Dual-path decoupling power inspection image classification method and system based on frequency band statistics

Abstract

The application belongs to the technical field of image processing, and particularly relates to a dual-path decoupling power inspection image classification method and system based on frequency band statistics. Under the premise of controllable calculation overhead, a dual-path processing mode composed of state space modeling processing and frequency domain filtering enhancement processing is constructed to extract features, and a frequency band statistical feature extraction and response allocation process based on discrete wavelet transform is introduced in the training stage to guide the dual-path to form a response allocation relationship facing different frequency band information, so as to adjust the contribution proportion of the dual-path generated features, improve the feature expression capability, recognition reliability and deployment adaptability in the power inspection image classification task. In the inference stage, the discrete wavelet analysis and response allocation coefficient generation are stopped, and only the trained and optimized dual-path structure is reserved to perform feature extraction and fusion, so as to realize the decoupling of explicit guidance in the training stage and light execution in the inference stage.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a dual-path decoupling power line inspection image classification method and system based on frequency band statistics. Background Technology

[0002] Images acquired by drones during power line inspections contain critical components such as insulators, vibration dampers, and clamps. The results of their automatic identification directly affect the accurate assessment of the transmission line's operational status. Due to the complex inspection environment, significant differences in component structures, and prominent background noise interference, extracting stable and discriminative features from high-resolution images is a pressing issue that needs to be addressed in current engineering applications.

[0003] Visual processing methods based on convolutional neural networks (CNNs) excel at extracting local details such as image texture and edges, thus typically achieving high classification accuracy in image classification tasks. However, due to their limited receptive field of view, they struggle to fully model long-distance spatial dependencies between components in power transmission towers, limiting their recall capabilities for critical power components and potentially leading to missed detections.

[0004] The visual Transformer method based on self-attention mechanism can achieve global contextual information modeling, but the computational complexity of this method increases sharply with image resolution, and it has high requirements for computing power and video memory resources in high-resolution power inspection image processing. In addition, due to the lack of explicit image inductive bias, it has shortcomings under limited sample conditions.

[0005] Visual state-space models (such as Vision Mamba), as emerging sequence modeling architectures, have shown great potential in long-range context modeling through selective scanning mechanisms and linearly complex sequence processing. However, existing research indicates that these methods do not always maintain a performance advantage in general image classification tasks, and their ability to represent local textures and fine-grained details is weaker than convolutional neural network models in some pure image classification scenarios.

[0006] Therefore, in power line inspection scenarios characterized by high-resolution input and complex backgrounds, existing single-path structures struggle to balance the collaborative stability between global modeling and local detail compensation, resulting in technical bottlenecks such as unbalanced feature extraction capabilities and insufficient overall recognition stability. There is an urgent need for a feature extraction method that can integrate the advantages of multiple feature representations and achieve a balance between modeling capabilities and computational efficiency. Summary of the Invention

[0007] To address the problems raised in the background art, this invention provides a dual-path decoupling power line inspection image classification method and system based on frequency band statistics.

[0008] The technical solution of the present invention is as follows: This invention provides a dual-path decoupling power line inspection image classification method based on frequency band statistics, comprising: S1: The power inspection image to be classified is divided into image blocks to obtain the features to be processed; S2: The features to be processed are processed through state space modeling to generate the first processed features; The features to be processed are enhanced by frequency domain filtering to generate second processed features, specifically: After the features to be processed are subjected to Fast Fourier Transform, learnable frequency domain filtering weights are applied, and then the inverse Fourier Transform is used to generate the second processed features. S3: The features to be processed are subjected to discrete wavelet transform to decompose them into statistical features of multiple frequency band sub-bands. For the statistical features of the frequency band sub-bands, after fully connected processing and constraint to a preset range, the first response allocation coefficient and the second response allocation coefficient are generated. S4: Multiply the first processing feature by the first response allocation coefficient, multiply the second processing feature by the second response allocation coefficient to obtain the corresponding modulation feature, then sum them by weight to generate the current feature; S5: The current feature is used as the feature to be processed. S2 is executed. After the preset conditions are met, classification processing is performed to obtain the probability of power component category.

[0009] Based on the dual-path decoupled power line inspection image classification method based on frequency band statistics described above, in step S3, the statistical features of the frequency band sub-bands, after being processed by a fully connected layer and constrained to a preset range, generate a first response allocation coefficient and a second response allocation coefficient, specifically as follows: The statistical features of each frequency band sub-band are concatenated to form a statistical vector, and then mapped through a fully connected layer to obtain two original response values ​​corresponding to state space modeling and frequency domain filtering enhancement, respectively. The original response values ​​are constrained by the Sigmoid function to obtain the first response allocation coefficient and the second response allocation coefficient.

[0010] Based on the dual-path decoupling power inspection image classification method based on frequency band statistics described above, in step S5, the preset conditions are met, and the structural parameters of the state space modeling processing and frequency domain filtering enhancement processing described in step S2 are optimized to a preset threshold by using backpropagation of the loss function.

[0011] Based on the above-described dual-path decoupled power inspection image classification method based on frequency band statistics, for the on-site power inspection image to be classified, after passing through S1 and S2 in sequence, the first processing feature and the second processing feature are weighted and summed, and the resulting fused feature is processed by residual connection to calculate the probability of power component category.

[0012] Based on the dual-path decoupling power line inspection image classification method based on frequency band statistics described above, the features to be processed in S3 undergo discrete wavelet transform to decompose them into statistical features of multiple frequency band sub-bands, specifically: The features to be processed are reconstructed into a two-dimensional feature map, and the two-dimensional feature map is decomposed by discrete wavelet transform to obtain multiple frequency band sub-bands; Calculate the statistical characterization of each frequency domain sub-band, which includes: wavelet coefficient energy, mean, and variance.

[0013] Based on the dual-path decoupling power line inspection image classification method based on frequency band statistics described above, the learningable frequency domain filtering weights applied in S2 are determined by the formula: To achieve; In the formula, , These represent the learnable frequency domain filter weights and learnable parameters, respectively.

[0014] This invention also provides a dual-path decoupling power line inspection image classification system based on frequency band statistics, comprising: Preprocessing module: The power inspection images to be classified are divided into image blocks to obtain the features to be processed; Dual-path processing module: The features to be processed are modeled in the state space to generate the first processed features; The features to be processed are enhanced by frequency domain filtering to generate second processed features, specifically: After the features to be processed are subjected to Fast Fourier Transform, learnable frequency domain filtering weights are applied, and then the inverse Fourier Transform is used to generate the second processed features. Frequency band statistics module: The features to be processed are subjected to discrete wavelet transform to decompose them into statistical features of multiple frequency band sub-bands. For the statistical features of the frequency band sub-bands, after fully connected processing and constraint to a preset range, the first response allocation coefficient and the second response allocation coefficient are generated. Response coefficient modulation module: The first processed feature is multiplied by the first response allocation coefficient, and the second processed feature is multiplied by the second response allocation coefficient to obtain the corresponding modulation feature. Then, the modulated features are weighted and summed to generate the current feature. Classification module: The current feature is taken as the feature to be processed and enters the dual-path processing module. After the preset conditions are met, classification processing is performed to obtain the probability of the power component category.

[0015] Based on the dual-path decoupled power line inspection image classification system based on frequency band statistics described above, in the frequency band statistics module, the statistical features of the frequency band sub-bands, after being processed by a fully connected layer and constrained to a preset range, generate a first response allocation coefficient and a second response allocation coefficient, specifically as follows: The statistical features of each frequency band sub-band are concatenated to form a statistical vector, and then mapped through a fully connected layer to obtain two original response values ​​corresponding to state space modeling and frequency domain filtering enhancement, respectively. The original response values ​​are constrained by the Sigmoid function to obtain the first response allocation coefficient and the second response allocation coefficient.

[0016] Based on the above-described dual-path decoupled power inspection image classification system based on frequency band statistics, in the classification module, reaching the preset condition means using backpropagation of the loss function to optimize the structural parameters of the state space modeling processing and frequency domain filtering enhancement processing of the dual-path processing module to a preset threshold.

[0017] Based on the frequency band statistics-based dual-path decoupled power inspection image classification system described above, in the dual-path processing module, the application of learnable frequency domain filtering weights is determined by the formula: To achieve; In the formula, , These represent the learnable frequency domain filter weights and learnable parameters, respectively.

[0018] Beneficial effects This invention, under the premise of controllable computational overhead, constructs a dual-path processing method consisting of state-space modeling and frequency-domain filtering enhancement to extract features. In the training phase, a frequency band statistical feature extraction and response allocation process based on discrete wavelet transform is introduced to guide the dual paths to form response allocation relationships oriented towards different frequency band information, thereby adjusting the contribution ratio of features generated by the dual paths. This can enhance the collaborative feature expression capability of the dual paths for different frequency band information while keeping the overall model complexity controllable, thereby improving the feature expression capability, recognition reliability, and deployment adaptability in power inspection image classification tasks.

[0019] During the inference phase, discrete wavelet analysis and response allocation coefficient generation are stopped. Only the dual-path structure optimized by training is retained for feature extraction and fusion. The path response allocation relationship formed during the training phase is solidified in the dual-path parameters through parameter optimization. The dual-path parameters are the parameters optimized by the response allocation coefficients during the training phase, thereby decoupling the explicit guidance of the training phase from the lightweight execution of the inference phase. Attached Figure Description

[0020] Figure 1 This is a flowchart of the frequency domain filtering enhancement process based on Fast Fourier Transform in Example 1.

[0021] Figure 2 This is a flowchart of frequency band statistical feature extraction and response allocation coefficient generation during the training phase of Example 1.

[0022] Figure 3This is a schematic diagram of the relationship of a single TriMamba layer in Example 2. Detailed Implementation

[0023] The following examples are intended to illustrate the present invention, and not to further limit the invention.

[0024] Example 1 This embodiment provides a dual-path decoupling power line inspection image classification method based on frequency band statistics, including: S1: The power inspection image to be classified is divided into image blocks to obtain the features to be processed.

[0025] The power inspection images to be classified are converted into image block embedding sequences through configurable convolutional projection. The image blocks can be non-overlapping or partially overlapping, and are mapped to image token sequences.

[0026] S2: The features to be processed are extracted via dual-path feature extraction, as detailed below.

[0027] One path is that the feature to be processed is processed through state space modeling to generate the first processed feature.

[0028] State-space modeling is used to extract long-range dependencies and overall structural information from image patch sequences.

[0029] The state space modeling process can be implemented using existing visual state space models to perform long-range dependency modeling on image patch embedding sequences.

[0030] Preferably, the state space modeling process employs a Vision Mamba-based state space scanning mechanism to perform sequence modeling on the image patch sequence in order to extract global contextual relationships and structural continuity information.

[0031] One approach is to enhance the features to be processed through frequency domain filtering to generate second-processed features, specifically: After the features to be processed undergo a fast Fourier transform, learnable frequency domain filtering weights are applied, and then an inverse Fourier transform is performed to generate the second processed features.

[0032] Among them, frequency domain filtering enhancement processing is used to perform frequency selective response adjustment on the features to be processed in order to compensate for the representation of local texture, edge details and periodic structural information of power components.

[0033] The frequency domain filtering enhancement processing described in this invention employs a frequency domain filtering structure based on Fast Fourier Transform, such as... Figure 1 ,include: The features to be processed are subjected to Fast Fourier Transform to obtain frequency domain features. Learnable frequency domain filtering weights are applied to the frequency domain features, and then the spatial features are recovered by Inverse Fourier Transform. These features are then used as local enhancement results in subsequent feature fusion.

[0034] Furthermore, to improve numerical stability, the frequency domain filtering weights can be constrained using the Sigmoid function.

[0035] Preferably, the application of learnable frequency domain filter weights is performed using the formula: To achieve this.

[0036] In the formula, , These represent the learnable frequency domain filter weights and learnable parameters, respectively.

[0037] The frequency domain filtering weights are trainable parameters that, after being constrained, are applied to the frequency domain features to enhance or suppress the corresponding frequency components, thereby achieving a compensatory expression of local high-frequency texture and edge information.

[0038] S3: The features to be processed are subjected to discrete wavelet transform to decompose them into statistical features of multiple frequency band sub-bands. For the statistical features of the frequency band sub-bands, after fully connected processing and constraint to a preset range, the first response allocation coefficient and the second response allocation coefficient are generated.

[0039] The discrete wavelet transform and response assignment coefficient generation process are not used as independent classification features, but rather as a response assignment guidance process during the training phase.

[0040] like Figure 2 Preferably, the features to be processed undergo discrete wavelet transform to decompose them into statistical features of multiple frequency band sub-bands, specifically: The features to be processed are reconstructed into a two-dimensional feature map, and the two-dimensional feature map is decomposed by discrete wavelet transform to obtain multiple frequency band sub-bands; Calculate the statistical characterization of each frequency domain sub-band, which includes: wavelet coefficient energy, mean, and variance.

[0041] In this embodiment, the energy of the sub-band wavelet coefficients Calculate using the following formula: .

[0042] in For the first Wavelet coefficients of each band, Let be the number of elements in this subband. This represents the square of the Frobenius norm.

[0043] Furthermore, the statistical characteristics of the frequency band sub-bands, after being processed by a fully connected layer and constrained to a preset range, generate a first response allocation coefficient and a second response allocation coefficient, specifically as follows: The statistical characteristics of each frequency band sub-band are concatenated to form a statistical vector. The two original response values, corresponding to state-space modeling and frequency-domain filtering enhancement, are obtained through mapping via a fully connected layer. and The first response allocation coefficient is obtained by constraining the original response values ​​using the Sigmoid function. Second response allocation coefficient The formula is as follows.

[0044] ; ; .

[0045] in, For learnable mapping parameters, For bias terms, This represents the Sigmoid function. and These are used to adjust the response of features obtained from state-space modeling and frequency-domain filtering enhancement, respectively.

[0046] S4: Multiply the first processing feature by the first response allocation coefficient, and multiply the second processing feature by the second response allocation coefficient to obtain the corresponding modulation feature. Then, sum the values ​​by weight to generate the current feature.

[0047] S5: The current feature is used as the feature to be processed. S2 is executed. After the preset conditions are met, classification processing is performed to obtain the probability of power component category.

[0048] The condition of reaching the preset condition is to optimize the structural parameters of the state space modeling and frequency domain filtering enhancement processes described in S2 to a preset threshold by using backpropagation of the loss function.

[0049] The classification method provided in Example 1 is for power line inspection labeled images with category labels, and is applicable to the corresponding classification model during the training phase. The S3 frequency band statistics and response allocation process generates response allocation coefficients corresponding to state-space modeling and frequency domain filtering enhancement processes, respectively, based on the frequency band statistics results, and adjusts the relative contributions of the dual-path generated features. Through backpropagation of the loss function during training, the dual-path structural parameters are jointly optimized within a unified framework under this process, thereby improving the stability and coordination of the dual-path contribution allocation.

[0050] Based on this, for the unlabeled on-site power inspection images to be classified, after passing through S1 and S2 in sequence, the first processing feature and the second processing feature are weighted and summed, and the resulting fused feature is processed by residual connection to calculate the probability of power component category.

[0051] This invention constructs a dual-path processing method consisting of state-space modeling and frequency-domain filtering enhancement to extract features. During the training phase, a frequency band statistical feature extraction and response allocation process based on discrete wavelet transform is introduced to guide the dual paths to form response allocation relationships for different frequency band information, thereby adjusting the contribution ratio of features generated by the dual paths. This can enhance the collaborative feature expression capability of the dual paths for different frequency band information while keeping the overall model complexity controllable, thereby improving the feature expression capability, recognition reliability, and deployment adaptability in power inspection image classification tasks.

[0052] During the inference phase, discrete wavelet analysis and response allocation coefficient generation are stopped. Only the dual-path structure optimized by training is retained for feature extraction and fusion. The path response allocation relationship formed during the training phase is solidified in the dual-path parameters through parameter optimization. The dual-path parameters are the parameters optimized by the response allocation coefficients during the training phase, thereby decoupling the explicit guidance of the training phase from the lightweight execution of the inference phase.

[0053] Example 2 Existing CNN, Transformer and visual state space models each have their own focus in feature extraction capabilities. Single-path structures usually cannot take into account the robust representation between global and local details, which will affect the stability and reliability of key power component identification.

[0054] To address the limitations of existing single-architecture models in power line inspection image classification tasks, which struggle to simultaneously handle global context modeling, local detail representation, and insufficient frequency domain stability, a classification model adapted to the classification method described in Example 1 includes: Image Patch Embedding Layer: The input power inspection image is converted into an image patch embedding sequence through a configurable convolutional projection method. The size and sliding step of each image patch can be set according to the model parameters. The image patches can be non-overlapping or partially overlapping structures, and are mapped to an image token sequence.

[0055] Multi-layer TriMamba backbone network: Consists of multiple stacked TriMamba layers. Each layer includes at least a dual-path structure consisting of a state-space modeling path and a frequency-domain filtering enhancement path, as well as frequency band statistical feature extraction and response allocation processes during the training phase. See [link to documentation]. Figure 3 .

[0056] Classification Head: Used to map the output features of the backbone network and output the probability of power component categories.

[0057] After training, the path response allocation relationship formed during the training phase of the model is solidified into the dual-path parameters through parameter optimization, meaning the dual-path structure parameters have been optimized. During the model inference phase, discrete wavelet transform and response allocation coefficient generation are stopped, while state-space modeling and frequency domain filtering enhancement are retained for feature extraction and fusion. This decouples the explicit guidance of the training phase from the lightweight execution of the inference phase.

[0058] Example 3 This embodiment provides a dual-path decoupled power line inspection image classification system based on frequency band statistics, including: Preprocessing module: The power inspection images to be classified are divided into image blocks to obtain the features to be processed; Dual-path processing module: The features to be processed are modeled in the state space to generate the first processed features; The features to be processed are enhanced by frequency domain filtering to generate second processed features, specifically: After the features to be processed are subjected to Fast Fourier Transform, learnable frequency domain filtering weights are applied, and then the inverse Fourier Transform is used to generate the second processed features. Frequency band statistics module: The features to be processed are subjected to discrete wavelet transform to decompose them into statistical features of multiple frequency band sub-bands. For the statistical features of the frequency band sub-bands, after fully connected processing and constraint to a preset range, the first response allocation coefficient and the second response allocation coefficient are generated. Response coefficient modulation module: The first processed feature is multiplied by the first response allocation coefficient, and the second processed feature is multiplied by the second response allocation coefficient to obtain the corresponding modulation feature. Then, the modulated features are weighted and summed to generate the current feature. Classification module: The current feature is taken as the feature to be processed and enters the dual-path processing module. After the preset conditions are met, classification processing is performed to obtain the probability of the power component category.

[0059] Preferably, in the dual-path processing module, the application of learnable frequency domain filtering weights is performed using the formula: To achieve; In the formula, , These represent the learnable frequency domain filter weights and learnable parameters, respectively.

[0060] In the frequency band statistics module, the statistical characteristics of the frequency band sub-bands, after being processed by a fully connected layer and constrained to a preset range, generate a first response allocation coefficient and a second response allocation coefficient, specifically as follows: The statistical features of each frequency band sub-band are concatenated to form a statistical vector, and then mapped through a fully connected layer to obtain two original response values ​​corresponding to state space modeling and frequency domain filtering enhancement, respectively. The original response values ​​are constrained by the Sigmoid function to obtain the first response allocation coefficient and the second response allocation coefficient.

[0061] In the classification module, reaching the preset condition means using backpropagation of the loss function to optimize the structural parameters of the state space modeling and frequency domain filtering enhancement processes in the dual-path processing module to a preset threshold.

[0062] Application Example 1 In this application example, the model input is a 512×512 power line inspection image. The image patch embedding layer uses a 16×16 non-overlapping convolution with a stride of 16 to convert the image into a sequence of image patch tokens of length 1024, and adds a classification token to form a 1025×384 input representation with an embedding dimension of 384.

[0063] This application example uses the publicly available insPLAD transmission line component image dataset. This dataset, released in 2023, was collected from UAV power line inspection scenarios and contains images of various typical transmission line components. This application example reorganizes it into an 18-class image classification task. The dataset has a total size of 10561 images, divided into a training set of 7935 images, a validation set of 1839 images, and a test set of 787 images. The images undergo the following processing during the training phase: The image was scaled to 512×512; standardized with a mean of [0.485, 0.456, 0.406] and a standard deviation of [0.229, 0.224, 0.225]; enhancement operations such as random horizontal flipping and random cropping were performed; only scaling and standardization were performed during the verification and testing phases.

[0064] Training was performed using the AdamW optimizer with a weight decay factor of 0.1 and an initial learning rate of 5e. -5 In conjunction with the cosine annealing strategy, the minimum learning rate is set to 1×10. -6 The training consisted of 300 epochs, with a batch size of 8. The gradient clipping threshold was set to 3.0. The loss function used was cross-entropy with class weights to mitigate the class imbalance problem. Mixed precision calculation was enabled during training.

[0065] To ensure the fairness and consistency of the comparison results, all models used the same data partitioning method, input resolution, and training strategy, and were trained and evaluated under the same hardware environment and number of training rounds.

[0066] This application example uses the following metrics to evaluate the model performance: Acc@1 represents the proportion of predicted classes that match the true classes in the model's prediction results, and is used to measure the overall classification accuracy.

[0067] Precision represents the proportion of samples predicted to be of a certain class that are actually of that class, reflecting the likelihood of model misclassification.

[0068] Recall This represents the proportion of samples that are correctly identified by the model in a certain category, reflecting the number of missed detections by the model.

[0069] F1 score is Precision A combined metric with Recall, used to evaluate the model's ability to balance false positives and false negatives.

[0070] FLOPs represent the number of floating-point operations required by the model in a single forward inference process, and are used to measure the computational complexity and inference efficiency of the model in engineering deployment scenarios.

[0071] Given that this application does not improve performance by adding independent inference branches, but rather guides the formation of response allocation relationships for different frequency band information through the frequency band statistical feature extraction and response allocation process during the training phase, and solidifies this response allocation relationship into the dual-path parameters during the inference phase, the dual-path parameters are parameters optimized under the constraint of response allocation coefficients during the training phase. To verify the synergistic gain effect between the dual-path structure and the frequency band statistical guidance and response allocation during the training phase, this application example sets up an ablation experiment to compare the model performance under different composition methods. The results are shown in Table 1. Among them, compared with the model of this application (i.e., the "complete dual-path response allocation model"), the "state space baseline model" only contains the state space modeling path; the "frequency domain filtering enhancement model only" adds the frequency domain filtering enhancement path to the state space baseline model; and the "frequency band statistical guidance model only" adds the training phase frequency band statistical guidance and response allocation mechanism to the state space baseline model.

[0072] Table 1 Ablation Experiment Results Table 1 shows that the complete dual-path response allocation model achieves the best overall performance, outperforming the frequency-domain filtering enhancement model, the frequency band statistical guidance model, and the state-space baseline model in terms of F1-score, Precision, Recall, and Acc@1. Compared to the state-space baseline model, the complete dual-path response allocation model in this application improves the F1-score from 0.41 to 0.54, the Recall from 0.41 to 0.60, and the Acc@1 from 0.69 to 0.79. This indicates that the improvement in this application does not originate from a single path enhancement, but rather from the synergistic effect formed between the state-space modeling path, the frequency-domain filtering enhancement path, and the frequency band statistical guidance and dual-path structure during the training phase.

[0073] Furthermore, to verify the performance advantages of this application among similar visual models, this application example selects VisionMamba (Baseline ViM), Swin-Tiny Transformer (Swin-T), EdgeViT, ShuffleNetV2, and MambaVision as comparison models. All models were trained from scratch on the same dataset, using the same data preprocessing methods, input resolution (512×512), training strategy, and number of training epochs (300 epochs) to ensure the fairness of the comparison. Because missing key components in power line inspection scenarios can lead to high engineering risks, this application example, in addition to using Acc@1, focuses more on metrics that reflect task reliability, such as Recall and F1-score. The results are shown in Table 2.

[0074] Table 2 Comparison of experimental results Table 2 shows that the model corresponding to the proposed method does not pursue absolute optimization of all classification metrics. Instead, it focuses on improving metrics more relevant to the risk of missed detections, such as Recall and F1-score, under conditions of lower computational complexity. This results in better recognition reliability and engineering applicability in power line inspection image classification tasks. Specifically, the F1-score of the proposed method reaches 0.54, which is better than Baseline VIM's 0.41, higher than MambaVision's 0.52, and also better than Swin-T's 0.47 and EdgeViT's 0.49. In terms of Recall, the proposed method achieves 0.60, which is higher than Baseline ViM's 0.41, higher than MambaVision's 0.54, and better than Swin-T and EdgeViT's 0.51. Compared to ShuffleNetV2, the proposed method achieves comparable levels in Recall and F1-score, but reduces FLOPs from 8.2G to 0.34G. The results show that this application guides response allocation through frequency band statistics during the training phase, enabling the dual-path collaborative feature representation capability to be maintained during the inference phase even without retaining additional control branches, thereby reducing the risk of missed detections.

[0075] It should be noted that BaselineVIM in Table 2 does not refer to all publicly available versions of the Vision Mamba model, but rather to the baseline configuration of a lightweight visual state space model built in this application example under the same input resolution, training process, and code framework. Its network width of 256, number of layers of 4, and classification head settings were uniformly set according to the engineering deployment constraints of this application example; therefore, its FLOPs, as measured by the actual code, are 0.31G. Correspondingly, after introducing a frequency domain filtering enhancement path and a frequency band statistical guidance mechanism during the training phase on this lightweight baseline, the FLOPs increased to 0.34G. This indicates that the technical effect of this application does not stem from a significant increase in inference computation, but rather from the feature representation optimization brought about by dual-path collaborative modeling and the frequency band statistical guidance and response allocation during the training phase.

[0076] While some comparative models slightly outperformed the model corresponding to the method in this application in Acc@1, their Recall and F1-score metrics were not generally superior. For power line inspection scenarios, missing critical components often poses a higher engineering risk than false positives; therefore, the performance of this application in terms of Recall and F1-score is more in line with practical application needs. Meanwhile, the FLOPs of the model corresponding to the method in this application are 0.34G, lower than ShuffleNetV2's 8.2G, Swin-T's 23.5G, EdgeViT's 22.1G, and MambaVision's 1.9G, achieving a more reasonable balance between recognition performance and computational overhead.

[0077] In summary, the experimental results show that this application does not improve accuracy by adding independent branches in the inference stage. Instead, it introduces a frequency band statistics-guided response allocation mechanism in the training stage, which enables the state space modeling path and the frequency domain filtering enhancement path to form a path contribution relationship controlled by response allocation in a unified backbone structure. This relationship is then solidified in the dual-path parameters optimized under the action of the response allocation coefficient in the training stage. This approach maintains low computational complexity in the inference stage while improving the recognition reliability and engineering applicability of power line inspection image classification tasks.

Claims

1. A dual-path decoupling power line inspection image classification method based on frequency band statistics, characterized in that, include: S1: The power inspection image to be classified is divided into image blocks to obtain the features to be processed; S2: The features to be processed are processed through state space modeling to generate the first processed features; The features to be processed are enhanced by frequency domain filtering to generate second processed features, specifically: After the features to be processed are subjected to Fast Fourier Transform, learnable frequency domain filtering weights are applied, and then the inverse Fourier Transform is used to generate the second processed features. S3: The features to be processed are subjected to discrete wavelet transform to decompose them into statistical features of multiple frequency band sub-bands. For the statistical features of the frequency band sub-bands, after fully connected processing and constraint to a preset range, the first response allocation coefficient and the second response allocation coefficient are generated. S4: Multiply the first processing feature by the first response allocation coefficient, multiply the second processing feature by the second response allocation coefficient to obtain the corresponding modulation feature, then sum them by weight to generate the current feature; S5: The current feature is used as the feature to be processed. S2 is executed. After the preset conditions are met, classification processing is performed to obtain the probability of power component category.

2. The dual-path decoupling power line inspection image classification method based on frequency band statistics according to claim 1, characterized in that, The statistical characteristics of the frequency band sub-bands described in S3, after being processed by a fully connected layer and constrained to a preset range, generate a first response allocation coefficient and a second response allocation coefficient, specifically as follows: The statistical features of each frequency band sub-band are concatenated to form a statistical vector, and then mapped through a fully connected layer to obtain two original response values ​​corresponding to state space modeling and frequency domain filtering enhancement, respectively. The original response values ​​are constrained by the Sigmoid function to obtain the first response allocation coefficient and the second response allocation coefficient.

3. The dual-path decoupling power line inspection image classification method based on frequency band statistics according to claim 1, characterized in that, When the preset conditions are met as described in S5, the structural parameters of the state space modeling and frequency domain filtering enhancement processes described in S2 are optimized to a preset threshold by using backpropagation of the loss function.

4. The dual-path decoupling power line inspection image classification method based on frequency band statistics according to claim 3, characterized in that, For the on-site power inspection images to be classified, after passing through S1 and S2 in sequence, the first processing feature and the second processing feature are weighted and summed. The resulting fused feature is then processed by residual connection to calculate the probability of power component category.

5. The dual-path decoupling power line inspection image classification method based on frequency band statistics according to claim 1, characterized in that, The features to be processed described in S3 undergo discrete wavelet transform to decompose them into statistical features of multiple frequency band sub-bands, specifically: The features to be processed are reconstructed into a two-dimensional feature map, and the two-dimensional feature map is decomposed by discrete wavelet transform to obtain multiple frequency band sub-bands; Calculate the statistical characterization of each frequency domain sub-band, which includes: wavelet coefficient energy, mean, and variance.

6. The dual-path decoupling power line inspection image classification method based on frequency band statistics according to claim 1, characterized in that, The learnable frequency domain filtering weights described in S2 are applied by the formula: To achieve; In the formula, , These represent the learnable frequency domain filter weights and learnable parameters, respectively.

7. A dual-path decoupled power line inspection image classification system based on frequency band statistics, characterized in that, include: Preprocessing module: The power inspection images to be classified are divided into image blocks to obtain the features to be processed; Dual-path processing module: The features to be processed are modeled in the state space to generate the first processed features; The features to be processed are enhanced by frequency domain filtering to generate second processed features, specifically: After the features to be processed are subjected to Fast Fourier Transform, learnable frequency domain filtering weights are applied, and then the inverse Fourier Transform is used to generate the second processed features. Frequency band statistics module: The features to be processed are subjected to discrete wavelet transform to decompose them into statistical features of multiple frequency band sub-bands. For the statistical features of the frequency band sub-bands, after fully connected processing and constraint to a preset range, the first response allocation coefficient and the second response allocation coefficient are generated. Response coefficient modulation module: The first processed feature is multiplied by the first response allocation coefficient, and the second processed feature is multiplied by the second response allocation coefficient to obtain the corresponding modulation feature. Then, the modulated features are weighted and summed to generate the current feature. Classification module: The current feature is taken as the feature to be processed and enters the dual-path processing module. After the preset conditions are met, classification processing is performed to obtain the probability of the power component category.

8. The dual-path decoupling power line inspection image classification system based on frequency band statistics according to claim 7, characterized in that, In the frequency band statistics module, the statistical characteristics of the frequency band sub-bands, after being processed by a fully connected layer and constrained to a preset range, generate a first response allocation coefficient and a second response allocation coefficient, specifically as follows: The statistical features of each frequency band sub-band are concatenated to form a statistical vector, and then mapped through a fully connected layer to obtain two original response values ​​corresponding to state space modeling and frequency domain filtering enhancement, respectively. The original response values ​​are constrained by the Sigmoid function to obtain the first response allocation coefficient and the second response allocation coefficient.

9. The dual-path decoupling power line inspection image classification system based on frequency band statistics according to claim 7, characterized in that, In the classification module, reaching the preset condition means using backpropagation of the loss function to optimize the structural parameters of the state space modeling and frequency domain filtering enhancement processes in the dual-path processing module to a preset threshold.

10. The dual-path decoupling power line inspection image classification system based on frequency band statistics according to claim 7, characterized in that, In the dual-path processing module, the application of learnable frequency domain filtering weights is done using the formula: To achieve; In the formula, , These represent the learnable frequency domain filter weights and learnable parameters, respectively.

Images