Lightweight detection method for foreign matter of power transmission line

Abstract

This invention discloses a lightweight method for detecting foreign objects (FOOs) in power transmission lines, relating to the fields of power line inspection and computer vision. A YOLOv10-based FEO detection model is constructed, comprising a feature extraction part, a feature interaction part, and a detection head part. The C2f module in the feature extraction and feature interaction parts is replaced with an expert dynamic extraction module, and the detection module in the detection head part is replaced with a parameter-sharing detection module. The trained FEO detection model is obtained by collecting and evaluating a calibrated FEO dataset of power transmission lines. FEO keyframes are input into the trained FEO detection model to obtain detection results. The expert dynamic extraction module acquires feature maps from multiple receptive fields, achieving rich representation in FEO detection environments with varying scales. The parameter-sharing detection module, based on this, reduces redundant computation and enhances the semantic representation of small targets through parameter sharing and grouping normalization, effectively improving the efficiency of FEO detection in power transmission lines.

Description

Technical Field

[0001] This invention relates to the field of power line inspection, specifically to a lightweight detection method for foreign objects in power transmission lines based on YOLOv10. Background Technology

[0002] Power transmission lines are the arteries of the power system, and their safe and stable operation is crucial to the national economy and people's livelihood. Because power transmission lines are exposed to the natural environment for extended periods, they are highly susceptible to being smothered, suspended, or entangled by external objects such as kites, bird nests, and agricultural film, leading to serious safety accidents such as short circuits and power outages. In recent years, with the popularization of drone technology, using drones equipped with high-definition cameras for power line inspection has become a mainstream trend. In terms of image processing for drone inspections, deep learning-based target detection algorithms (such as the YOLO series) are widely used due to their feature extraction capabilities. However, for the specific application scenario of power transmission lines, existing target detection algorithms still have the following significant shortcomings in practical deployment: Firstly, there is a contradiction between the computational complexity of the model and the limited computing power of the drone's edge. Existing high-precision target detection models typically rely on extremely deep network structures and multi-branch independent detection heads to handle multi-scale targets. This design results in a large number of model parameters and high computational redundancy. However, the edge computing devices of power inspection drones are strictly limited in terms of computing power and power consumption, making it difficult to meet the high frame rate real-time inference requirements of complex models. Forcibly compressing the model size can lead to feature loss, severely affecting detection accuracy. On the other hand, the complex outdoor background and extreme scale variations lead to serious problems of missed and false detections of small targets. Transmission line images have complex backgrounds, and different types of foreign objects vary greatly in scale. Traditional feature extraction networks often use static convolutional kernels with fixed receptive fields, lacking the ability to adaptively perceive input image features. This not only makes them susceptible to interference from invalid information in cluttered backgrounds during feature extraction, but also easily dilutes the semantic information of small targets during cross-scale feature fusion, resulting in poor robustness in detection under complex conditions.

[0003] Therefore, those skilled in the art are dedicated to developing a lightweight detection method for foreign objects in power transmission lines. Summary of the Invention

[0004] To address the technical problems of existing foreign object detection models for power transmission lines, such as large number of parameters, difficulty in real-time reasoning at the edge device end, and easy omission or false detection of small target foreign objects in complex field backgrounds, this invention provides a lightweight detection method for foreign objects in power transmission lines.

[0005] To achieve the above objectives, the present invention provides a lightweight detection method for foreign objects in power transmission lines. The method is characterized by comprising:

[0006] Step 1: Acquire and preprocess the image to be detected; Step 2: Input the preprocessed image into a pre-trained foreign object detection model, and perform multi-scale feature extraction through the expert dynamic extraction module in the foreign object detection model; wherein, the expert dynamic extraction module performs feature enhancement on some channels by splitting the channels and using dynamic convolution kernels; Step 3: Predict the extracted multi-scale features using a parameter-shared detection head, wherein the parameter-shared detection head shares convolution weights between output layers at different scales and uses a scale adaptive factor to compensate for feature differences at different scales; Step 4: Determine the location, category, and confidence level of the foreign object in the transmission line based on the output of the parameter-shared detection head.

[0007] To address interference from complex outdoor backgrounds, this invention's expert dynamic extraction module introduces an input-adaptive dynamic convolution kernel generation mechanism. This mechanism dynamically adjusts the attention weight coefficients of different expert convolution kernels based on the global context of the input image, enabling the network to actively suppress background noise feature responses and focus on the high-frequency features of subtle foreign objects, effectively solving the representation bottleneck of traditional static convolution in complex scenes. A parameter-sharing detection head is employed. By forcibly sharing convolution weights among detection branches at different scales, parameter redundancy in multi-branch structures is effectively eliminated. Simultaneously, in conjunction with the channel splitting mechanism of the expert dynamic extraction module in the backbone network, complex operations are performed only on a subset of channels. The synergistic effect of these two mechanisms significantly reduces the total number of model parameters and floating-point operations, perfectly meeting the deployment requirements of computationally limited platforms such as power line inspection drones. This invention effectively improves the accuracy and robustness of foreign object detection on transmission lines in complex backgrounds, exhibiting high applicability and real-time performance.

[0008] Technical effect

[0009] The present invention provides a lightweight detection method for foreign objects in power transmission lines. The method is reasonably designed and highly practical, and can effectively and efficiently detect and identify foreign objects in the complex environment of existing power transmission lines.

[0010] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0011] Figure 1 This is a flowchart of the foreign object detection method for power transmission lines based on YOLOv10 of the present invention;

[0012] Figure 2 This is a network structure diagram of the foreign object detection model for power transmission lines based on YOLOv10 in this invention;

[0013] Figure 3 This is a schematic diagram of the expert dynamic extraction module structure of the present invention;

[0014] Figure 4 This is a schematic diagram of the parameter sharing detection module structure of the present invention. Detailed Implementation

[0015] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0016] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0017] Specific implementation method one: Combining Figures 1 to 4 This embodiment describes the following steps:

[0018] Step 1: Acquire and preprocess the image to be detected;

[0019] Step 2: Input the preprocessed image into the pre-trained foreign object detection model, and perform multi-scale feature extraction through the expert dynamic extraction module in the foreign object detection model; wherein, the expert dynamic extraction module performs feature enhancement on some channels by splitting the channels and using dynamic convolution kernels;

[0020] Step 3: Use a parameter-sharing detection head to predict the extracted multi-scale features. The parameter-sharing detection head shares convolution weights between output layers at different scales and uses a scale adaptation factor to compensate for feature differences at different scales.

[0021] Step 4: Determine the location, type, and confidence level of foreign objects in the transmission line based on the output of the parameter sharing detection head.

[0022] Specific Implementation Method Two: The preprocessing method of Specific Implementation Method One includes the following steps:

[0023] Step 2.1: Perform bilinear interpolation downsampling on the acquired original transmission line image to uniformly map it to the standard input dimension (640×640 pixels) of the foreign object detection model. For each target pixel in the 640×640 output image, use it to map back to the original high-resolution image. The weighted average of the four surrounding adjacent pixels is used to fill the gap.

[0024] Define a pixel in the target downsampled image as... The corresponding original transmission line image The floating-point coordinates on are In the original image Find it on The four nearest integer quantized pixels in terms of geometric distance are: top left corner Top right corner Bottom left corner and bottom right corner In this context, each quadrant point is defined as follows: The initial pixel grayscale value at is Perform linear interpolation along the x-axis. First, determine the target point. The coordinates are relative to the two adjacent integer quantized pixels. and The relative horizontal distance ratio is used as the weight. This weight is then used to... and ,as well as and Perform horizontal linear fusion to obtain two intermediate transition pixels. and :

[0025]

[0026]

[0027] Then, using the target point coordinates relative to and The vertical distance ratio for the two intermediate transition pixels mentioned above. and Perform the final linear fusion. This yields the pixel values ​​in the target downsampled image. final pixel value :

[0028]

[0029] By employing the aforementioned two-stage linear superposition formula, each pixel in the 640×640 pixel target downsampled image incorporates the grayscale information of the four pixels surrounding its corresponding spatial location in the original high-definition inspection image. This method effectively alleviates the image jaggedness caused by traditional nearest-neighbor interpolation, not only reducing the computational load on model inference at the hardware level but also preserving the edge topological details of foreign objects in the original high-definition image of the transmission line to the greatest extent possible. This lays a data foundation for the expert dynamic extraction network to efficiently capture the representational features of foreign objects under multiple receptive fields in subsequent steps.

[0030] Step 2.2: Perform channel-by-channel normalization on the downsampled image. Let the image be... One channel Middle position The pixel value at that location is Its normalization formula is:

[0031]

[0032] In the formula, These are the normalized pixel values; and These are the statistical results obtained from the foreign object dataset of transmission lines. The mean and standard deviation of the channels. This step maps pixel values ​​to a standard normal distribution, effectively reducing the interference of extreme inspection conditions such as backlight and strong light on foreign object features.

[0033] Step 2.3: Simultaneously transform the image containing the foreign object and its corresponding ground truth bounding box using a random rotation matrix. Let the pixel coordinates of a certain foreign object in the original image be... The center of rotation is The rotation angle is Then the transformed coordinates The following affine transformation formula applies:

[0034]

[0035] Specific Implementation Method Three: The expert dynamic extraction module of Specific Implementation Method One includes the following parts:

[0036] Step 3.1: Let the initial multi-scale feature map input to the dynamic extraction module be... ( For the number of channels, and (For spatial resolution). The feature map first passes through the first batch normalization layer (BN1). Considering that UAVs often face extreme lighting changes such as direct sunlight and local shadows during field inspections, leading to drastic shifts in the internal covariance of the input feature map, BN1 aims to physically decouple the lighting and contrast of the feature channels. Let the mini-batch mean be... The variance is The first normalized output The formula is:

[0037]

[0038] In the formula, and The scale factor and translation parameters are learned by the network during training. To prevent the use of tiny constants with a denominator of zero, normalization was applied to suppress anomalous activation values ​​caused by high reflectivity of transmission line insulators or tree shadows. Subsequently, the normalized characteristics were... The input is fed into the dynamic receptive field module to extract adaptive features specific to the foreign object. To prevent the loss of the rigid topology of power transmission towers and conductors in the original image during deep semantic extraction, the module introduces a first residual connection. This first residual connection combines the dynamically extracted output with the original, unprocessed input feature map. By performing element-wise addition, the first-level fusion feature is obtained. :

[0039]

[0040] Step 3.2: To further integrate the superimposed features at the channel dimension and eliminate the semantic gap between foreign object features and background features, the first-level fusion features... Proceed to the second level of processing. First, features... After a second batch normalization layer (BN2) for re-centering and scaling, the output is... .then, The input is sent to the channel interaction module. This channel interaction module consists of a single 3×3 standard convolutional kernel and a Rectified Linear Activation Function (ReLU). Let the weight matrix of the 3×3 convolutional layer be... The bias term is Then the nonlinear channel interaction output at this stage The calculation formula is:

[0041]

[0042] In the formula, This represents a two-dimensional cross-correlation convolution operation. This represents the ReLU activation operation. Its core mechanism utilizes a 3×3 receptive field to capture the local geometric texture of foreign objects on the transmission line, and through the truncation characteristic of ReLU, filters out negative feedback noise irrelevant to the detection task, achieving nonlinear sharpening of feature boundaries. Finally, through a second residual connection, the locally interactively enhanced features are... Compared with the previous level feature map Perform element-wise addition again to output the final feature map of the current dynamic extraction module. :

[0043]

[0044] Specific implementation method four: Combination Figure 3 and Figure 4 This embodiment describes the dynamic receptive field module of specific embodiment three, which includes the following parts:

[0045] Step 4.1: Let the initial multi-scale feature map input to the dynamic extraction module be... ( For the number of channels, and (for spatial resolution). The channel segmentation unit first proceeds along the channel dimension. Input tensor The first sub-feature map is divided into equal parts. With the second sub-feature map First sub-feature map The data is fed into subsequent dynamic branches for enhanced extraction of foreign object sensitive features; while the second sub-feature map... This is then fully preserved as an identity mapping. The floating-point computation time for dynamic convolution is halved, while preserving... This maintains the low-frequency, rigid global spatial topology of transmission lines, insulator strings, etc., and prevents the loss of physical location information caused by excessive nonlinear transformation.

[0046] Step 4.2: For the first sub-feature map Global average pooling is used to extract the global context vector. .vector The Middle The specific calculation formula for each channel is as follows:

[0047]

[0048] Step 4.3: Transfer the context vector The input is fed into a dynamic routing network consisting of fully connected layers and nonlinear activation function mappings. Assume the system is pre-configured... If there are several expert candidate convolutional kernels (3×3 and 5×5 convolutional kernels) with different receptive field characteristics, then the network is the first... Dynamic kernel weight coefficients generated from candidate convolutional kernels Normalization calculation is performed using the Softmax function:

[0049]

[0050] In the formula, For transformation functions containing fully connected mappings, the normalization operation ensures... Dynamic weighting coefficients This indicates which receptive field scale the model depends on most under the current specific transmission line context.

[0051] Step 4.4: Based on the dynamic kernel weight coefficients calculated above. Some dynamic convolutional units, within the parameter space, are based on preset parameters. candidate convolutional kernels Perform linear weighted summation to synthesize a target dynamic convolution kernel that highly matches the characteristics of the current image in real time. :

[0052]

[0053] Dynamic convolution kernel for the target For the first sub-feature map Performing convolution operations yields an output with significantly enhanced foreign object features. :

[0054]

[0055] Step 4.5: The feature fusion unit will process the output enhanced by the dynamic target convolution kernel and rich in high-frequency foreign object features. Compared with the initially preserved second sub-feature map rich in low-frequency background structural features. Perform splicing operations at the channel level:

[0056]

[0057] The final output feature map after splicing The initial channel dimension C was perfectly restored. This closed-loop design, which combines partial processing with backend splicing, ensures an extremely lightweight transmission line foreign object detection network while achieving a highly efficient unification of local salient features and global structural features.

[0058] Specific Implementation Method Five: The parameter-sharing detection head and scale adaptation factor of Specific Implementation Method One include the following parts:

[0059] Step 5.1: After the feature map passes through the feature extraction part and the feature interaction part in sequence, the output is a multi-scale feature map. First, each feature map enters its own independent channel adjustment layer. To eliminate statistical differences in feature distributions at different scales before parameter sharing, a grouped normalization layer is used. Let the number of feature map channels be C, and it is divided into G groups (each group contains C / G channels). For the pixel features within the g-th group... Its normalized calculation formula is:

[0060]

[0061] in, and The mean and standard deviation are calculated within the pixel group. The underlying principle is that, unlike batch normalization (BN), GN calculates statistics within the channel group, thus avoiding the limitation of single-batch input during UAV-side inference. This grouped statistical approach enhances the model's robustness to the semantic representation of foreign objects at different scales against the background of power transmission lines.

[0062] Step 5.2, as follows Figure 2 As shown in the middle section, this invention reconstructs the originally independent detection branches into weight-shared convolutional layers. Let... This is a pre-defined set of shared convolutional kernel parameters. For input feature maps of three different resolutions (P3, P4, and P5), the following convolutional mapping is uniformly performed:

[0063]

[0064] This mechanism directly reduces the number of parameters in the detection head to one-third of that in a traditional standalone detection head. The model no longer learns feature maps separately for each scale, significantly improving its running speed on embedded inspection terminals.

[0065] Since P3 corresponds to shallow, high-resolution features, while P5 corresponds to deep, low-resolution features, forcibly sharing weights would lead to an imbalance in the output response. Therefore, this embodiment introduces a learnable scaling factor at the output of each detection branch. For the output response at the i-th scale The calculation formula is as follows:

[0066]

[0067] in, Scale the parameters for three independent branches. During training, these parameters are automatically optimized along with the loss function. By adjusting... The model can automatically compensate for the feature strength bias caused by weight sharing, ensuring that the confidence of the predicted box and the regression amplitude are on the same baseline, whether in a high-frequency small target area or a low-frequency large target area.

[0068] Step 5.3, as follows Figure 2 As shown on the right, the feature output at each scale is further divided into two parallel weight-sharing branches: the category prediction branch and the regression prediction branch. The prediction branch... Used to determine the type of foreign object, regression branch Used to determine the bounding box coordinates (x, y, w, h) of the foreign object.

[0069] Specific Implementation Method Six: The training process of the pre-trained foreign object detection model in Specific Implementation Method One aims to optimize network parameters so that the model can accurately fit the visual features of foreign objects in the background of power transmission lines. The specific training principle and loss function calculation process are as follows: First, construct the joint loss function. This is used to guide the model to achieve a balance between classification accuracy and localization accuracy. The joint loss function is expressed as:

[0070]

[0071] In the formula, , , They are respectively preset weight balance coefficients, which are used to adjust the contribution of different tasks to the model convergence. When the sample is a positive sample, the classification loss Its calculation method is:

[0072]

[0073] When the sample is a negative sample, the classification loss Its calculation method is:

[0074]

[0075] The bounding box regression loss The calculation formula is:

[0076]

[0077] In the formula, is the overlap area ratio between the predicted box and the ground truth box, is the distance between the center points, is the diagonal length of the smallest closed region containing the two boxes, is a parameter to measure the aspect ratio consistency. This formula ensures that the model can still maintain the stability of bounding box regression when the scale of foreign objects fluctuates violently, especially for slender foreign objects such as kite strings, and has better positioning effects. is the distribution focus loss. Its principle is to model the position of the bounding box as a general probability distribution rather than a single value. By optimizing the cross-entropy between the predicted distribution and the true distribution, the model can more accurately locate the boundary when the edge of the foreign object is blurred. Finally, the total loss value is calculated based on the joint loss function, and the network parameters of the expert dynamic extraction module and the parameter-sharing detection head are continuously iteratively updated through the backpropagation algorithm until the loss function tends to be stable and meets the convergence conditions.

[0078] Specific Embodiment 7. The final result determined in Specific Embodiment 1 is achieved by performing logical screening and spatial redundancy removal on the output of the parameter-sharing detection head.

[0079] Step 7.3. Perform confidence filtering on all candidate prediction boxes output. Let the classification score of the i-th prediction box be Si, and the preset confidence threshold be T. If Si < T is satisfied, then this prediction box is regarded as environmental noise interference and is removed. This step eliminates a large number of invalid prediction results generated by background false alarms through preliminary screening.

[0080] Step 7.2. Perform non-maximum suppression operation. This operation aims to solve the problem that the same foreign object target is repeatedly covered by multiple detection windows due to the parameter-sharing detection head outputting at multiple scales (P3, P4, P5) simultaneously. The specific calculation logic is as follows:

[0081] Step 7.2.1: Sort all remaining candidate prediction boxes in descending order of confidence score;

[0082] Step 7.2.2: Select the prediction box with the highest confidence level. This serves as a baseline result and is stored in the final set of foreign object locations.

[0083] Step 7.2.3: Calculate the remaining prediction boxes With baseline prediction box Spatial overlap between them, i.e., intersection-union ratio Value. Its calculation formula is:

[0084]

[0085] In the formula, The area of ​​the intersection of the two frames. This represents the area of ​​the merged region between the two frames.

[0086] Step 7.2.4: The preset overlap threshold is Tiou. If the calculated... >Tiou, then determine To prevent repeated predictions of the same foreign object target, it is completely removed from the candidate prediction box set.

[0087] Step 7.2.5: Repeat the above sorting and comparison steps until all predicted boxes in the candidate set have been traversed.

[0088] Through the above steps, the present invention can filter out the unique detection box with the highest confidence from the dense prediction output, and finally output a detection report reflecting the accurate location, type and confidence of foreign objects on the transmission line, effectively solving the problem of multiple box overlap in complex working conditions of lightweight models.

[0089] In the eighth specific implementation method, in order to support the application of the above-mentioned lightweight detection method for foreign objects on power transmission lines in real power inspection scenarios, this invention provides corresponding computer equipment and computer-readable storage media, and designs a tensor reconstruction deployment process adapted to the underlying hardware.

[0090] Step 8.1: This embodiment provides a computer device, which is mainly deployed in the onboard computing platform of a power line inspection drone. The computer device includes a processor and a memory. The memory stores the computer program, and its internal structure includes high-speed random access memory and non-volatile flash memory chips to ensure stable program residency in complex electromagnetic environments. The processor employs a tensor operation core with high parallel computing capabilities or an embedded microprocessor. When the processor executes the computer program stored in the memory, it can fully schedule and implement all the steps of the lightweight detection method for foreign objects on power transmission lines as described in any one of claims 1 to 8 of this invention.

[0091] Step 8.2: This embodiment provides a computer-readable storage medium. The compiled instruction set of the aforementioned foreign object detection algorithm is physically embedded within this medium. When this computer program is read and executed by the processor of an edge computing device, the aforementioned model inference and post-processing flow can also be implemented.

[0092] Step 8.3: The trained dynamic extraction network and parameter-sharing detection head are exported in an open neural network exchange format. Then, an edge computing device-specific tensor inference engine is used for operator fusion and precision quantization to generate a serialization engine file adapted to the underlying hardware. During inference prediction, since the output of the parameter-sharing detection head needs to interact with the post-processing algorithm in memory, this embodiment performs deep memory optimization on the tensor dimension of the network output layer. After the backbone and detection head are calculated, the initial output result of the model is a three-dimensional prediction tensor. Let the batch size be B, the total number of prediction boxes be N, and the sum of the number of detected target categories and the coordinate dimensions of the bounding boxes be A (specifically, the sum of the number of categories and the four coordinate offsets). The dimension arrangement order of the initial output tensor corresponds to the spatial memory index sequence defined as [0,1,2], and its physical shape is [B,A,N]. To accelerate the memory addressing efficiency of the non-maximum suppression algorithm and avoid cache misses caused by step-wise reads, the processor calls a memory rearrangement instruction at the output end to perform a dimension transpose operation on the initial output tensor. The process involves keeping the zero-dimensional index of the batch size unchanged, and swapping the first-dimensional index representing the attribute channel with the second-dimensional index representing the number of prediction boxes. After the transpose operation, the dimensional index sequence of the output tensor is transformed to [0,2,1]. The final reconstructed output tensor is strictly reshaped into a [B,N,A] structure. This transpose operation ensures that all classification confidence scores and coordinate parameters of the same prediction box exhibit a strictly continuous distribution in the underlying physical memory. When the processor traverses 8,400 candidate boxes for threshold filtering and overlap removal, it can maximize the use of the processor's single-instruction multiple-data-stream prefetch mechanism, significantly reducing the memory access latency of the edge computing platform, thereby establishing a real-time advantage for foreign object detection in power transmission lines at the hardware level.

[0093] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A lightweight detection method for foreign objects in power transmission lines, characterized in that, The method includes: Step 1: Acquire and preprocess the image to be detected; Step 2: Input the preprocessed image into the pre-trained foreign object detection model, and perform multi-scale feature extraction through the expert dynamic extraction module in the foreign object detection model; wherein, the expert dynamic extraction module performs feature enhancement on some channels by splitting the channels and using dynamic convolution kernels; Step 3: Use a parameter-sharing detection head to predict the extracted multi-scale features. The parameter-sharing detection head shares convolution weights between output layers at different scales and uses a scale adaptation factor to compensate for feature differences at different scales. Step 4: Determine the location, type, and confidence level of foreign objects in the transmission line based on the output of the parameter sharing detection head.

2. The lightweight detection method for foreign objects in transmission lines according to claim 1, characterized in that, Step 1 includes first performing multi-scale downsampling and normalization processing on the acquired original image to construct a test sequence that conforms to the input dimension of the foreign object detection model, and then using random rotation to obtain foreign object images of transmission lines at different spatial locations.

3. The lightweight detection method for foreign objects in transmission lines according to claim 1, characterized in that, The dynamic extraction module adopts a residual structure, specifically: before entering the channel segmentation unit, the feature map first passes through the first batch normalization layer (BN), the dynamic receptive field module, and the first residual connection in sequence. The first residual connection adds the dynamically fused output to the original input element by element. The output of the first residual connection passes through a second batch normalization (BN) layer, then through a channel interaction module consisting of a 3×3 convolutional layer and a linear rectified activation function (ReLU), and finally is fused with the previous level feature map through a second residual connection.

4. The lightweight detection method for foreign objects in transmission lines according to claim 3, characterized in that, The dynamic receptive field module includes: a channel segmentation unit, used to segment the input feature map along the channel dimension into a first sub-feature map and a second sub-feature map; a dynamic weight generation branch, used to extract a context vector from the first sub-feature map through global average pooling, and to obtain a set of dynamic kernel weight coefficients through a fully connected layer; a partial dynamic convolution unit, used to perform linear weighted fusion of multiple preset candidate convolution kernels based on the dynamic kernel weight coefficients to generate a target dynamic convolution kernel, and to use the target dynamic convolution kernel to extract features from the first sub-feature map; and a feature fusion unit, used to concatenate the first sub-feature map after feature extraction with the second sub-feature map to restore the channel dimension.

5. The lightweight detection method for foreign objects in transmission lines according to claim 1, characterized in that, The parameter-sharing detection head in step 3 includes: a weight-sharing convolutional layer, set in the detection branches for feature maps of different scales (P3, P4, P5), and all detection branches share the same set of convolutional kernel parameters; a scale compensation unit, which configures a learnable scale scaling factor for each detection branch, used to align the feature responses of different scales after weight sharing; and a grouping normalization layer, used to enhance the robustness of the model to the semantic representation of foreign objects at different scales by grouping and statistically analyzing the feature maps based on parameter sharing.

6. The lightweight detection method for foreign objects in transmission lines according to claim 5, characterized in that, The parameter-sharing detection head adopts a decoupled structure, that is, for the feature input at each scale, class prediction and regression prediction are performed through two parallel weight-sharing branches to eliminate feature conflicts between the classification task and the localization task.

7. The lightweight detection method for foreign objects in transmission lines according to claim 1, characterized in that, The pre-trained foreign object detection model is trained as follows: a training sample set labeled with ground truth bounding boxes and foreign object categories is obtained; the training sample set is input into the initial foreign object detection model, which then passes through the expert dynamic extraction module and the parameter-shared detection head to output prediction results; a joint loss function is constructed, which includes a classification loss for calculating category differences, a bounding box regression loss for evaluating the overlap between the predicted and ground truth boxes, and a distribution focusing loss for optimizing bounding box details; the total loss value is calculated based on the joint loss function, and the network parameters of the expert dynamic extraction module and the parameter-shared detection head are updated using the backpropagation algorithm until the convergence condition is met.

8. The lightweight detection method for foreign objects in transmission lines according to claim 1, characterized in that, Step 4, which determines the final result based on the output of the parameter-sharing detection head, specifically includes: extracting all candidate prediction boxes and their corresponding confidence scores from the output of the parameter-sharing detection head; removing candidate prediction boxes with confidence scores below a preset threshold; using a non-maximum suppression algorithm to filter the remaining candidate prediction boxes, eliminating redundant prediction boxes for the same foreign object target, and outputting the final location of the foreign object on the transmission line.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the lightweight detection method for foreign objects in transmission lines as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the lightweight detection method for foreign objects on transmission lines as described in any one of claims 1 to 8.

Images