A low-illumination power small insulator detection method based on cross-domain cooperative gating fusion

Abstract

The application discloses a kind of based on cross-domain collaborative gating fusion low-illumination electric power small insulator detection method, comprising step one, constructing MFN-Net network model, step two, the visible light image and infrared image are preprocessed;Step three, obtain visible light feature by the backbone network of CSPDarknet 53 construction, step four, obtain infrared feature by cascaded HCT module, step five, visible light feature and infrared feature are fused to obtain deep semantic feature, step six, after fusion enhancement, the accurate boundary box coordinates of each detected insulator and the corresponding confidence score are output by multi-scale adaptive detection head module to deep semantic feature.This application can effectively solve the problems of small insulator feature extraction difficulty, modal interference, high miss detection and false detection rate in low-illumination scene, significantly improve the detection accuracy and robustness, provide technical support for intelligent inspection of power transmission line, with the advantages of high detection accuracy and strong environmental adaptability.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, deep learning, and intelligent operation and maintenance of power equipment, specifically to a low-light power small insulator detection method based on cross-domain collaborative gating fusion, which is particularly suitable for high-precision, real-time automatic detection of small insulators in transmission lines by UAVs under low-light conditions. Background Technology

[0002] Insulators are critical components of power transmission lines, and their condition directly affects power grid safety. With the development of intelligent inspection, vision-based drone-based automatic inspection has become mainstream. However, low-light environments pose significant challenges to visual inspection: due to insufficient illumination, visible light images exhibit reduced feature saliency, blurred details, and are easily interfered with by strongly similar background objects; simultaneously, small insulators themselves lack features, further increasing the difficulty of feature extraction under low light conditions, leading to a significant decrease in the accuracy of traditional visible light detection algorithms and prominent issues of missed detections and false detections.

[0003] Infrared thermal imaging technology, based on the principle of thermal radiation, can highlight the thermal characteristics of insulators under low illumination and avoid light interference, naturally complementing visible light images. However, infrared images lack rich texture and edge details, making it difficult to accurately extract the geometry of insulators. Using them alone cannot achieve high-precision positioning and category differentiation. The inherent limitations of single-mode algorithms make low-illuminance insulator detection technology unable to meet the high-precision requirements of actual inspections, becoming a key bottleneck restricting the intelligent operation and maintenance of power systems around the clock. Existing solutions mainly have the following shortcomings: (1) The single-modal enhancement detection method follows the separate process of "enhancement first, detection later". Its limitations are: noise and artifacts are easily introduced in the enhancement stage, and key features lost in the imaging process cannot be recovered; at the same time, when dealing with small targets and complex scenes, the target and background may be enhanced at the same time, increasing the difficulty of differentiation.

[0004] (2) Early fusion-based method: Low-light visible light and infrared images are directly stitched together in the channel dimension to form a four-channel input single network. This method has a simple model, but ignores the huge domain differences between the two imaging modes. The fusion features that have not been aligned and interacted with have a lot of interference, and the effect is unstable.

[0005] (3) Attention-based approach based on intermediate fusion: An attention mechanism is introduced into the intermediate layer of the network for feature interaction. The limitations of this approach are: the attention mechanism is simple to design and it is difficult to capture deep semantic associations across modalities; the interaction process is fixed and unidirectional and lacks dynamic adaptive ability; at the same time, it is sensitive to registration errors between modalities, which reduces the discriminativeness of the fused features.

[0006] Existing methods often employ simple channel or spatial attention, resulting in insufficient intermodal interaction depth. Furthermore, they fail to specifically address the characteristics of small targets and weak features on insulators under low illumination, leading to low parameter utilization and poor real-time inference performance. Therefore, there is an urgent need for a low-illumination insulator detection method that can deeply mine and adaptively fuse complementary information from visible and infrared modes, while specifically enhancing the representation of small target features, and balancing detection accuracy and inference efficiency. Summary of the Invention

[0007] To address the aforementioned issues, this invention provides a method for detecting low-light power insulators based on cross-domain collaborative gating fusion, which accurately quantifies the impact of drive circuit parameters on drive signal interference, providing theoretical support for the design reliability of high-frequency drives for wide-bandgap power devices.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows: A method for detecting small electrical insulators in low-light conditions based on cross-domain collaborative gating fusion includes the following steps: Step 1: Construct and train the MFN-Net network model to obtain the trained MFN-Net network model. The MFN-Net network model consists of three main parts: a bimodal feature extraction network, a cross-modal fusion network, and a multi-scale adaptive detection head. The bimodal feature extraction network includes a feature extraction structure built with CSPDarknet53 and a feature extraction structure built with cascaded HCT modules. The cross-modal fusion network includes four identical DMFT modules, which include SCMT, IFF, and FGM modules. The multi-scale adaptive detection head includes AFF, SFF modules, and detection heads for different target sizes. Step 2: Visible light images collected during transmission line inspections. With infrared images Preprocessing is required; Step 3: Process the pre-processed visible light image Inputting a backbone network constructed using CSPDarknet 53, the system extracts multi-level spatial detail features through the CSPDarknet 53 backbone network connection structure, and outputs visible light features at four scales. , , , ; , , , The scales are 160×160×128, 80×80×256, 40×40×512, and 20×20×1024, respectively, in pixels; Step 4: Process the pre-processed infrared image Input cascaded HCT modules; where each HCT The module operates in parallel. CWSA module and AKAC module Deeply analyze the preprocessed infrared images Thermal radiation characteristics of medium insulators, outputting infrared characteristics at four scales , , , ; , , , The scales are 160×160×128, 80×80×256, 40×40×512, and 20×20×1024, respectively, in pixels; Step 5: Identify visible light characteristics , , , With infrared features , , , One-to-one correspondence and splicing along the channel dimension to obtain splicing features , , , splicing features , , , Input to four identical DMFT modules, outputting high-resolution features with the same size but double the number of channels. Shallow features Mid-layer characteristics Deep features ; Step Six , , , Input to the multi-scale adaptive detection head module: where, , , After convolutional upsampling and efficient fusion and enhancement using the C3K2 module, fused and enhanced features at three scales are output. , , The AFF module will , , Upsampling to The same size yields the feature map Then the SFF module will and Adaptive weighted fusion is performed to obtain weighted fusion features for small target detection. ; Step Seven , , , The data is input into detection heads with dimensions of 160×160×40, 80×80×256, 40×40×512, and 20×20×1024 respectively to complete the bounding box regression and category prediction of small target insulators.

[0009] In a further improvement, in step one, the MFN-Net network model is trained until the composite loss function is minimized, thus obtaining the trained MFN-Net network model. The formula for the composite loss function is as follows:

[0010] Total loss function, The contribution of bounding box positioning accuracy to the total loss. The contribution of the accuracy of the control target classification to the total loss. The contribution of bounding box regression stability to the total loss is controlled. Bounding box loss, Classification loss, Distribution focus loss.

[0011] In a further improvement, in step two, the preprocessing involves processing the visible light image. With infrared images The size has been adjusted to 512×512 pixels, and all data are in JPG format.

[0012] In a further improvement, in step three, the feature extraction structure constructed by CSPDarknet53 is applied to visible light images. By progressively abstracting image information through multi-level feature extraction, visible light images... Through convolutional blocks and cross-stage modules, the spatial resolution is halved at each stage, while the number of channels increases exponentially, forming a feature pyramid from edge texture to high-level semantics:

[0013] This represents the i-th visible light feature. These are visible light images. Height and width, Based on the number of channels, The channel spread factor is used, and the spatial resolution is calculated according to... Decreasing proportionally, number of channels according to Increasing exponentially; Indicates the first The number of channels in the layer, R indicates that each element in the feature map is a real number, and CSPDarknet53 indicates the feature extraction structure constructed by CSPDarknet53.

[0014] Further improvements are made, and the specific steps of step four are as follows:

[0015] in, Representation layer normalization, It was through Layer output, It was through CWSA Branches and AKAC The output of the branch, It was through and Output, For layer normalization; Indicates the first One infrared feature, , This represents the i-th visible light feature.

[0016] In a further improvement, the data processing method of the SCMT module in the DMFT module in step five is as follows:

[0017]

[0018]

[0019] ; in, Representing visible light and infrared light value, Represents visible light value, Indicating infrared value, Represents visible light value, Indicating infrared value, Represents visible light value, Represents visible light value; Represents the normalized exponential function, This represents DropPath regularization; Pi represents the i-th concatenated feature; This indicates the output of the infrared features after the interaction. This represents the output of the visible light features after the interaction, where T represents the transpose operation in the matrix. Represents a 1×1 convolution. This indicates the position encoding, and d represents the scaling factor. This indicates a splicing operation. Indicates the features after interaction; The data processing method of the IFF module is as follows:

[0020] in, This represents depthwise separable convolution. Represents a gated linear unit. Indicates enhanced features; Will , Features are obtained by concatenating the components. ; The data processing method for the FGM module is as follows:

[0021] Indicates channel attention weights. Indicates spatial attention weights; E This indicates features that have undergone interaction and enhancement; T i Representing features at different levels .

[0022] A further improvement is made in step six, , , After passing through the enhanced PANet structure, multi-scale features are output and shallow features are fused via the bidirectional paths of FPN and PAN. , , :

[0023] , , All are intermediate features of PANet. This refers to the C3K2 feature extraction module. Number of input channels Number of output channels This indicates a 2x upsampling operation. This indicates a 2x downsampling operation. Channel splicing operation; out=256 or 512 or 1024; in=768 or 1536; The AFF module first uses convolution operations and upsampling techniques to convert three shallow features... , , The size is expanded to perfectly match the shallow detail features, solving the feature misalignment problem caused by size differences in traditional fusion. Under this premise, an adaptive fusion mechanism is introduced, which fuses the matched features through a dynamic weight allocation strategy, fully mining deep semantic information and shallow detail information, avoiding information redundancy and loss problems in traditional fusion, and outputting semantically rich features. ; The feature pyramid outputs three feature maps at different scales. , , In the middle, located at the pixel point ( , The eigenvectors at position ) are respectively , , The fusion formula for corresponding hierarchical features is as follows:

[0024] Here, , , The weights correspond to the three feature layers output by the pyramid module, and the following conditions must be met: , , , The range is [0,1]; The SFF module calculates the fused feature map through weighted summation; it introduces two learnable scalar parameters. and Then, the Sigmoid function is used to map it to the range of 0-1 to obtain the fusion weights. , Shallow features and Multiplication, fusion of deep features and Multiply them, then add them together to get the output feature. This overcomes the limitations of insufficient semantics from single shallow features or ambiguous spatial localization from single deep features. The implementation process is as follows:

[0025]

[0026]

[0027] for Activation function These are learnable deep weights. These are learnable shallow weights.

[0028] A further improvement is made in step seven, , , , The data is fed to four detection heads at different scales. Each detection head is connected to a neck network and is used to predict the target category and regress the bounding box of the input fused feature map. The output includes the target category, confidence score, and location information. The detection head adopts a multi-scale structure, independently predicting feature maps with resolutions of 160×160×40, 80×80×256, 40×40×512, and 20×20×1024. The preliminary prediction results from each scale are then integrated, and a non-maximum suppression algorithm is applied to filter redundant detection boxes based on a set confidence score of 0.25 and an intersection-over-union (IoU) threshold of 0.7. The precise bounding box coordinates of each detected insulator and the corresponding confidence score are output, thus completing the bounding box regression and category classification prediction for the small target insulator.

[0029] Advantages of this invention: (1) A multi-scale feature extraction network was constructed. The network consists of two parts: the CSPDarknet 53 structure and the multi-level HCT module. The CSPDarknet 53 structure extracts the geometric shape and texture details of low-light visible light images. The cascaded HCT module mines the global and local enhancement features of thermal radiation in infrared images. The hybrid convolution-transformer (HCT) module improves the feature capture capability of small targets through the parallel structure of multi-head channel attention (CWSA) and adaptive kernel aggregation convolution (AKAC).

[0030] (2) After extracting visible light features and infrared features at different scales, the dual-modal features are efficiently fused through the dual-modal fusion module (DMFT). The splicing feature expression is enhanced by the improved feedforward network (IFF). The bidirectional interactive channel of infrared and visible light features is established by the simple cross-modal Transformer (SCMT). The final gate module (FGM) is combined to suppress redundant information and modal interference.

[0031] (3) Relying on the multi-scale adaptive detection head, the adaptive feature fusion (AFF) and shallow feature fusion (SFF) strategies are adopted to achieve accurate complementarity between deep semantic features and shallow detail features, thereby improving the detection accuracy of small targets in power insulators.

[0032] (4) Model training and optimization are completed based on a self-built insulator dataset. The model can perform inference on low-light images and output accurate insulator location results. This invention effectively solves the problems of difficult feature extraction, large modal interference, and high false negative and false positive rates of small insulators in low-light scenarios, significantly improving detection accuracy and robustness, providing technical support for intelligent inspection of transmission lines, and has the advantages of high detection accuracy and strong environmental adaptability. Attached Figure Description

[0033] Figure 1 : Overall system flowchart of the method of the present invention.

[0034] Figure 2 : A schematic diagram of the overall architecture of the MFN-Net network proposed in this invention.

[0035] Figure 3 : Schematic diagram of the dual-modal fusion module (DMFT) proposed in this invention.

[0036] Figure 4 Visualization of the detection results of this invention and various advanced algorithms on a self-built dataset. Detailed Implementation

[0037] The technical solution of the present invention will be specifically described below through specific embodiments and in conjunction with the accompanying drawings.

[0038] A method for detecting small power insulators in low-light conditions based on cross-domain collaborative gating fusion, comprising the following steps: Step 1: Adjust the size of the input images for the model to 512×512 pixels. All data should be in JPG format. Input visible light images of power transmission line inspections that were synchronously acquired under low-light conditions. With infrared images .

[0039] Step 2: The preprocessed visible light image The input is a backbone network built on the CSPDarknet 53 structure. Through its cross-stage local connectivity structure, it extracts multi-level spatial detail features and outputs features. , , , : The CSPDarknet53 architecture is used for input visible light images. It employs a hierarchical convolutional architecture, progressively abstracting image information through multi-level feature extraction. The input image sequentially passes through convolutional blocks and cross-stage modules, with the spatial resolution halved at each level and the number of channels multiplied, forming a feature pyramid from edge texture to high-level semantics. Output... , , , This approach addresses the detection needs of objects at different scales, laying the foundation for subsequent feature fusion and target prediction in the network. The visible light feature extraction representation is as follows:

[0040] in , The input image has a height and width. Based on the number of channels, The channel spread factor is used, and the spatial resolution is calculated according to... Decreasing proportionally, number of channels according to Increasing exponentially; Step 3: Process the pre-processed infrared image The input is a feature extraction network constructed from cascaded HCT modules. Each HCT module, through parallel CWSA and AKAC modules, deeply mines the thermal radiation features of insulators in infrared images, and outputs feature extraction features. , , , : The HCT module, building upon the CWSA and FFN modules, introduces the AKAC module, which runs parallel to the CWSA module, to enhance local representation capabilities. A convolution is then introduced after the FFN module to adjust the number of channels, generating feature maps at different scales. Firstly... Output after LN standardization. The CWSA module is a highly efficient self-attention mechanism based on convolution, combining the local feature extraction capability of convolutional operations with the global context modeling capability of self-attention to establish relationships between targets and solve the problem of small targets being occluded. The AKAC module is a mechanism that dynamically adjusts the importance of different positions within the receptive field of the convolutional kernel. Unlike the fixed weights of traditional convolution, this module adaptively learns the weights of each position in the receptive field based on the input content, enhancing the edge and texture features of small targets, allowing the network to focus more on key regions for output features. The features processed by the parallel modules are added to the initial features to strengthen the initial information, resulting in the final feature set. The adjusted features are then normalized and input into the FFN module to further capture local and global contextual information, which is then compared with... Perform residual connection and output. Finally, the image features are resized using a 1×1 convolution to be used for spatial and channel feature extraction in the next stage of the deeper HCT module, and then combined with... Residual linking preserves shallow features and prevents information loss during layer-by-layer propagation. Finally, the extracted multi-level features are output. , , The infrared feature extraction is described as follows:

[0041] in, It was through Layer output, It was through and Output It was through and Output, For layer normalization; Step 4: Combine the visible light characteristics output from Step 2 and Step 3 , , , With infrared features , , , Features are obtained by splicing along the channel dimension , , , The data is then input into four identical DMFT modules. Each DMFT module consists of three parts; firstly, features are obtained through the Feature Interaction Module (SCMT). Enhanced features are obtained through the improved feedforward network (IFF) module. At the same time , Performing a concatenation operation yields interactively enhanced features. Finally, the FGM module enhances the extraction and representation of key features of the insulator through a parallel dual-path attention mechanism. , , , The DMFT module outputs features at four scales. , , , : The SCMT, IFF, and FGM modules in the DMFT module, such as Figure 3 As shown. , , , The SCMT module first processes the spliced ​​features LN layer normalization is performed to eliminate differences in numerical distribution between modalities, laying a stable foundation for subsequent calculations. Based on this, a positional encoding mechanism is introduced to inject learnable encoded information into each spatial location in the feature map, enhancing the model's understanding of feature spatial relationships and avoiding spatial misalignment of cross-modal features. Subsequently, 1×1 convolutions are used to obtain the values ​​belonging to the visible light and infrared modalities respectively. , , , , , To meet the needs of attention calculation, asymmetric bidirectional attention is then employed to obtain features resulting from the full interaction between visible and infrared light. , Finally, the interactive bimodal features are concatenated along the channel dimension, and the feature dimensions are adjusted using 1×1 convolutional projection. DropPath regularization is then applied to prevent overfitting and enhance generalization ability. The interactive features are then output. The feature fusion process formula is as follows:

[0042]

[0043]

[0044]

[0045] middle , Represents the normalized exponential function, express .

[0046] The IFF module first processes the concatenation features of the input. , , , A channel expansion operation is performed, doubling the number of feature channels through a 1×1 convolution. This operation provides ample dimensionality for subsequent feature refinement and extraction, while effectively avoiding information loss due to insufficient channel dimension. Subsequently, the module uses only the 3×3 depthwise convolution portion of the depthwise separable convolution for local feature enhancement. This design significantly reduces computational redundancy while accurately capturing the local spatial information of the insulator target, thereby significantly enhancing the local representation capability of the spliced ​​features. To further optimize feature quality, a GLU module is introduced. This unit filters the features output by depthwise convolution channel by channel, adaptively suppressing unimportant features such as background noise and redundant information, while enhancing key features related to the insulator target, thus improving the target-background discriminability of the features. Finally, the module passes through another 1×1 convolutional layer to restore the number of filtered feature channels to the initial dimension, outputting enhanced features. This step ensures that the output features match the input dimensions of subsequent processing modules, providing high-quality, dimension-consistent single-modal feature inputs for the alignment and fusion of bimodal features. The IFF calculation process is described below:

[0047] This represents depthwise separable convolution. Represents a gated linear unit. express .

[0048] The FGM module uses a parallel dual-path attention mechanism to process features. Enhancements are made. Channel gating branches generate channel weights through global pooling and shared full connection. Spatial gating branches generate spatial weights through channel pooling and convolution. Input features respectively with and Element-wise multiplication is used to recalibrate two-dimensional features, enhancing key insulator features and suppressing background interference. The module further introduces DropPath regularization, which improves the model's generalization ability in complex low-light scenes by randomly discarding output paths during training. The final output is optimized features enhanced by attention and regularization. , , , This process can be formally represented as:

[0049] Indicates channel attention weights. This represents the spatial attention weights. middle .

[0050] Step 5: Analyze the features output in Step 4. , , , Input is given to the multi-scale adaptive detection head module. First, the features are processed... , , After convolutional upsampling and efficient fusion and enhancement using the C3K2 module, features at three scales are output. , , This enables the network to better handle objects of different sizes and improves detection accuracy. Secondly, it enhances feature... First, the AFF module is used to... , , Feature map upsampling to Feature maps of the same size are obtained Then, the SFF module is used to extract the detailed shallow features. With semantically rich deep features Perform adaptive weighted fusion to obtain features This enhances the ability to characterize small insulators. The feature enhancement modules, AFF and SFF modules, in the multi-scale adaptive detection head. Input features. , , Multi-scale feature fusion and output are achieved through a bidirectional path composed of convolutional upsampling and C3K2. , , The AFF module first uses convolution operations and upsampling techniques to expand the size of the three deep semantic features to perfectly match the shallow detail features, solving the feature misalignment problem caused by size differences in traditional fusion. Building on this, an adaptive fusion mechanism is introduced, using a dynamic weight allocation strategy to fuse the matched deep and shallow features, fully mining both deep semantic and shallow detail information, avoiding information redundancy or loss in traditional fusion, and outputting the feature... .

[0051] The feature pyramid outputs three feature maps at different scales. , , In the middle, located at the pixel point ( , The eigenvectors at position ) are respectively , , The fusion formula for corresponding hierarchical features is as follows:

[0052] Here, , , The weights correspond to the three feature layers output by the pyramid module, and the following conditions must be met: , , , The range is [0,1].

[0053] The SFF module calculates the fused feature map through weighted summation. It introduces two learnable scalar parameters. and Then, the Sigmoid function is used to map it to the range of 0-1 to obtain the fusion weights. , shallow features and Multiplication will fuse the deep features. and Multiply them, then add them together to get the output feature. This overcomes the limitations of insufficient semantics from single shallow features or ambiguous spatial localization from single deep features. The implementation process is as follows:

[0054]

[0055]

[0056] for Activation function These are learnable deep weights. These are learnable shallow weights; Step Six: Extract the features output from Step Five. , , , Inputting data into detection heads of different target sizes, the bounding box regression and category classification prediction for small insulator targets are completed. Output features , , , The data are fed into four detection heads of different scales, and the entire network demonstrates significant performance in small target detection. The network integrates preliminary predictions from all scales and applies a non-maximum suppression algorithm to filter redundant detection boxes based on a preset confidence level of 0.25 and an intersection-over-union (IoU) threshold of 0.7. This process ultimately outputs the precise bounding box coordinates of each detected insulator, along with the corresponding confidence score, enabling efficient and robust automatic detection of small power insulators on transmission lines in complex low-light environments.

[0057] On both a self-built insulator dataset and a public dataset, the low-light transmission line small insulator detection method using a multimodal fusion network provided by this invention, compared with currently leading detection networks, demonstrates superior performance in detecting small targets under low light conditions, occlusion, and strong interference from similar objects. To verify the applicability of the method in specialized fields, tests were conducted on a self-built power insulator dataset, such as... Figure 4 As shown. In complex situations where there are background interference objects and the insulators are partially obscured by the transmission tower, see... Figure 4 Lines 1 to 3 show that this method achieves stable target detection. Particularly noteworthy is its ability to detect targets even in background interference environments with highly similar characteristics. Figure 4 In line 4, this method accurately identified all insulator targets without false detections, fully demonstrating its superior scene discrimination capability and detection reliability. Quantitative indicators show that the method's MAP and Precision reach 97.29% and 95.31% respectively, surpassing the comparative method MFFNet, which has a MAP of 96.62% and a Precision of 94.68%. This clearly demonstrates that this method has higher detection reliability and stability when dealing with small, occluded targets under strongly similar background interference.

[0058] The above is only one specific implementation method of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing the protection scope of the present invention.

Claims

1. A method for detecting low-light power small insulators based on cross-domain collaborative gating fusion, characterized in that, Includes the following steps: Step 1: Construct and train the MFN-Net network model to obtain the trained MFN-Net network model. The MFN-Net network model consists of three main parts: a bimodal feature extraction network, a cross-modal fusion network, and a multi-scale adaptive detection head. The bimodal feature extraction network includes a feature extraction structure built with CSPDarknet53 and a feature extraction structure built with cascaded HCT modules. The cross-modal fusion network includes four identical DMFT modules, which include SCMT, IFF, and FGM modules. The multi-scale adaptive detection head includes AFF, SFF modules, and detection heads for different target sizes. Step 2: Visible light images collected during transmission line inspections. With infrared images Preprocessing is required; Step 3: Process the pre-processed visible light image Inputting a backbone network constructed using CSPDarknet 53, the system extracts multi-level spatial detail features through the CSPDarknet 53 backbone network connection structure, and outputs visible light features at four scales. , , , ; , , , The scales are 160×160×128, 80×80×256, 40×40×512, and 20×20×1024, respectively, in pixels; Step 4: Process the pre-processed infrared image Input cascaded HCT modules; where each HCT The module operates in parallel. CWSA Module and AKAC module Deeply analyze the preprocessed infrared images Thermal radiation characteristics of medium insulators, outputting infrared characteristics at four scales , , , ; , , , The scales are 160×160×128, 80×80×256, 40×40×512, and 20×20×1024, respectively, in pixels; Step 5: Identify visible light characteristics , , , With infrared features , , , One-to-one correspondence and splicing along the channel dimension to obtain splicing features , , , splicing features , , , Input to four identical DMFT modules, outputting high-resolution features with the same size but double the number of channels. Shallow features Mid-layer characteristics Deep features ; Step Six , , , Input to the multi-scale adaptive detection head module: where, , , After convolutional upsampling and efficient fusion and enhancement using the C3K2 module, fused and enhanced features at three scales are output. , , The AFF module will , , Upsampling to The same size yields the feature map Then the SFF module will and Adaptive weighted fusion is performed to obtain weighted fusion features for small target detection. ; Step Seven , , , The data is input into detection heads with dimensions of 160×160×40, 80×80×256, 40×40×512, and 20×20×1024 respectively to complete the bounding box regression and category prediction of small target insulators.

2. The method for detecting low-light power small insulators based on cross-domain collaborative gating fusion as described in claim 1, characterized in that, In step one, the MFN-Net network model is trained until the composite loss function is minimized, which yields the trained MFN-Net network model. The formula for the composite loss function is as follows: ； Total loss function, The contribution of bounding box positioning accuracy to the total loss. The contribution of the accuracy of the control target classification to the total loss. The contribution of bounding box regression stability to the total loss is controlled. Bounding box loss, Classification loss, Distribution focus loss.

3. The method for detecting low-light power small insulators based on cross-domain collaborative gating fusion as described in claim 1, characterized in that, In step two, the preprocessing involves processing the visible light image. With infrared images The size has been adjusted to 512×512 pixels, and all data are in JPG format.

4. The method for detecting low-light power small insulators based on cross-domain collaborative gating fusion as described in claim 1, characterized in that, In step three, the feature extraction structure constructed by CSPDarknet53 is used for visible light images. By progressively abstracting image information through multi-level feature extraction, visible light images... Through convolutional blocks and cross-stage modules, the spatial resolution is halved at each stage, while the number of channels increases exponentially, forming a feature pyramid from edge texture to high-level semantics: ； This represents the i-th visible light feature. These are visible light images. Height and width, Based on the number of channels, The channel spread factor is used, and the spatial resolution is calculated according to... Decreasing proportionally, number of channels according to Increasing exponentially; Indicates the first The number of channels in the layer, R indicates that each element in the feature map is a real number, and CSPDarknet53 indicates the feature extraction structure constructed by CSPDarknet53.

5. The method for detecting low-light power small insulators based on cross-domain collaborative gating fusion as described in claim 1, characterized in that, The specific steps of step four are as follows: ； in, Representation layer normalization, It was through Layer output, It was through CWSA Branches and AKAC The output of the branch, It was through and Output, For layer normalization; Indicates the first One infrared feature, , This represents the i-th visible light feature.

6. The method for detecting low-light power small insulators based on cross-domain collaborative gating fusion as described in claim 1, characterized in that, In step five, the data processing method of the SCMT module in the DMFT module is as follows: ； ； ； ； in, Representing visible light and infrared light value, Represents visible light value, Indicating infrared value, Represents visible light value, Indicating infrared value, Represents visible light value, Represents visible light value; Represents the normalized exponential function, This represents DropPath regularization; Pi represents the i-th concatenated feature; This indicates the output of the infrared features after the interaction. This represents the output of the visible light features after the interaction, where T represents the transpose operation in the matrix. Represents a 1×1 convolution. This indicates the position encoding, and d represents the scaling factor. This indicates a splicing operation. Indicates the features after interaction; The data processing method of the IFF module is as follows: ； in, This represents depthwise separable convolution. Represents a gated linear unit. Indicates enhanced features; Will , Features are obtained by concatenating the components. ; The data processing method for the FGM module is as follows: ； Indicates channel attention weights. Indicates spatial attention weights; E This indicates features that have undergone interaction and enhancement; T i Representing features at different levels .

7. The method for detecting low-light power small insulators based on cross-domain collaborative gating fusion as described in claim 1, characterized in that, In step six, , , After passing through the enhanced PANet structure, multi-scale features are output and shallow features are fused via the bidirectional paths of FPN and PAN. , , : ； , , All are intermediate features of PANet. This refers to the C3K2 feature extraction module. Number of input channels Number of output channels This indicates a 2x upsampling operation. This indicates a 2x downsampling operation. Channel splicing operation; out=256 or 512 or 1024; in=768 or 1536; The AFF module first uses convolution operations and upsampling techniques to convert three shallow features... , , The size is expanded to perfectly match the shallow detail features, solving the feature misalignment problem caused by size differences in traditional fusion. Under this premise, an adaptive fusion mechanism is introduced, which fuses the matched features through a dynamic weight allocation strategy, fully mining deep semantic information and shallow detail information, avoiding information redundancy and loss problems in traditional fusion, and outputting semantically rich features. ; The feature pyramid outputs three feature maps at different scales. , , In the middle, located at the pixel point ( , The eigenvectors at position ) are respectively , , The fusion formula for corresponding hierarchical features is as follows: ； Here, , , The weights correspond to the three feature layers output by the pyramid module, and the following conditions must be met: , , , The range is [0,1]; The SFF module calculates the fused feature map through weighted summation; it introduces two learnable scalar parameters. and Then, the Sigmoid function is used to map it to the range of 0-1 to obtain the fusion weights. , Shallow features and Multiplication, fusion of deep features and Multiply them, then add them together to get the output feature. This overcomes the limitations of insufficient semantics from single shallow features or ambiguous spatial localization from single deep features. The implementation process is as follows: ； ； ； for Activation function These are learnable deep weights. These are learnable shallow weights.

8. The method for detecting low-light power small insulators based on cross-domain collaborative gating fusion as described in claim 1, characterized in that, In step seven, , , , The data is fed to four detection heads at different scales. Each detection head is connected to a neck network and is used to predict the target category and regress the bounding box of the input fused feature map. The output includes the target category, confidence score, and location information. The detection head adopts a multi-scale structure, independently predicting feature maps with resolutions of 160×160×40, 80×80×256, 40×40×512, and 20×20×1024. The preliminary prediction results from each scale are then integrated, and a non-maximum suppression algorithm is applied to filter redundant detection boxes based on a set confidence score of 0.25 and an intersection-over-union (IoU) threshold of 0.

7. The precise bounding box coordinates of each detected insulator and the corresponding confidence score are output, thus completing the bounding box regression and category classification prediction for the small target insulator.

Images