - constructed Mamba - YOLO target detection model to detect the location and identify the type of defects in the power grid infrastructure image; Among them, the Mamba - YOLO target detection model includes a backbone feature extraction network, a neck path aggregation network, and a head detection network; The pre - processed power grid infrastructure image first enters the optimized deep state space block in the backbone feature extraction network for layer - by - layer feature extraction to obtain a multi - scale feature map; The optimized deep state space block consists of a local space block, a two - dimensional selective scanning module, and a residual gate control block. The multi - scale feature map is input into the neck path aggregation network for bidirectional fusion, and the fused feature map is further input into the head detection network for target detection tasks.

[0007] Furthermore, the processing steps of the local space block are as follows: For a given input feature First, it undergoes depth - wise separable convolution, which operates independently on each input channel without mixing channel information; Then, batch normalization is performed to provide a regularization effect while reducing over - fitting, and the resulting intermediate state is defined as:

[0008] In the formula is a three - dimensional real - number array of the input, is the intermediate state, represents a three - dimensional real - number array, C is the number of channels, H is the height, W is the width, BN represents batch normalization, is the depth - wise separable convolution with a convolution kernel size of 3×3; Finally, the original input and the processed features are fused through residual splicing; Allowing the model to understand and integrate features of different dimensions in the image; ⊕

[0009] where is the output feature, Φ represents the activation function, ⊕ is the residual connection, is the depthwise separable convolution, and the kernel size is 1×1.

[0010] Furthermore, the processing steps of the two-dimensional selective scanning module are as follows. By implementing a scanning strategy in multiple directions including horizontal, vertical, and diagonal, the global context information contained in the input feature map is comprehensively mined; these feature information after being scanned and unfolded are fed into the two-dimensional selective scanning module, and through a merging strategy, i.e., the attention mechanism, the feature information from different directions is effectively fused to form a comprehensive and rich global feature representation, providing a more comprehensive decision-making basis for the model.

[0011] Furthermore, the residual gate control block combines the advantages of the gating mechanism and the deep convolutional residual structure. Through non-linear activation, pointwise convolution realizes the deep fusion of features between channels, and the gating function dynamically adjusts the feature weights according to the feature context, realizing the focusing on key target features and the suppression of background noise; The residual gate control block creates two branches from the input and and implements a fully connected layer in the form of 1×1 convolution on each branch; ,

[0012] The depthwise separable convolution is used as the position encoding module on the branch, and the gradient is more effectively backflowed during the training process through residual concatenation and the performance is significantly improved by retaining and utilizing the spatial structure information of the image; The residual gate control block uses non-linear GeLU as the activation function to control the information flow at each level, and then merges with one branch of through element-wise multiplication, then refines with the global feature through 1x1 convolution to mix the channel information, and finally adds the original input and the features in the hidden layer through residual concatenation; The residual gate control block captures more global features, and the generated output feature

[0013] where represents element-wise multiplication, Φ represents the non-linear GeLU activation function; the gating mechanism in the RG block retains the spatial information by integrating convolutional operations, while making the model more sensitive to the fine-grained features in the image.

[0014] Furthermore, the neck path aggregation network module processes multi-scale features and improves the model's detection ability for targets of different sizes. This module consists of a Feature Pyramid Network (FPN) layer and a bidirectional fusion strategy: The role of the Feature Pyramid Network (FPN) layer is to perform upsampling processing on multi-scale feature maps layer by layer to construct a rich feature pyramid structure; the Feature Pyramid Network (FPN) layer receives feature maps of different levels from the backbone feature extraction network as inputs, and these feature maps have different scales and semantic information; Upsampling operation: For each layer of input feature map, the Feature Pyramid Network (FPN) layer uses the upsampling method to gradually increase its spatial resolution and make it consistent with the feature map of higher resolution in scale; Feature feedback and reinforcement learning: The result obtained by upsampling is divided into two parts for processing; one part of the feature map is fed back to the previous feature extraction module of the backbone feature extraction network; by fusing or recalculating the upsampled feature map with the input features of the previous feature extraction module, the model can better capture the correlation information between multi-scale features, further enrich the semantic content of the feature representation, and improve the model's perception ability for targets of different scales; The other part of the upsampled feature map is then passed into the subsequent module of the neck path aggregation network, that is, the bidirectional fusion strategy module, for deeper feature fusion and optimization.

[0015] Furthermore, the bidirectional fusion strategy is another key component of the neck path aggregation network module. It further improves the expression ability of features through the bidirectional fusion operation between the upsampled feature map and the original input feature map to meet the requirements of multi-scale object detection; Fusion of the upsampled feature map and the original input feature map: In the bidirectional fusion strategy, the upsampled feature map and the original input feature map first perform bidirectional feature fusion; Through the fusion operation, the model can combine the high-level semantic information contained in the upsampled feature map with the low-level detail information in the original input feature map to obtain a more representative fused feature map; Downsampling module: The downsampling module further processes the fused feature map. Through the downsampling operation, the spatial resolution of the feature map is adjusted to adapt to the detection requirements of targets of different scales.

[0016] Furthermore, the head detection network is designed with a decoupled head structure, including a classification head and a detection head; The classification head classifies each feature map through Global Average Pooling (GAP) and outputs the probability distribution of each class, reducing the feature dimension and improving the classification efficiency; The detection head includes a series of convolutional layers and transposed convolutional layers, which are used to generate the bounding box regression values and object confidences for each anchor box, and pointwise convolution and non-linear activation are adopted to improve the regression accuracy.

[0017] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the above method.

[0018] In yet another aspect, the present invention also discloses a computer device including a memory and a processor, where the memory stores a computer program, which, when executed by the processor, causes the processor to execute the steps of the above method.

[0019] As can be seen from the above technical solutions, the present invention discloses a power grid infrastructure target detection method based on Mamba-YOLO. This method combines the Mamba architecture and YOLO target detection technology, and optimizes the YOLO model by introducing the State Space Model (SSM), solving the bottleneck problems of target detection accuracy and computational efficiency in power grid infrastructure images. The method first extracts feature maps of different levels from the input data through a feature extraction module, and then uses a neck path aggregation network for feature fusion. The fused feature maps are fed into a target localization and classification module for object classification and bounding box regression. This method can efficiently identify targets such as equipment, facilities, and lines in power grid infrastructure. Through a multi-directional scanning and feature fusion strategy, the present invention can accurately detect targets of different scales in complex backgrounds, especially having significant advantages in small target detection. Compared with traditional YOLO models, Mamba-YOLO of the present invention significantly reduces the computational complexity while ensuring the detection accuracy, meeting the real-time requirements of power grid infrastructure target detection. This method can be widely applied to fields such as intelligent monitoring, equipment inspection, and fault diagnosis in the power industry.

[0020] Compared with the prior art, the present invention has the following advantages: 1. Efficient multi-scale feature extraction: By means of the optimized deep state space block module in the backbone feature extraction network, combined with the local feature extraction of the local space block, the global modeling ability of the two-dimensional selective scanning module, and the depth channel fusion of the residual gate control block, the detection ability for small targets and complex backgrounds is significantly improved; 2. Feature bidirectional fusion strategy: In the neck path aggregation network, through a bidirectional feature fusion mechanism combining upsampling and downsampling, the global context information and local detail information are effectively balanced, improving the detection accuracy; 3. Optimized head detection network: By decoupling the design to separate the classification and detection tasks, reducing the interference between tasks, and improving the classification accuracy and the robustness of bounding box regression; 4. Grid infrastructure specific target detection ability: It can effectively identify key equipment, transmission lines, transmission towers in the power grid, and small defects in complex scenarios, comprehensively covering the detection requirements in the grid infrastructure scenario; 5. Strong real-time performance: Through the linear complexity characteristics of the Mamba architecture and the efficient framework design of YOLO, the present invention achieves real-time detection performance while maintaining high accuracy, meeting the requirements of high efficiency and accuracy of target detection in the grid infrastructure scenario.

[0021] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present invention, nor is it used to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. Of course, any product implementing the present invention does not necessarily need to achieve all the above-mentioned advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The specification drawings forming a part of this application are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 is the network structure diagram of the Mamba-YOlO power grid target detection model in the present invention; Figure 2 is the network structure of the optimized deep state space block in the backbone feature extraction network module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the protection scope of the present invention.

[0024] The power grid infrastructure target detection method based on Mamba-YOLO in the embodiments of the present invention includes inputting the preprocessed power grid infrastructure image into a pre-constructed Mamba-YOLO target detection model to detect the position and identify the type of defects in the power grid infrastructure image. Among them, the Mamba-YOLO target detection model includes a backbone feature extraction network, a neck path aggregation network, and a head detection network. The preprocessed power grid infrastructure image first enters the optimized deep state space block module of the backbone feature extraction network for feature extraction to obtain a multi-scale feature map. The optimized deep state space block is composed of a local space block, a two-dimensional selective scanning module, and a residual gate control block.

[0025] The obtained multi-scale feature maps are input into the neck path aggregation network for upsampling and downsampling. The neck path aggregation network first performs upsampling layer by layer through the feature pyramid network. The upsampling results are transmitted back to the backbone feature extraction network on one hand to observe the image multiple times to obtain richer feature information, and on the other hand, they are input into the neck path aggregation network for downsampling. The downsampling results are fused bidirectionally with the upsampling results of the feature pyramid network layers. The feature maps obtained from the bidirectional feature fusion are input into the head detection network to perform the object detection task.

[0026] Among them, the core modules of the optimized deep state space block include: 1) Local spatial block: It uses depthwise separable convolution technology to refine the feature extraction process. This technology first processes each channel independently through pointwise convolution, and then integrates information across channels through depthwise convolution. It not only significantly reduces the computational burden of the model, but also enhances the ability to capture local texture and structure of the image, especially showing excellent adaptability when identifying small-sized objects and complex background scenes. The design of the local spatial block enables the model to fully utilize local feature information while maintaining efficient computation, laying a solid foundation for subsequent feature fusion and object detection.

[0027] 2) Two-dimensional selective scanning module: It further expands the model's field of view. By implementing a multi-directional scanning strategy, it comprehensively mines the global context information contained in the input feature maps. These feature information expanded by scanning are sent into the two-dimensional selective scanning module. Through a clever merging strategy (such as weighted average or attention mechanism), the feature information from different directions is effectively fused to form a comprehensive and rich global feature representation, providing a more comprehensive decision-making basis for the model.

[0028] 3) Residual gate control block: It combines the advantages of the gating mechanism and the deep convolutional residual structure. Through a gating function (such as GLU or GSGU), it dynamically adjusts the feature weights according to the feature context to achieve focusing on key object features and suppressing background noise. At the same time, the deep convolutional residual structure strengthens the feature propagation path through skip connections, alleviates the gradient vanishing problem in deep networks, and enhances the consistency and stability of features.

[0029] The head detection network of the embodiment of the present invention adopts a decoupled head structure, separating the classification and detection heads to make the processing task more efficient and accurate.

[0030] The head detection network part is responsible for the final object detection and classification tasks, including a detection head and a classification head. The detection head contains a series of convolutional layers and deconvolutional layers for generating detection results. These layers are responsible for predicting the bounding box regression values and the confidence of object presence for each anchor box. The classification head uses global average pooling (GAP) to classify each feature map, and by reducing the dimension of the feature map, outputs the probability distribution of each class. The design of the classification head enables effective handling of multi-class classification tasks.

[0031] The neck path aggregation network includes multiple levels of feature fusion modules, adopting a bidirectional fusion strategy to perform upsampling and downsampling between feature maps of different scales, enhancing the detection ability for complex objects.

[0032] The local spatial block fuses the local features with the original input features through residual connections, and the fused features enhance the detection ability for small targets in power grid infrastructure.

[0033] The scan merging module in the two-dimensional selective scanning module can efficiently restore the original dimension of the feature map, enabling the model to maintain high performance when processing high-resolution images.

[0034] The following is a specific description: A power grid infrastructure object detection method based on Mamba-YOLO. The object detection method is to input the preprocessed power grid infrastructure image into a pre-constructed Mamba-YOLO object detection model to detect the location and identify the type of defects in the power grid infrastructure image. Among them, the Mamba-YOLO object detection model includes a backbone feature extraction network, a neck path aggregation network, and a head detection network. The preprocessed power grid infrastructure image first enters the optimized deep state space block module in the backbone feature extraction network for layer-by-layer feature extraction to obtain multi-scale feature maps. The optimized deep state space block is composed of a local spatial block, a two-dimensional selective scanning module, and a residual gate control block. The multi-scale feature maps are input into the neck path aggregation network for bidirectional fusion, and the fused feature maps are further input into the head detection network for object detection tasks.

[0035] The optimized deep state space block module in the backbone feature extraction network is one of the core designs of the present invention and includes the following modules: 1) Local Spatial Block: The depthwise separable convolution technique is utilized to refine the feature extraction process. This technique independently processes each channel through channel-wise convolution and then integrates information across channels via pointwise convolution. It not only significantly reduces the computational burden of the model but also enhances the ability to capture local textures and structures in images, especially demonstrating excellent adaptability when identifying small-sized objects and complex background scenes. The design of the local spatial block enables the model to fully utilize local feature information while maintaining efficient computation, laying a solid foundation for subsequent feature fusion and object detection. Specifically, for a given input feature , it first undergoes depthwise separable convolution, which operates independently on each input channel without mixing channel information. It effectively extracts the local spatial information of the input feature map while reducing the computational cost and the number of parameters. Then, batch normalization is performed, providing a certain degree of regularization effect while reducing overfitting. The resulting intermediate state is defined as:

[0036] In the formula is the input three-dimensional real number array, is the intermediate state, represents a three-dimensional real number array, C is the number of channels, H is the height, W is the width, BN represents batch normalization, is the depthwise separable convolution with a kernel size of 3×3; Finally, the original input is fused with the processed features through residual concatenation. This allows the model to understand and integrate features in different dimensions of the image, thereby enhancing the robustness to scale changes.

[0037] ⊕

[0038] where is the output feature, Φ represents the activation function, ⊕ is the residual connection, is the depthwise separable convolution with a kernel size of 1×1.

[0039] The local spatial block can effectively capture and represent the local spatial information of the input feature map and fuse it with the original input to enhance the representation of features, providing Mamba-YOLO with a powerful ability to handle scale changes and context information in visual data.

[0040] 2) 2D Selective Scanning Module: Further expands the model's field of view. By implementing multi-directional (including horizontal, vertical, diagonal, etc.) scanning strategies, it comprehensively mines the global context information contained in the input feature map. These feature information after being scanned and expanded are sent into the 2D Selective Scanning Module. Through a clever merging strategy (attention mechanism), the feature information from different directions is effectively fused to form a comprehensive and rich global feature representation, providing a more comprehensive decision-making basis for the model.

[0041] 3) Residual Gate Control Block: Combines the advantages of the gating mechanism and the deep convolutional residual structure. Through non-linear activation and pointwise convolution, it realizes the deep fusion of features between channels, and the gating function dynamically adjusts the feature weights according to the feature context to achieve the focusing on key target features and the suppression of background noise. For example, Figure 2 In it, the simple design of the Residual Gate Control Block aims to improve the model's performance at low computational cost. The Residual Gate Control Block creates two branches from the input and and and implements a fully connected layer in the form of 1×1 convolution on each branch.

[0042]

[0043]

[0044] Depthwise separable convolution is used as the position encoding module on the branch. By residual concatenation, it more effectively backflows gradients during the training process and significantly improves the performance by retaining and utilizing the spatial structure information of the image. The Residual Gate Control Block uses non-linear GeLU as the activation function to control the information flow at each level, and then merges with one branch of through element-wise multiplication, then refines with global features through 1x1 convolution to mix channel information, and finally adds with the original input and the features in the hidden layer through residual concatenation. The Residual Gate Control Block can capture more global features while only slightly increasing the computational cost. The resulting output feature

[0045] is defined as:

[0046] where Φ represents the activation function (non-linear GeLU). The gating mechanism in the RG block retains spatial information by integrating convolutional operations, while making the model more sensitive to fine-grained features in the image. 1) Feature Pyramid Network (FPN) The Feature Pyramid Network is one of the core components of the neck path aggregation network module. Its main function is to perform upsampling processing on multi-scale feature maps layer by layer to construct a rich feature pyramid structure. Specifically, the FPN layer receives feature maps of different levels from the backbone feature extraction network as inputs, and these feature maps have different scales and semantic information.

[0047] Upsampling operation: For each layer of input feature map, the FPN layer adopts a specific upsampling method, such as bilinear interpolation or deconvolution, etc., to gradually increase its spatial resolution and make it consistent with the feature map of higher resolution in scale. This upsampling operation can not only restore the detailed information of the feature map but also provide a basis for subsequent feature fusion.

[0048] Feature feedback and reinforcement learning: The result obtained by upsampling is processed in two parts. One part of the feature map is fed back to the previous feature extraction module of the backbone feature extraction network. The significance of this operation lies in strengthening the feature learning process. By fusing or recomputing the upsampled feature map with the input features of the previous feature extraction module, the model can better capture the correlation information between multi-scale features, further enrich the semantic content of the feature representation, and improve the model's perception ability of targets at different scales. The other part of the upsampled feature map is then passed into the subsequent module of the neck path aggregation network, that is, the bidirectional fusion strategy module, for deeper feature fusion and optimization.

[0049] 2) Bidirectional fusion strategy The bidirectional fusion strategy is another key component of the neck path aggregation network module. It further enhances the feature expression ability through the bidirectional fusion operation between the upsampled feature map and the original input feature map to meet the requirements of multi-scale object detection.

[0050] Fusion of the upsampled feature map and the original input feature map: In the bidirectional fusion strategy, the upsampled feature map and the original input feature map first perform bidirectional feature fusion. This fusion method can be simple element addition, concatenation, or more complex fusion operations, such as weighted fusion with attention mechanism, etc. Through the fusion operation, the model can combine the high-level semantic information contained in the upsampled feature map with the low-level detailed information in the original input feature map to obtain a more representative fused feature map.

[0051] Downsampling Module: The downsampling module plays a crucial role in the entire bidirectional fusion strategy. It further processes the fused feature maps. Through downsampling operations such as max pooling, average pooling, etc., it adjusts the spatial resolution of the feature maps to adapt to the detection requirements of targets at different scales. At the same time, the downsampling module can also introduce more context information to help the model better understand the relationship between the target and the surrounding environment, thereby improving the accuracy and robustness of object detection.

[0052] Through the above bidirectional fusion and downsampling operations, the model can further optimize the feature expression ability. On the one hand, the fused multi-scale features can more comprehensively describe the characteristics of the target; on the other hand, the downsampling operation helps to reduce the redundant information in the feature maps, highlight the key features, and make the model more focused on the essential features of the target, thus improving the performance of object detection.

[0053] The design of the head detection network adopts a decoupled head structure, including a classification head and a detection head: The classification head classifies each feature map through global average pooling (GAP), outputs the probability distribution of each class, reduces the feature dimension and improves the classification efficiency. The detection head includes a series of convolutional layers and transposed convolutional layers, which are used to generate the bounding box regression values and object confidence levels of each anchor box, and uses pointwise convolution and non-linear activation to improve the regression accuracy.

[0054] The embodiments of the present invention design the following detection capabilities for the special requirements in the power grid infrastructure scenario: 1) Power grid equipment detection: The model can accurately identify key equipment in the substation, such as transformers, circuit breakers, and disconnectors, etc., locate the equipment position and judge its operating status.

[0055] 2) Cable line identification: Through the multi-scale feature extraction and fusion of Mamba-YOLO, it detects high-altitude and ground cable lines, accurately calibrates the line position and identifies possible damaged areas.

[0056] 3) Transmission tower status monitoring: Utilizing the local information capture ability of the local spatial block, it identifies the structural components of the transmission tower and detects problems such as tower body cracks and component detachment.

[0057] 4) Small target detection: For small faults in power grid infrastructure (such as bolt loosening, cable damage, etc.), through the global information fusion and high-resolution processing ability of the two-dimensional selective scanning module, high-precision detection is achieved.

[0058] 5) Separation of complex background targets: Facing the common complex backgrounds in the power grid environment (such as vegetation occlusion, high-altitude background interference, etc.), the remaining gate control block enhances the ability to distinguish between the target and the background by deeply fusing the feature information.

[0059] The Mamba-YOLO model of the present invention adopts multiple optimization techniques, significantly improving the object detection performance: 1) An optimized deep state space block module is introduced into the backbone feature extraction network, combining the local feature extraction of the local space block, the global modeling ability of the two-dimensional selective scanning module, and the depth channel fusion of the residual gate control block, significantly improving the small object detection ability and the feature capture efficiency in complex scenarios; 2) The neck path aggregation network adopts a bidirectional fusion strategy, fully extracting multi-scale information in upsampling and downsampling to enhance the model's perception ability of objects; 3) In the head detection network part, the detection and classification tasks are separated through decoupled design, reducing interference between tasks and improving classification accuracy and the robustness of bounding box regression; 4) Considering the particularity of power grid infrastructure targets, through the cooperation of multiple modules, the detection of large and small targets is taken into account to adapt to diverse power grid scenarios; 5) Through the linear complexity design of the state space model, the processing ability of high-resolution images is optimized while maintaining real-time performance.

[0060] The preprocessing method includes: using an image annotation tool to annotate the target positions and defect types in the power grid infrastructure images to generate Pascal VOC format files for training and validating the Mamba-YOLO object detection model.

[0061] The following is an example for illustration: Example 1: A method for power grid infrastructure object detection based on Mamba-YOLO, which is specifically carried out according to the following steps: Step S1, construct a Mamba-YOLO object detection model (1) The network structure of the Mamba-YOLO object detection model is as Figure 1 shown. It is composed of three major parts: a backbone feature extraction network, a partial path aggregation network, and a head detection network. The preprocessed power grid infrastructure images first enter the backbone network of the backbone feature extraction network for layer-by-layer feature extraction to obtain multi-scale feature maps; the backbone feature extraction network contains several optimized deep state space block modules, and each optimized deep state space block is composed of a local space block, a two-dimensional selective scanning module, and a residual gate control block in series, which can effectively model remote dependence features while retaining local details.

[0062] The obtained multi-scale feature maps are further input into the neck path aggregation network for feature fusion. First, they go through the feature pyramid network layer for upsampling. On one hand, the upsampling results of the feature pyramid network layer are fed back to the previous layer of the backbone feature extraction network to observe the original image multiple times. On the other hand, they are input into the neck path aggregation network for downsampling. The downsampling results are fused with the upsampling results of the feature pyramid network layer in a two-way manner to obtain the fused feature maps, which are then passed to the head detection network to perform the object detection task.

[0063] (2) Aiming at the characteristics of multiple scenarios, multiple defect categories, and dense small targets in power grid infrastructure construction, various improvements are introduced in the optimized deep state space block module of the backbone feature extraction network: Local space block: Depthwise separable convolution is used to extract the local details of the input image, point convolution is used to enhance the channel fusion ability, and the original feature information is retained through residual connection. Two-dimensional selective scanning module: Based on the multi-directional scanning expansion and merging of the state space model, the long-range dependencies in high-resolution images are processed in parallel to improve the model's adaptability to complex scenarios. Residual gate control block: A gating mechanism is introduced into the residual structure. By combining channel attention with a non-linear activation function, the attention to key target features is enhanced and background noise is suppressed, thereby reducing missed detections and false detections in the multi-interference environment of the power grid.

[0064] (3) To better balance multi-scale object detection and real-time performance, the neck path aggregation network adopts an up-and-down two-way feature fusion strategy: The upsampling results of the feature pyramid network layer are fed back to the early feature layer to strengthen the ability of the shallow layer to capture small targets; the downsampling layer fuses the high-level semantics with the shallow layer detail features to improve the modeling ability for large targets and global information.

[0065] (4) A decoupled structure is adopted in the head detection network part: The position regression branch and the class prediction branch are split, so that box regression and class classification are trained independently to reduce mutual interference; for confidence prediction, the same backbone convolutional layer output as the regression branch is used to determine whether there is an object in the anchor box and its credibility.

[0066] Step S2: Model training First, input the power grid infrastructure images and annotation information into the Mamba-YOLO object detection model constructed in step S2 for training; use common object detection loss strategies (such as IoU loss, Focal Loss, or cross-entropy) to optimize the bounding box position regression, class prediction, and confidence prediction respectively to obtain a trained Mamba-YOLO object detection model. Then, input the real-time collected power grid infrastructure images into the trained Mamba-YOLO object detection model after the same preprocessing steps, and the object detection and recognition results, including the defect category and specific location, can be output.

[0067] Step S3: Performance detection 1. Detection results of different defect types To verify the effectiveness of the power grid infrastructure object detection method based on Mamba-YOLO described in the present invention, input the real-scene images collected by a certain power grid operation and maintenance department into the Mamba-YOLO model trained in Embodiment 1, and the model identifies different defect positions and types in the images. The obtained detection results are shown in Table 1:

[0068] As can be seen from Table 1, the Mamba-YOLO object detection model trained by the present invention can achieve high accuracy, recall rate, and mean average precision for various power grid infrastructure defects. Among them, the detection performance of insulator defects and fitting defects is particularly stable, and it has good detection capabilities for small targets and complex backgrounds.

[0069] 2. Comparison of detection effects of different models To comprehensively evaluate the performance of Mamba-YOLO, compare and test the trained model with YOLOv4, YOLOv5, and YOLOv8. The results are shown in Table 2:

[0070] As can be seen from Table 2, YOLOv4 has a certain advantage over YOLOv5 because the region proposal network has stronger recognition ability for small targets, but its detection speed is relatively slow; YOLOv8 can also achieve a balance between speed and accuracy through multi-scale feature extraction and attention mechanism; the Mamba-YOLO proposed in the present invention reaches the highest level in terms of accuracy, recall rate, mean average precision, and detection speed. Especially, it maintains good detection performance under small targets and complex interference backgrounds, and can meet the requirements of power grid inspection for real-time and accuracy.

[0071] Generally speaking, the Mamba-YOLO model of the embodiments of the present invention can perform high-precision detection on the positions of power grid infrastructure defects in complex backgrounds and shows good recognition capabilities for different defect types (such as tower defects, insulator defects, fitting defects, etc.); through multi-level feature fusion of the backbone feature extraction network and the neck path aggregation network, the model achieves faster convergence, and both the detection speed and accuracy are significantly better than other typical detection algorithms. The present invention can be flexibly deployed on edge devices or server sides to meet the requirements of various power grid scenarios, which is of great significance for improving the efficiency and safety of power grid inspection.

[0072] On the other hand, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the above method.

[0073] On yet another hand, the present invention also discloses a computer device including a memory and a processor, where the memory stores a computer program, which, when executed by the processor, causes the processor to execute the steps of the above method.

[0074] In another embodiment provided by the present application, there is also provided a computer program product containing instructions, which, when running on a computer, causes the computer to execute any one of the above methods for power grid infrastructure target detection based on MAMBA-YOLO.

[0075] It can be understood that the system provided by the embodiments of the present invention corresponds to the method provided by the embodiments of the present invention. The explanations, examples, and beneficial effects of related contents can refer to the corresponding parts in the above method.

[0076] The embodiments of the present application also provide an electronic device including a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete mutual communication through the communication bus. The memory is used to store a computer program. The processor is used to implement the above method for power grid infrastructure target detection based on MAMBA-YOLO when executing the program stored in the memory.

[0077] The communication bus mentioned in the above electronic device may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc.

[0078] The communication interface is used for communication between the above-mentioned electronic device and other devices.

[0079] The memory may include a random access memory (Random Access Memory, RAM) or a non-volatile memory (Non-Volatile Memory, NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located far from the aforementioned processor.

[0080] The aforementioned processor may be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; it may also be a digital signal processor (Digital Signal Processing, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field-programmable gate array (Field-Programmable Gate Array, FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0081] It should also be noted that the electronic device also includes a terminal device, which can also be called a terminal, a user equipment (UE), a mobile station (MS), a mobile terminal (MT), etc. The terminal device can be a mobile phone, a smart TV, a wearable device, a tablet computer (Pad), a computer with wireless transceiver function, a virtual reality (VR) terminal device, an augmented reality (AR) terminal device, a wireless terminal in industrial control, a wireless terminal in self-driving, a wireless terminal in remote medical surgery, a wireless terminal in smart grid, a wireless terminal in transportation safety, a wireless terminal in smart city, a wireless terminal in smart home, and so on. The embodiments of the present application do not limit the specific technologies and specific device forms adopted by the terminal device.

[0082] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from a website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that the computer can access or a data storage device such as a server, data center, etc. that includes one or more available media integrated. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid state disk (SSD)), etc.

[0083] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0084] In addition, it should be noted that if there are directional indications (such as up, down, left, right, front, back...) involved in the embodiments of the present invention, the directional indications are only used to explain the relative position relationship and movement conditions between components in a specific posture. If the specific posture changes, the directional indications will also change accordingly.

[0085] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, such descriptions of "first", "second", etc. are for descriptive purposes only, and should not be construed as indicating or implying their relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In addition, the meaning of "and / or" appearing throughout the text includes three parallel scenarios. Taking "A and / or B" as an example, it includes Scenario A, or Scenario B, or the scenario where both A and B are satisfied simultaneously. In addition, in the embodiments of the present invention, "a plurality of" means two or more. In addition, the technical solutions between various embodiments can be combined with each other, but it must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

Claims

1. A power grid infrastructure target detection method based on MAMBA - YOLO, characterized in that, Including the following steps, Input the preprocessed power grid infrastructure image into the pre-constructed Mamba-YOLO object detection model to detect the location and identify the type of defects in the power grid infrastructure image; Among them, the Mamba-YOLO object detection model includes a backbone feature extraction network, a neck path aggregation network, and a head detection network; The preprocessed power grid infrastructure image first enters the optimized deep state space block in the backbone feature extraction network for layer-by-layer feature extraction to obtain multi-scale feature maps; The optimized deep state space block consists of a local space block, a two-dimensional selective scanning module, and a residual gate control block. The multi-scale feature maps are input into the neck path aggregation network for bidirectional fusion, and the fused feature maps are further input into the head detection network for object detection tasks.

2. The method for detecting grid infrastructure targets based on MAMBA-YOLO according to claim 1, wherein, The processing steps of the local space block are as follows, For a given input feature First, it undergoes depthwise separable convolution, which operates separately on each input channel without mixing channel information; Then batch normalization is performed, providing a regularization effect while reducing overfitting, and the intermediate state obtained is defined as: In the formula is a three-dimensional real number array as input, is an intermediate state, represents a three-dimensional real number array, C is the number of channels, H is the height, W is the width, and BN represents batch normalization, is a depthwise separable convolution with a convolution kernel size of 3×3; Finally, the original input and the processed features are fused through residual splicing; Allowing the model to understand and integrate features in different dimensions of the image; ⊕ wherein is the output feature, Φ represents the activation function, and ⊕ is the residual connection, is the depthwise separable convolution, and the convolution kernel size is 1×1.

3. The method for detecting power grid infrastructure targets based on MAMBA - YOLO according to claim 2, wherein, The processing steps of the two-dimensional selective scanning module are as follows, By implementing a scanning strategy in multiple directions including horizontal, vertical, and diagonal, the global context information contained in the input feature map is comprehensively mined; these feature information after scanning and expansion are sent into the two-dimensional selective scanning module. Through a merging strategy, that is, the attention mechanism, the feature information from different directions is effectively fused to form a comprehensive and rich global feature representation, providing a more comprehensive decision-making basis for the model.

4. The method for detecting grid infrastructure targets based on MAMBA-YOLO according to claim 1, wherein, The residual gate control block combines the advantages of the gating mechanism and the deep convolutional residual structure. Through non-linear activation and pointwise convolution, it realizes the deep fusion of features between channels, and the gating function dynamically adjusts the feature weights according to the feature context, achieving the focusing on key target features and the suppression of background noise; The remaining gate control block receives input Create two branches and and implement a fully connected layer in the form of 1×1 convolution on each branch; ， Depthwise separable convolutions are used as the position encoding module on the branches, which can more effectively backflow gradients during training through residual cascading and significantly improve performance by preserving and exploiting the spatial structure information of the images; The remaining gate control block uses the non-linear GeLU as the activation function to control the information flow at each level, and then merges with one branch of through element-wise multiplication, and then refines with global features through 1x1 convolution to mix channel information, and finally adds the original input and the features in the hidden layer through residual concatenation; The remaining gate control block captures more global features, and the generated output features are defined as: where denotes element-wise multiplication, and Φ denotes the non-linear GeLU activation function; the gating mechanism in the RG block preserves spatial information by integrating convolutional operations, while making the model more sensitive to fine-grained features in the image.

5. The method for detecting grid infrastructure targets based on MAMBA - YOLO according to claim 1, wherein, The neck path aggregation network module processes multi-scale features and improves the model's detection ability for targets of different sizes. This module consists of a Feature Pyramid Network layer FPN and a bidirectional fusion strategy: The role of the Feature Pyramid Network layer FPN is to perform layer-by-layer upsampling on the multi-scale feature maps to construct a rich feature pyramid structure; the Feature Pyramid Network layer FPN receives feature maps at different levels from the backbone feature extraction network as inputs, and these feature maps have different scales and semantic information; Upsampling operation: For each layer of input feature map, the Feature Pyramid Network layer FPN uses the upsampling method to gradually increase its spatial resolution to make it consistent with the feature map at a higher resolution in scale; Feature backpropagation and reinforcement learning: The result obtained by upsampling is divided into two parts for processing; a part of the feature map will be backpropagated to the previous feature extraction module of the backbone feature extraction network; by fusing or recalculating the upsampled feature map with the input features of the previous feature extraction module, the model can better capture the correlation information between multi-scale features, further enrich the semantic content of the feature representation, and improve the model's perception ability for targets of different scales; The feature map after upsampling of the other part is then fed into the subsequent module of the neck path aggregation network, namely the bidirectional fusion strategy module, for deeper feature fusion and optimization.

6. The method for detecting power grid infrastructure targets based on MAMBA-YOLO according to claim 5, characterized in that, The bidirectional fusion strategy is another key component of the neck path aggregation network module. It further enhances the feature expression ability through the bidirectional fusion operation between the upsampled feature map and the original input feature map to meet the requirements of multi-scale object detection. Fusion of the upsampled feature map and the original input feature map: In the bidirectional fusion strategy, the upsampled feature map and the original input feature map first perform bidirectional feature fusion. Through the fusion operation, the model can combine the high-level semantic information contained in the upsampled feature map with the low-level detail information in the original input feature map to obtain a more representative fused feature map. Downsampling module: The downsampling module further processes the fused feature map. Through the downsampling operation, it adjusts the spatial resolution of the feature map to adapt to the detection requirements of different-scale objects.

7. The method for detecting grid infrastructure targets based on MAMBA - YOLO according to claim 1, wherein, The head detection network is designed with a decoupled head structure, including a classification head and a detection head. The classification head classifies each feature map through global average pooling (GAP), outputs the probability distribution of each class, reduces the feature dimension, and improves the classification efficiency. The detection head includes a series of convolutional layers and transposed convolutional layers, which are used to generate the bounding box regression values and object confidence levels of each anchor box, and adopt pointwise convolution and non-linear activation to improve the regression accuracy.

8. A computer-readable storage medium, characterized in that, A computer program is stored, and when the computer program is executed by a processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 7.

9. A computer device, characterized in that, It includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 7.

Images