Power tower component detection method for unmanned aerial vehicle scene

Abstract

The invention discloses a power tower component detection method for an unmanned aerial vehicle scene. According to the method, a target detection model comprising a self-adaptive multi-receptive field module, a hierarchical wavelet interaction unit and a distributed perception confidence optimization head is constructed, so that key technical problems in unmanned aerial vehicle power inspection are solved. In the feature extraction stage, a parallel multi-expansion deep convolution and channel attention mechanism is adopted, and the multi-scale feature perception capability is enhanced; in the feature fusion stage, semantic alignment and edge detail retention are realized through wavelet transform decomposition and cross-scale channel interaction; in the prediction stage, the classification confidence coefficient is adaptively calibrated based on statistical analysis of bounding box distribution, and the consistency of the positioning precision and the classification score is improved. According to the method, the problems of missing detection and false detection caused by large component scale difference and complex background are effectively solved, the accuracy and reliability of a detection result are remarkably improved, and an effective technical support is provided for unmanned aerial vehicle power inspection.

Description

Technical Field

[0001] This invention relates to a method for detecting power pole components in drone scenarios, belonging to the field of computer vision. Background Technology

[0002] As the core energy source for modern society, the stable supply of electricity directly depends on the safe operation of high-voltage transmission lines. As critical infrastructure of these lines, power poles are constantly exposed to complex natural environments, requiring regular inspection and maintenance to ensure grid reliability. Traditional manual inspection methods suffer from inherent drawbacks such as high labor intensity, low inspection efficiency, and high safety risks, especially in complex terrains like mountainous and forested areas where manual inspections are even more challenging. In recent years, with the rapid development of drone technology, drone-based power inspection solutions have demonstrated significant advantages. Drones, with their flexible mobility, wide monitoring perspective, and efficient operating modes, provide a completely new technological means for power inspection. This inspection method not only greatly improves operational efficiency but also effectively reduces personnel safety risks, becoming an important development direction for power facility inspection.

[0003] In UAV power line inspection missions, accurately identifying various key components on power poles, including insulators, fittings, and bolts, is a fundamental step in assessing equipment condition. Compared to overall pole identification, component-level inspection faces more severe technical challenges: First, components vary significantly in size, ranging from several-meter-long insulator strings to centimeter-sized fasteners coexisting in the same scene, placing extremely high demands on the multi-scale feature perception capabilities of the inspection model; second, components have diverse shapes, making them susceptible to changes in angle, partial obstruction, and lighting conditions from an aerial perspective; furthermore, complex environmental backgrounds such as mountains, fields, and buildings, as well as weather changes, further increase the difficulty of accurate inspection. These factors collectively constitute the technical bottleneck of power component inspection.

[0004] Currently, deep learning-based object detection methods have been widely applied in the field of power component detection. Researchers have proposed various improvement strategies, including introducing attention mechanisms to enhance feature representation capabilities, using feature pyramid networks to fuse multi-scale information, and optimizing detection performance by improving network structure and loss functions. However, these methods still have significant shortcomings in several key aspects: at the feature extraction level, the receptive field design of traditional convolutional neural networks is relatively fixed, making it difficult to adapt to the feature learning needs of components at different scales. This limitation leads to the model's inability to simultaneously and effectively capture significantly different target features such as large insulators and micro bolts, resulting in a high false negative rate for small-scale components, while the feature representation of large-scale components is also insufficient. Tan et al. used directional object detection and deformable convolutional structures to achieve high-precision identification of transmission towers in remote sensing images. Li et al. proposed a cross-scale spatial attention detector, which effectively identifies minute components in the mechanical connections of transmission lines, but it does not involve adaptive adjustment of the receptive field. In terms of feature fusion, existing multi-scale fusion methods mostly use simple upsampling and feature concatenation operations, lacking in-depth mining of semantic relationships between cross-scale features. This fusion method is prone to feature mismatch problems in complex scenarios, resulting in the loss of key details during transmission, which directly affects the ability to locate the precise boundaries of components.

[0005] At the prediction output level, there is a significant mismatch between classification confidence and localization quality. Existing methods lack an accurate evaluation mechanism for the quality of predicted bounding boxes, leading to situations where correct detections with accurate localization but low confidence are filtered out, while incorrect detections with inaccurate localization but high confidence are retained. This severely impacts the reliability of the detection results. Wang et al. proposed a dual prediction head structure and coordinate attention mechanism to further optimize the accuracy and efficiency of small target detection. Furthermore, existing improvement schemes often focus on optimizing single technical aspects, lacking a systematic solution. While improving the performance of one aspect, most methods may negatively impact other aspects or significantly increase computational complexity, making it difficult to achieve ideal overall results in practical inspection scenarios. These problems restrict further improvement in the efficiency of power line inspection, necessitating an innovative technical solution that can systematically optimize the entire process from feature extraction to result prediction.

[0006] In summary, current methods for detecting power pole components still have significant shortcomings in the three core stages of feature extraction, feature fusion, and detection prediction: First, the receptive field structure design in the feature extraction stage is simplistic, lacking multi-scale feature perception capabilities and failing to adapt to the vast scale differences of power pole components. Second, the feature fusion stage lacks a semantic alignment mechanism for cross-scale features, resulting in insufficient detail preservation and decreased feature discrimination in complex backgrounds. Third, the detection prediction stage fails to effectively align classification confidence with localization accuracy and lacks accurate evaluation of the predicted bounding box quality, affecting the reliability of the detection results. Therefore, there is an urgent need for a precise detection method for power pole components in UAV scenarios, which can solve the above-mentioned technical challenges through systematic structural innovation and meet the practical application needs of automated power inspection. Summary of the Invention

[0007] The purpose of this invention is to provide a method for detecting power pole components in unmanned aerial vehicle (UAV) scenarios, comprising the following steps:

[0008] This invention is implemented using the following scheme:

[0009] S1: Obtain a dataset of drone aerial images of power poles;

[0010] S2: Preprocess the power pole and tower dataset;

[0011] S3: Establish a target detection model for power pole components in a drone scenario;

[0012] S3.1: In the feature extraction stage, an adaptive multi-receptive field module is adopted. This module performs multi-scale feature extraction through its internal parallel multi-dilation depth convolution sub-module. This sub-module covers receptive fields of different scales on the same channel dimension through multiple pre-set convolution branches with fixed dilation rates, so as to enhance the ability to model the context of the target. Subsequently, the feature is reorganized by combining the channel attention sub-module and the dimension transformation sub-module, and with the help of residual connections, the enhanced multi-scale features are output.

[0013] S3.2: In the feature fusion stage, a hierarchical wavelet interaction unit is used to perform wavelet transform decomposition, detail enhancement, and channel interaction on features at adjacent scales. First, wavelet transform decomposition is performed on large-scale features to obtain low-frequency and high-frequency components in the horizontal, vertical, and diagonal directions. High-frequency details are enhanced by fusion, and low-frequency components are interactively compressed with small-scale features in the channel dimension. Then, the fused features are reconstructed into a unified scale through inverse wavelet transform, thereby achieving semantic alignment and preserving edge details in complex backgrounds.

[0014] S3.3: In the prediction stage, a distribution-aware confidence optimization head is used. This module performs statistical analysis on the distribution of the predicted bounding boxes, adjusts the initial classification score based on the Top-k probabilities and their mean statistical features, and thus optimizes the confidence of the target. This process uses a multilayer perceptron to process the statistical features, generates the optimized confidence, and combines the bounding box coordinate distribution features to output the final detection box and class probability.

[0015] S4: Train the model using the training set of the dataset to obtain the optimal model;

[0016] S5: Use the optimal model to test the test set images of the dataset to obtain the final test results.

[0017] Furthermore, the InsPLAD remote sensing dataset is cropped into 640x640 image blocks without sliding window slicing or block overlap, and zero-padding is applied to areas smaller than 640 pixels to 640x640. The label files are converted to YOLO detection format and coordinate mapping and normalization are performed based on the scaling and padding relationship.

[0018] Furthermore, the adaptive multi-receptive field module consists of a multi-expanded depth convolution submodule, a channel attention submodule, and a dimension transformation submodule.

[0019] Furthermore, the adaptive multi-receptive field module is an improvement on the C3k2 framework, replacing the bottleneck unit with a multi-expanded depth convolution submodule. This submodule is customized for different convolution kernel sizes. When the convolution kernel size is ≥7, multi-branch expanded depth convolution is used, and its expansion rate combination is adapted according to the convolution kernel size. When the convolution kernel size is 7, a multi-branch convolution structure with an expansion rate of [1,2,3] is used. For a size of 13, an expansion rate combination of [1,2,3,4,5] is used to adapt to the detection requirements of power pole tower components of different sizes.

[0020] Furthermore, the adaptive multi-receptive field module also includes a channel attention submodule and a dimension transformation submodule. The channel attention submodule is configured to adaptively generate and weight the channel weights of the input feature map. It compresses the spatial dimension of the feature map through a global average pooling layer, generates a weight vector representing the importance of each channel using convolution operations and ReLU and Sigmoid activation functions, and multiplies the weight vector with the original input feature map channel by channel. The module also includes a dimension transformation submodule. This submodule transforms the feature tensor from batch-channel-height-width format to batch-height-width-channel format through a layout transformation from NCHW to NHWC. It expands the feature dimension using a linear layer and introduces GELU activation. It then normalizes the global response of the expanded features through a global response normalization layer to stabilize the feature distribution, and finally compresses the feature dimension through a linear layer. Finally, it restores the feature format to batch-channel-height-width format through an inverse transformation from NHWC to NCHW.

[0021] Furthermore, the hierarchical wavelet interaction unit is configured to fuse a large-scale feature and a small-scale feature. It decomposes the large-scale feature into low-frequency components and high-frequency components in three directions (horizontal, vertical, and diagonal) through wavelet transform, enhances the high-frequency components, performs channel interaction between the low-frequency components and the small-scale feature, and finally reconstructs the processed components into a unified-scale fused feature through inverse wavelet transform.

[0022] Furthermore, the wavelet transform in the hierarchical wavelet interaction unit decomposes the large-scale features into approximately low-frequency components and high-frequency components in three directions: horizontal, vertical, and diagonal. The high-frequency components in the three directions are concatenated and fused and enhanced through a residual block containing continuous 3x3 convolutions. The approximately low-frequency components and the small-scale features are concatenated in the channel dimension and interacted and compressed through a channel transform module composed of 1x1 convolutions, compressing the size to three times the dimension of the large-scale features to meet the input dimension requirements of the inverse wavelet transform. The enhanced high-frequency features are concatenated with the interactively compressed low-frequency small-scale features and fused through the inverse wavelet transform to obtain the fused features.

[0023] Furthermore, the distribution-aware confidence optimization head is configured to perform the following operations: Parallel processing of the input features; generating discrete coordinate distribution features of the bounding box through a first convolutional branch, the first convolutional branch containing two consecutive 3x3 standard convolutional layers and a final 1x1 convolutional layer; and generating an initial classification score through a second convolutional branch; reshaping the discrete coordinate distribution features into a tensor structure of [B, M, 4, H, W], where M is the regression maximum value, and processing it with a Softmax function to obtain the probability distribution of the bounding box coordinates; and extracting the coordinates of each coordinate dimension from the probability distribution. The first K maximum probability values ​​are calculated, and the mean of these K probability values ​​is calculated. The first K probability values ​​are concatenated with the mean to form a distribution statistical feature. A multilayer perceptron is used to perform a nonlinear transformation on the distribution statistical feature to generate a quality score adjustment value. The quality score adjustment value is added to the initial classification score to output the optimized classification confidence. The second convolutional branch adopts a hierarchical structure that includes depthwise separable convolutions and standard convolutions. The value of K is 4. The multilayer perceptron consists of two 1x1 convolutional layers, which are used to further perform nonlinear transformations on the statistical features. Its hidden layer dimension is 64.

[0024] The beneficial effects of this invention are:

[0025] This invention proposes a novel adaptive multi-receptive field module. Through the collaborative design of customized dilation rate combinations ([1,2,3] for kernel=7, [1,2,3,4,5] for kernel=13) and a channel attention mechanism, it simultaneously captures features of large-scale structures such as power poles and towers, as well as small-scale components such as insulators and bolts. This overcomes the limitation of traditional models' difficulty in multi-scale adaptation and significantly enhances multi-scale feature extraction capabilities. Furthermore, this invention proposes a novel hierarchical wavelet interaction unit. It achieves high- and low-frequency feature decoupling through Haar wavelet transform, enhances high-frequency details with residual blocks, and optimizes cross-scale interaction with 1×1 convolution. This invention effectively mitigates interference from complex backgrounds such as sky and vegetation in drone aerial photography, significantly improving feature discrimination in complex scenes. It proposes a novel distribution-aware confidence optimization head, which mines the Top-4 statistical features of the bounding box probability distribution and adaptively calibrates the initial classification score using MLP, addressing the pain point of "mismatch between confidence and positioning accuracy" for small targets and improving positioning stability. The effectiveness and superiority of the method are verified on the InsPLAD power pole component detection dataset. Comprehensive evaluation metrics are used to assess model accuracy: CDP, CDR, CDA@50, and CDR@50-95. Experimental results show that this invention improves CDA@50 by 4.4% and CDR@50-95 by 7.0% compared to the YOLOv11 baseline model, and outperforms mainstream models such as YOLOv3-Tiny, YOLOv8n, and RT-DETR-L, fully demonstrating the effectiveness of the proposed method in power pole component detection tasks. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0027] Figure 1 This is a flowchart illustrating the overall implementation of this invention.

[0028] Figure 2 This is a framework diagram of the overall network model in an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of the adaptive multi-receptive field module structure in an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram of the hierarchical wavelet interaction unit structure in an embodiment of the present invention;

[0031] Figure 5 This is a schematic diagram of the distributed sensing confidence optimization head structure in an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions in the embodiments of the present invention will be further described below with reference to the accompanying drawings, which are not intended to limit the present invention.

[0033] like Figure 1 As shown, Figure 1 As shown, an embodiment of the present invention provides a method for detecting power pole components in a drone scenario, including steps S1-S6:

[0034] S1: Obtain the drone aerial image dataset of power pole towers, preprocess it, and input it into the built model;

[0035] S2: The backbone network design uses an adaptive multi-receptive field module to extract multi-scale features through parallel multi-dilated convolutions;

[0036] The adaptive multi-receptive field module includes a multi-expansion depth convolution submodule, a channel attention submodule, and a dimensionality transformation submodule, along with residual connections. The multi-expansion depth convolution submodule covers receptive fields of different scales through multi-branch dilated convolution, enhancing the context modeling capability for power pole components of different sizes. The channel attention submodule strengthens identity-related features and suppresses background interference by generating channel weights. The dimensionality transformation submodule stabilizes feature distribution to adapt to subsequent processing through feature format and dimensionality optimization. Finally, residual connections are combined to preserve the original feature information, alleviating the problem of deep training. Overall, it is used to adapt to multi-scale feature extraction of power pole components in UAV scenarios, outputting enhanced multi-scale features.

[0037] S3: Feature fusion network, which uses hierarchical wavelet interaction units to fuse multi-scale features, combining high-frequency and low-frequency features;

[0038] The hierarchical wavelet interaction unit includes a wavelet transform decomposition module, a high-frequency enhancement module, a cross-scale interaction module, and a wavelet inverse transform reconstruction module. The wavelet transform decomposition module performs multi-directional decomposition on large-scale features, separating low-frequency structural information from high-frequency detail information. The high-frequency enhancement module strengthens target edge and texture features in high-frequency components. The cross-scale interaction module realizes channel association and adaptation between low-frequency information and small-scale features. Finally, the wavelet inverse transform reconstruction module fuses the processed high and low frequency features into unified scale features, enhancing the semantic consistency and detail preservation capability of power pole tower component features in complex backgrounds.

[0039] S4: In the prediction phase, a distributed-aware confidence optimization head is used to optimize the confidence of the detection boxes;

[0040] The distributed perception confidence optimization includes a parallel feature processing branch, a distribution statistics module, a confidence adjustment module, and a detection output fusion module. The parallel feature processing branch generates bounding box coordinate distribution features and initial classification scores, respectively. The distribution statistics module performs probability analysis and feature extraction on the coordinate distribution to capture positioning quality correlation information. The confidence adjustment module calibrates the initial classification score based on the distribution statistics features. Finally, the detection output fusion module integrates the optimized confidence and bounding box information to improve the matching degree between the confidence and positioning accuracy of power pole component detection in UAV scenarios.

[0041] S5: Train the model using the training set and verify the detection accuracy using the test set.

[0042] Furthermore, in step 1, the InsPLAD remote sensing dataset is cropped into 640x640 image blocks without sliding window slicing or block overlap, and zero-padding is used to bring the area to 640x640 for areas less than 640 pixels; the label file is converted to YOLO detection format and coordinate mapping and normalization are completed according to the scaling and padding relationship.

[0043] like Figure 2 The diagram shown illustrates the overall framework of an improved target detection method for detecting power pole components for unmanned aerial vehicles (UAVs) provided by this invention. It mainly comprises three parts: a backbone network, a feature fusion section, and a detection head. The backbone network includes multiple convolutional layers, an adaptive multi-receptive field module, a C3k2 module, an SPPF module, and a C2PSA module for multi-scale feature extraction and enhancement. The feature fusion section includes a splicing layer, convolutional layers, a C3k2 module, and a hierarchical wavelet interaction unit for semantic alignment and detail fusion of cross-scale features. The detection head section features a distributed perceptual confidence optimization head, used to optimize classification confidence and output the final detection result for the power pole components.

[0044] The present invention provides a preferred embodiment for performing S2 to obtain an enhanced multi-scale feature map. For example... Figure 3 As shown, it includes the following steps:

[0045] S2-1: The multi-dilation depthwise convolutional submodule extracts multi-scale basic features from the input feature map. Where C hid The number of hidden channels is determined by the number of output channels C. out C is calculated using the compressibility factor e = 0.5. hid =C out ×e, employing an expanded depthwise convolutional structure with main branches and multiple auxiliary branches to fuse multi-scale receptive fields: the process can be represented as:

[0046] The main branch is: F base1 =BN(DepthConv 7×7，1 (Xhid )),

[0047] Auxiliary branch calculation:

[0048] Multi-branch fusion: F base =F base1 +F base2-1 +F base2-2+ F base2-3 ,

[0049] Where, k i ∈{5, 3, 3} represents the kernel size of the three auxiliary branches, k1 = 5, k2 = 3, k3 = 3, d i ∈{1, 2, 3} represents the expansion rate of the corresponding auxiliary branch, where d1 = 1, d2 = 2, d3 = 3, and DepthConv k×k，d (·) represents a depthwise convolution with a kernel size of k×k and an expansion rate of d; BN(·) is the batch normalization operation; F base This is the multi-scale basic feature map output after multi-branch fusion.

[0050] S2-2: The channel attention submodule enhances key features, focusing on the multi-scale basic feature map F. base After batch normalization, channel weight calibration is performed to enhance the characteristics of tower components and suppress background interference.

[0051] First, global average pooling compresses the spatial dimension: F gap =GAP(F base GAP(·) is global average pooling, which compresses the spatial dimension to 1×1 and focuses on channel-level information;

[0052] Next, a 1×1 convolution is used to reduce the number of channels from C. hid Compressed to the number of internal neurons Conv2d 1×1 (C in →C out ), which is a 1×1 convolution operation;

[0053] Next, a nonlinear representation is introduced through ReLU to enhance feature selectivity: F act =ReLU(F down );

[0054] Next, channel-expanded convolution is performed: a 1×1 convolution is used to expand the number of channels from... Expand back to C hid Restore the original channel dimensions:

[0055] Next, weight normalization is performed, and the channel weights are mapped to the [0,1] interval using the Sigmoid function to generate interpretable attention weights: W ch =Sigmoid(F up );

[0056] Finally, feature weighting fusion is performed: the original features are multiplied element-wise by the channel weights to enhance the key channel features F. att =F base ⊙W ch ⊙ represents element-wise multiplication. Where F att This is the attention-weighted feature map. The S2-3 dimensional transformation submodule optimizes the feature distribution and applies it to the attention-weighted feature F. att The feature distribution is optimized through feature format conversion, dimensionality scaling, and normalization operations. The specific process is as follows:

[0057] First, the dimension order of the feature tensor is adjusted from batch, channel, height, width to batch, height, width, channel to adapt to the linear transformation of the channel dimension; then, the number of channels is changed from Cb to Cc through a linear transformation. hid Expand to 2C hid The process can be represented as:

[0058] F exp =GELU(Linear(Transpose(F) att ,NCHW→NHWC),C hid →2C hid )),

[0059] Where Transpose(·,NCHW→NHWC) is the format conversion, and Linear(C in →C out ) represents the channel linear transformation, and GELU(·) is the activation function.

[0060] Then, through channel compression and inverse format conversion: the linear layer reduces the number of channels from 2C. hid Compress back to C hid Finally, it is converted back to NCHW format, and after compression, it needs to undergo additional batch normalization to stabilize the training.

[0061] F linear2 =Linear(F norm 2C hid →C hid ),

[0062] F nchw =Transpose(F linear2 ,NHWC→NCHW),

[0063] F nchw =BN(Fnchw ),

[0064] The final output is the feature map F after dimension optimization. nchw This process enhances the global discriminative power of key features through global response normalization, and combines dimensional scaling and normalization to stabilize training, providing high-quality input for subsequent residual fusion.

[0065] S2-4 Residual Fusion Output: Features F after dimensionality optimization nchw With the original input features X hid Residual fusion is performed to preserve the original feature information and enhance multi-scale feature representation. The specific process is as follows:

[0066] F out =X hid +F nchw ,

[0067] Final output F out This process preserves the basic information of the original features through the residual structure, while integrating the optimized multi-scale features, so that the output features can be more accurately adapted to the feature extraction needs of power poles.

[0068] This invention provides a preferred embodiment for executing S3, which is used for the fusion enhancement of multi-scale feature maps. It achieves deep interaction between high-frequency details of large-scale features and fine-grained information of small-scale features through wavelet transform, such as... Figure 4 As shown, it includes the following steps:

[0069] S3-1 wavelet transform module for large-scale feature multi-component separation:

[0070] First, the high-rise large-scale feature map F high The input wavelet transform submodule is decomposed into 4 groups of semantically clear feature components through 2×2 grouped convolution with fixed weights.

[0071] wt_decomp = WT(F high ,rev=False);

[0072] The core of the wavelet transform submodule WT consists of four 2×2 convolution kernels, with weights as follows: K for the low-frequency component convolution kernel. a : Horizontal high-frequency component convolution kernel K h : Vertical high-frequency component convolution kernel K v : Diagonal high-frequency component convolution kernel K d :

[0073] The channel separation operation splits wt_decomp into four components;

[0074] [a, h, v, d]=ChannelSplit(wt_decomp, split_dim=1, split_size=C high );

[0075] Where the dimensions of a, h, v, and d are all It achieves a 50% reduction in space size while retaining key information.

[0076] S3-2 high-frequency component residual enhancement adds horizontal, vertical, and diagonal high-frequency components element-wise along the channel dimension and inputs them into the residual block for detail enhancement.

[0077] F hvd =h + v + d;

[0078] The residual block is implemented as two 3×3 convolutions with padding = 1 to maintain the same dimensionality.

[0079] F hvd-enhanced =Conv(C high C high ,3,padding=1)(Conv(C high C high ,3,padding=1)(E hvd ));

[0080] Output Only high-frequency detail features are enhanced, and the number of channels is the same as the input.

[0081] S3-3 Cross-layer feature fusion and channel transformation, transforming low-level small-scale features F low Concatenated with the low-frequency component 'a' along the channel dimension, and input into a channel transformation module consisting of two 1×1 convolutional layers, cross-layer information interaction and dimension adaptation are achieved:

[0082] F concat =Concat(F low (a, dim = 1);

[0083] F trans =Conv1x1(F concat in_channels=C low +C high out_channels = C low +C high )

[0084] F fusion =Conv1x1(F trans in_channels=C low +C highout_channels = 3 × C high )

[0085] Output The number of channels has been adjusted to 3C. high , and the enhanced high-frequency component C high After splicing, the total number of channels is 4C. high This matches the input requirements of the inverse WT transform.

[0086] S3-4 Inverse wavelet transform, feature reconstruction output: the enhanced high-frequency component F hvd-enhanced F after channel transformation fusion The input is concatenated by channel and fed into the inverse wavelet transform submodule. The resulting output, reconstructed using a 2×2 transposed grouped convolution, is the same size as the original large-scale features.

[0087] F wt-input =Concat(F hvd-enhanced ,F fusion dim=1)

[0088] F out =WT(F wt-input ,rev=True)

[0089] Final output F out This feature integrates the global structure of large-scale features, high-frequency details in multiple directions, and fine-grained information of small-scale features, and can accurately adapt to the detection needs of multi-scale components of power poles.

[0090] This invention provides a preferred embodiment that executes step S4 to obtain the detection box and class probability. For example... Figure 5 As shown, it includes the following steps:

[0091] The first convolutional branch of S4-1 generates and statistically optimizes the distribution of bounding box corner points, fusing features across multiple scales. Input the left branch, and perform the following operations in sequence:

[0092] Through successive 3×3 convolution, 3×3 convolution, and 1×1 convolution operations, the discrete coordinate distribution features of the bounding box, pred_corners, are output, as shown in the formula:

[0093] pred_corners=Conv 1×1 (Conv 3×3 (Conv 3×3 (F)));

[0094] in reg_max is the maximum value of the bounding box regression;

[0095] S4-2 performs distribution statistics and quality score generation on the results of the first convolution branch: Distribution statistics analysis is performed on `pred_corners` to generate adjusted quality score values.

[0096] Tensor reshaping: Reshape pred_corners into a structure of [B, reg_max, 4, H, W];

[0097] Softmax probabilistics: Calculate the probability distribution for each corner dimension. The formula is prob = Softmax(pred_corners) reshape dim = 1;

[0098] Top-K statistics: Extract the top 4 probabilities (prob) for each corner dimension. top4 And calculate its mean prob mean ;

[0099] Feature splicing: combining prob top4 +prob mean splicing into statistical features

[0100] MLP Transformation: Input the stat data into an MLP consisting of two 1×1 convolutional layers to generate a quality score adjustment value. Furthermore, the weights and biases of the last layer of the MLP are initialized to 0 to ensure that the original score is not interfered with in the initial stage.

[0101] The second convolutional branch of S4-3 is a composite structure of depthwise separable convolution and standard convolution, which generates the initial classification score. Its structure consists of three consecutive sub-modules, and the specific process is as follows:

[0102] The first submodule to F fusion Perform a 3×3 depthwise separable convolution, extracting spatial features independently for each channel, and then use a 1×1 convolution to reduce the number of channels from C. in Adjust to c3, output features

[0103] The second submodule performs a 3×3 depthwise separable convolution on the F1 score to enhance spatial feature representation, followed by a 1×1 convolution to maintain the number of channels c3, and outputs the features.

[0104] The third submodule: Performs a 1×1 convolution on F2, adjusts the number of channels from c3 to the number of classes nc, and finally outputs the initial classification score:

[0105] scores init =Conv 1×1 (F2) where, In the table, nc=17 represents the number of categories.

[0106] S4-4: Classification Score Optimization

[0107] Initial classification scores init The optimized classification confidence scores are obtained by adding the quality score adjustment value (quality_score) element by element. refined The formula is:

[0108] scores refined =scores init +quality_score, where

[0109] S4-5 Feature Concatenation and Detection Output combines the bounding box corner distribution features (pred_corners) with the optimized classification confidence scores. refined Concatenate the data along the channel dimension to obtain the input features of the detection head.

[0110] F detect =Concat(pred_corners,scores refined (dim=1);

[0111] The final output detection boxes and category probabilities can accurately adapt to the detection needs of multi-scale components of power poles. By optimizing distributed perception, the detection confidence of small components is improved, background false detections are suppressed, and the reliability of detection results in UAV inspection scenarios is significantly enhanced.

[0112] In this embodiment, the proposed method is compared with other advanced methods on the InsPLAD dataset for UAV power pole component detection, and commonly used evaluation metrics are selected: P, R, mAP50, mAP50-95, Params (M), GFLOPs, Size (MB), and Inference. The experimental results are shown in Table 1.

[0113] Table 1 compares the proposed methods on the InsPLAD dataset.

[0114]

[0115] As shown in Table 1, this embodiment outperforms existing methods in multiple metrics on the dataset, demonstrating the effectiveness of the method in this embodiment.

[0116] The above is a detailed description of specific embodiments of the present invention. However, the present invention is not limited to the specific embodiments described above. Those skilled in the art can make various modifications or variations to the above embodiments without affecting the essential content of the present invention, and all such modifications or variations should be within the protection scope of the present invention.

Claims

1. A power tower component detection method for unmanned aerial vehicle scenarios, characterized by, The method comprises the following steps: S1: obtaining unmanned aerial vehicle aerial power tower image dataset; S2: preprocessing the power tower dataset; S3: establishing a power tower component target detection model for the unmanned aerial vehicle scene; S3.1: in the feature extraction stage, an adaptive multi-receptive field module is adopted, which extracts multi-scale features through its internal parallel multi-dilated depth convolution submodule. The submodule covers different scale receptive fields in the same channel dimension through a plurality of fixed dilation rate convolution branches, so as to enhance the context modeling capability of the target. Then, the channel attention submodule and the dimension transformation submodule are combined to complete feature reconstruction, and the enhanced multi-scale features are output through residual connection. S3.2: in the feature fusion stage, a hierarchical wavelet interaction unit is adopted to perform wavelet transform decomposition, detail enhancement and channel interaction on adjacent scale features; first, the large-scale features are decomposed by wavelet transform to obtain low-frequency and high-frequency components in horizontal, vertical and diagonal directions, the high-frequency details are enhanced and fused, and the low-frequency components are interacted and compressed with the small-scale features in the channel dimension; then, the unified scale fusion features are reconstructed through inverse wavelet transform, so as to realize semantic alignment and retain edge details in complex background. S3.3: in the prediction stage, a distribution-aware confidence optimization head is adopted, which adjusts the initial classification score based on the Top-k probability and its mean statistical features by statistically analyzing the distribution of the predicted bounding boxes, so as to optimize the confidence of the target; the process is processed by a multi-layer perceptron to generate an optimized confidence, and the final detection box and class probability are output in combination with the bounding box coordinate distribution features. S4: training the model using the training set of the dataset to obtain an optimal model; S5: using the optimal model to test the test set images of the dataset to obtain the final test result.

2. The method for unmanned airfield scene-oriented power tower component detection according to claim 1, characterized in that: The InsPLAD remote sensing dataset is cropped into 640x640 image blocks, without sliding window slicing and block overlap. The zero padding is used to fill the area less than 640 pixels to 640x640. The label file is converted into YOLO detection format and the coordinate mapping and normalization are completed according to the scaling and padding relationship.

3. The method of claim 1, wherein: The adaptive multi-receptive field module is composed of a multi-dilated depth convolution submodule, a channel attention submodule and a dimension transformation submodule.

4. The method of claim 3, wherein: The adaptive multi-receptive field module is improved on the basis of the C3k2 framework, and the multi-dilated depth convolution submodule is used to replace the bottleneck unit. The submodule is customized for different convolution kernel sizes. When the convolution kernel size is greater than or equal to 7, the multi-branch dilated depth convolution is used, and the dilation rate combination is adapted according to the convolution kernel size. When the convolution kernel size is 7, the multi-branch convolution structure with dilation rates of [1, 2, 3] is used. For the size of 13, the dilation rate combination [1, 2, 3, 4, 5] is used, so as to adapt to the detection requirements of different size power tower components.

5. The method of claim 3, wherein: The adaptive multi-receptive field module further comprises a channel attention sub-module and a dimension transformation sub-module; the channel attention sub-module is configured to perform adaptive generation and weighting of channel weights on the input feature map, compress the spatial dimension of the feature map through a global average pooling layer, generate a weight vector representing the importance of each channel by using a convolution operation and a Sigmoid activation function, and multiply the weight vector and the original input feature map channel by channel; and comprises a dimension transformation sub-module; the sub-module converts the feature tensor from batch-channel-height-width format to batch-height-width-channel format through NCHW to NHWC layout transformation; uses a linear layer to realize feature dimension expansion and introduce GELU activation, performs global response normalization on the expanded feature through a global response normalization layer to stabilize the feature distribution, and then completes feature dimension compression through a linear layer; finally, the feature format is restored to batch-channel-height-width format through NHWC to NCHW inverse transformation.

6. The method of claim 1, wherein: The hierarchical wavelet interaction unit is configured to fuse a large-scale feature and a small-scale feature, decompose the large-scale feature into a low-frequency component and high-frequency components in horizontal, vertical and diagonal directions through wavelet transform, respectively enhance the high-frequency components, channel-interact the low-frequency component and the small-scale feature, and finally reconstruct the processed components into a unified scale of fusion features through inverse wavelet transform.

7. The method of claim 6, wherein the method is for an unmanned airfield scenario. The wavelet transform in the hierarchical wavelet interaction unit decomposes the large-scale feature into an approximate low-frequency component and high-frequency components in horizontal, vertical and diagonal directions; the three-direction high-frequency components are spliced and high-frequency fusion enhanced through a residual block comprising consecutive 3x3 convolutions; the approximate low-frequency component and the small-scale feature are spliced in the channel dimension and cross-channel interacted through a channel transformation module comprising a 1x1 convolution, compressed to 3 times the dimension of the large-scale feature, and adapted to the input dimension requirement of the inverse wavelet transform; the enhanced high-frequency features and the low-frequency small-scale features after interaction and compression are spliced, and the fusion features are obtained through inverse wavelet transform.

8. The method of claim 1, wherein: The distribution-aware confidence optimization head is configured to perform the following operations: parallel processing of input features, generating discrete coordinate distribution features of the bounding box through a first convolution branch, the first convolution branch comprising two consecutive 3x3 standard convolution layers and a final 1x1 convolution layer, and generating initial classification scores through a second convolution branch; the discrete coordinate distribution features are reshaped into a tensor structure of [B, M, 4, H, W], where M is the maximum regression value, and processed through a Softmax function to obtain the probability distribution of the bounding box coordinates; the top K maximum probability values in each coordinate dimension are extracted from the probability distribution, and the mean value of the K probability values is calculated; the top K probability values and the mean value are spliced to form distribution statistical features; The distribution statistical features are nonlinearly transformed by using a multilayer perceptron to generate a quality score adjustment value; the quality score adjustment value is added to the initial classification score to output an optimized classification confidence; the second convolution branch adopts a hierarchical structure containing a depth separable convolution and a standard convolution; the value of K is 4; the multilayer perceptron is composed of two 1x1 convolution layers, which are used to further nonlinearly transform the statistical features, and the hidden layer dimension is 64.

Images