Small target detection method and system for infrared image of power equipment

Abstract

The invention relates to a small target detection method and system for an infrared image of power equipment, and the method comprises the following steps: collecting a data sample of the infrared image of the power equipment, carrying out the image preprocessing, and constructing a data set; training the improved YOLOv8-EED model by using the data set, and optimizing and adjusting model parameters to obtain a target detection model; and pre-processing the current infrared image of the power equipment, inputting the pre-processed infrared image into the target detection model, and outputting to obtain a corresponding detection result. According to the improved YOLOv8-EED model, a small target enhancement layer is additionally arranged, a shallow large feature map rich in details and an up-sampled deep semantic feature map are spliced and fused to construct a feature extraction structure, the detail expression ability of the small target can be improved, and a deformable convolution detection head is adopted to adaptively capture fine features of the small target in a complex scene. Compared with the prior art, the method can achieve the efficient and precise detection of the small target of the infrared power equipment.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and deep learning technology, and in particular to a small target detection method and system for infrared images of power equipment. Background Art

[0002] As power systems continue to evolve and the number of transmission equipment continues to increase, ensuring its smooth operation is crucial. Power equipment is often exposed at high altitudes, and some are located in remote areas, making manual inspections challenging. Traditional manual inspections using image analysis are not only cumbersome and inefficient, but also prone to missed anomalies due to fatigue and negligence, resulting in high costs.

[0003] At present, the use of drones and infrared thermal imaging technology for inspections has gradually become an important mode of power inspection operations. Infrared image detection technology has been widely used in the monitoring and diagnosis of power equipment due to its advantages such as non-contact and fast speed. The infrared thermal imaging equipment carried by drones can not only quickly scan long-distance lines, but also capture the thermal status of distribution equipment in a non-contact manner, thereby timely discovering potential faults, problems and abnormalities, and realizing live detection. This innovative inspection method has greatly improved the convenience and efficiency of distribution equipment inspections.

[0004] Although drones combined with infrared thermal imaging technology have become an important tool for power inspections, using infrared images for power equipment inspection still faces numerous challenges. This is primarily due to the fact that, compared to visible light images of ordinary targets, infrared images of power equipment have complex backgrounds, low contrast, similar target features, and a large aspect ratio. This results in poor infrared image detection and low resolution. Furthermore, power equipment is located in complex environments, such as forests, near buildings, or in densely populated areas. Furthermore, drones have variable shooting angles, heights, and distances, and most power equipment captured is small, significantly increasing the difficulty of detection. Existing research has used deep learning models for target detection, such as YOLOv8n. While this model has achieved some success in detecting small targets, it still suffers from insufficient granularity in expressing high-level semantic features when dealing with extremely small targets such as power equipment, leading to missed detections and false detections. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a small target detection method and system for infrared images of power equipment. By using the improved YOLOv8-EED model, efficient and accurate detection of small targets of infrared power equipment can be achieved.

[0006] The object of the present invention can be achieved by the following technical solution: A small target detection method for infrared images of power equipment, comprising the following steps:

[0007] S1. Collect infrared image data samples of power equipment, perform image preprocessing, and construct a data set;

[0008] S2. Use the dataset to train the improved YOLOv8-EED model and obtain the target detection model by optimizing and adjusting the model parameters.

[0009] S3. Pre-process the current infrared image of the power equipment and input it into the target detection model to output the corresponding detection result.

[0010] Furthermore, the image preprocessing in step S1 includes image screening, cropping, normalization and denoising operations, wherein the image screening specifically screens the infrared image to remove blurred, damaged or irrelevant images of the power equipment;

[0011] The image cropping is specifically to crop the filtered image, remove irrelevant parts of the image edge, and only retain the area where the power equipment is located;

[0012] The image normalization specifically maps the pixel values ​​of the cropped image to the interval [0, 1];

[0013] The image denoising specifically involves using a filtering algorithm to remove noise from an image.

[0014] Furthermore, the improved YOLOv8-EED model specifically introduces an EMA parallel attention module at the end of the YOLOv8 backbone network to dynamically enhance the channel and spatial weights of the target area;

[0015] The EC2f module is used in the neck network of YOLOv8 for feature fusion to reduce the loss of position information and enhance the ability to capture subtle features of small targets;

[0016] A dynamic deformable convolutional detection head is used in the head network of YOLOv8 to deal with target overlap and occlusion problems.

[0017] Furthermore, the improved YOLOv8-EED model includes a backbone feature extraction unit, a neck feature fusion unit, and a detection unit connected in sequence. The neck feature fusion unit is provided with a plurality of small target feature enhancement modules UCE, and the small target feature enhancement module UCE includes an upsampling module Upsample, a concatenation module Concat, and an EC2f module connected in sequence.

[0018] The detection unit is provided with a plurality of DyHead-D3 detection heads, and the DyHead-D3 detection heads are provided with a scale perception attention module, a spatial position perception module and a task perception module.

[0019] Furthermore, the working process of the small target feature enhancement module UCE is as follows: the shallow large feature map Fshallow large is fused with the deep feature map Fdeep. The fusion process is as follows: first, upsample Fdeep by an upsampling multiple of M, and obtain Fdeep by bilinear interpolation. 深上采样 ;

[0020] Then F 浅大 With F 深上采样 Through the concatenation operation Concat fusion, F 融合2 =F 浅大 ,F 深上采样 ;

[0021] After the EC2f convolution operation, the enhanced feature map is obtained:

[0022]

[0023] Among them, i and j are the coordinates of the output feature map, K EC2f is the convolution kernel parameter, β is the bias term, and m and n are the position offsets of the convolution kernel sliding.

[0024] Furthermore, the scale-aware attention module specifically weights features of different scales through the scale-aware attention mechanism. Assuming the dimension of the feature tensor F is L×S×C, the scale-aware attention function is applied to F to capture and process information related to targets of different scales. The calculation process is: first, the linear function f(x) is approximated by 1×1 convolution, and the activation function σ(x) uses high-sigmoid. The output of the scale-aware attention module is:

[0025] F 尺度 =σ(f(F))·π L (F)

[0026]

[0027] Among them, π L is the scale-aware attention function, w l is the scale-related weight, F l Features of different scales.

[0028] Furthermore, the spatial position perception module specifically uses deformable convolution to perform sparse processing on the attention learning process. Assume that the number of sparse sampling positions is K and the spatial offset of self-learning is p k +Δp k , for the initial position Δp k Adjust, position p k The self-learning importance scalar is Δm k , extracted from the middle layer of feature F, and after being operated by the spatial position perception module, the output is:

[0029]

[0030] Furthermore, the task perception module specifically introduces an intelligent control mechanism to dynamically switch the on and off states of the feature channels according to different task requirements. Let the task requirement vector be T and the feature channel set be C. 集合 , using the channel control function based on the threshold judgment, if the value of an element in the task demand vector T is greater than the set threshold θ, then C 集合 The corresponding channel is opened, otherwise it is closed.

[0031] Furthermore, step S2 includes the following steps:

[0032] S21. Use the dataset to construct and train the feature fusion network: build the feature fusion network architecture and determine the weight coefficient for the fusion of large feature maps and backbone feature maps;

[0033] According to the size and characteristics of the small target of the power equipment, the convolution kernel size, number and step size are set, and the multiple and method of the upsampling operation are set;

[0034] Set the convolution kernel parameters and bias terms for the EC2f operation;

[0035] The preprocessed dataset is divided into a training set and a validation set. The Adam optimizer is selected and the initial learning rate is set. During the training process, the model parameters of the feature fusion network architecture are adjusted according to the validation set indicators. After multiple rounds of iterative training, the model learns the feature representation of small infrared targets and obtains a trained feature fusion network.

[0036] S22. Build and optimize a dynamic deformable convolutional detection head using the dataset: Use Python combined with a deep learning framework to write code and design a scale-aware attention mechanism;

[0037] The attention learning process is sparsely processed using deformable convolution. The number of sparse sampling positions, the self-learning spatial offset, and the position self-learning importance scalar parameters are set to build a spatial position perception module.

[0038] For task perception, we write code logic to dynamically switch the feature channel on and off. By setting thresholds and control functions, we can achieve flexible allocation of feature channel resources.

[0039] Implement the deep cross-talk method, self-masking operation, and tri-BCE loss weight calculation method of the DCNV3 (Deformable Convolution v3) module in the code, fuse the Dynamic Head and DCNV3 into the DyHead-D3 detection head, adjust the parameters and replace the original detection head, integrate and debug the entire model, and obtain the optimized DyHead-D3 detection head model;

[0040] S23. Based on the YOLOv8n algorithm, combined with the trained feature fusion network and the optimized DyHead-D3 detection head model, the YOLOv8-EED model was constructed. During the YOLOv8-EED model training process, different convolution kernel sizes and step sizes were tried. The number of training rounds was determined based on the model convergence and training time. The early stopping method was used to avoid overtraining. Finally, the target detection model was trained.

[0041] A small target detection system for infrared images of power equipment, comprising:

[0042] Image preprocessing module, used to perform image preprocessing operations on infrared images of power equipment, including image screening, cropping, normalization and denoising operations;

[0043] The large feature fusion module is used to splice and fuse the shallow large feature map in the infrared image with the upsampled deep semantic feature map to obtain enhanced features;

[0044] Dynamic deformable convolution module, used to perform scale-aware, spatial-aware, and task-aware operations on infrared images to obtain subtle features;

[0045] The model building and optimization module is used to combine the large feature fusion module and the dynamic deformable convolution module to train the target detection model using the dataset;

[0046] The verification module is used to receive the target detection results output by the target detection model and evaluate and provide feedback on the detection accuracy and robustness.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] This paper uses a constructed dataset to train an improved YOLOv8-EED model and optimizes model parameters to obtain a target detection model. The improved YOLOv8-EED model introduces an EMA parallel attention module at the end of the YOLOv8 backbone network to dynamically enhance the channel and spatial weights of the target area. The EC2f module is used in the YOLOv8 neck network for feature fusion, reducing position information loss and enhancing the ability to capture subtle features of small targets. Finally, a dynamic deformable convolutional detection head is used in the YOLOv8 head network to address target overlap and occlusion. This enables efficient and accurate detection of small targets in infrared power equipment.

[0049] The present invention sets up multiple small target feature enhancement modules UCE in the neck feature fusion unit of the YOLOv8-EED model. The small target feature enhancement module UCE includes an upsampling module Upsample, a splicing module Concat and an EC2f module connected in sequence to fuse, splice and convolve shallow large features with deep feature maps, thereby obtaining enhanced features, which can improve the ability to express small target details and help improve the accuracy of small target detection.

[0050] The present invention sets multiple DyHead-D3 detection heads (dynamically deformable convolutional detection heads) in the detection unit of the YOLOv8-EED model. The DyHead-D3 detection head integrates scale perception, spatial perception and task perception attention mechanisms, can adaptively capture the subtle features of small targets in complex scenes, can accurately identify targets at different scales, shapes, positions and viewing angles, and can process various representation forms of targets. The present invention adopts the DyHead-D3 detection head to add a branch for feature fusion to enhance the information step capability, and comprehensively improves the detection capability of small targets in complex working conditions from the three dimensions of scale perception, spatial perception and task perception.

[0051] The improved YOLOv8-EED model proposed in the present invention optimizes the feature fusion network structure based on YOLOv8, removes redundant detection heads, and adopts a lightweight splicing strategy, that is, retaining the original Concat module of the feature fusion part. The Concat module directly splices the feature map, completely retains the feature information, reduces the amount of calculation, and ensures the real-time performance of the model in an environment with limited hardware resources. In this way, the computational complexity is reduced while ensuring the detection accuracy, thereby achieving efficient and accurate detection of small targets of infrared power equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of the method flow of the present invention;

[0053] Figure 2 Schematic diagram of the application process of the embodiment;

[0054] Figure 3 Schematic diagram of the network architecture of the improved YOLOv8-EED model;

[0055] Figure 4 This is a schematic diagram of the network architecture of the DyHead-D3 detection head;

[0056] Figure 5 Schematic diagram of target detection results in the embodiment;

[0057] Figure 6a Schematic diagram of the experimental results of target detection robustness under Gaussian noise in the embodiment;

[0058] Figure 6b Schematic diagram of the experimental results of target detection robustness under salt and pepper noise in the embodiment;

[0059] Figure 6c Schematic diagram of the experimental results of target detection robustness under luminance noise in the embodiment. DETAILED DESCRIPTION

[0060] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] Example

[0062] like Figure 1 As shown, a small target detection method for infrared images of power equipment includes the following steps:

[0063] S1. Collect infrared image data samples of power equipment, perform image preprocessing, and construct a data set;

[0064] S2. Use the dataset to train the improved YOLOv8-EED model and obtain the target detection model by optimizing and adjusting the model parameters.

[0065] S3. Pre-process the current infrared image of the power equipment and input it into the target detection model to output the corresponding detection result.

[0066] Based on the above method, a small target detection system for infrared images of power equipment is built, including:

[0067] Image preprocessing module, used to perform image preprocessing operations on infrared images of power equipment, including image screening, cropping, normalization and denoising operations;

[0068] The large feature fusion module is used to splice and fuse the shallow large feature map in the infrared image with the upsampled deep semantic feature map to obtain enhanced features;

[0069] Dynamic deformable convolution module, used to perform scale-aware, spatial-aware, and task-aware operations on infrared images to obtain subtle features;

[0070] The model building and optimization module is used to combine the large feature fusion module and the dynamic deformable convolution module to train the target detection model using the dataset;

[0071] The verification module is used to receive the target detection results output by the target detection model and evaluate and provide feedback on the detection accuracy and robustness.

[0072] This embodiment applies the above solution, such as Figure 2 As shown, the main process includes:

[0073] 1. Data collection and preprocessing: Collect infrared images of power equipment and perform data collection and preprocessing operations, including image screening, cropping, normalization and denoising operations.

[0074] 1.1): The image screening operation screens the large number of collected infrared images, removes blurry, damaged or irrelevant images of power equipment, and processes the filtered images for the next step;

[0075] 1.2): The image cropping operation crops the filtered image, removes irrelevant parts of the image edge, and retains only the area where the power equipment is located. The cropped image is then normalized;

[0076] 1.3): Image normalization maps the pixel values ​​of the cropped image to the interval [0, 1] to enhance the comparability between images. After completion, denoising is performed;

[0077] 1.4): The image denoising operation uses a filtering algorithm to remove noise from the image, improve image quality, and pass the preprocessed image to the feature fusion network module and the detection head module.

[0078] 2. Establish a feature fusion network module based on the large feature fusion mechanism, receive the preprocessed image data, and perform feature fusion network construction and training operations.

[0079] 2.1): Based on a deep learning framework (such as PyTorch), the network is built. According to the designed neck network architecture, a feature fusion network based on the large feature map fusion mechanism is built, and the weight coefficient of the fusion of the large feature map and the backbone feature map is determined;

[0080] 2.2): Set the convolution parameters. According to the size and characteristics of the small target of the power equipment, select the appropriate convolution kernel size, number and step size, and set the multiple and method of the upsampling operation;

[0081] 2.3) Design the convolution kernel parameters and bias terms for the EC2f operation to enhance the representation of small object features, and then train the constructed network;

[0082] 2.4): The preprocessed dataset is divided into a training set and a validation set in a ratio of 7:3. The Adam optimizer is selected and the initial learning rate is set. During the training process, the model parameters are adjusted according to the validation set indicators. After multiple rounds of iterative training, the model learns the feature representation of small infrared targets. The relevant results of the trained feature fusion network model are passed to the model construction and optimization module.

[0083] 3. Establish a dynamic deformable convolution detection head module, receive the preprocessed image data, and perform construction and optimization operations of the dynamic deformable convolution detection head.

[0084] 3.1): Based on the design principles of scale perception, spatial perception, and task perception, use Python combined with a deep learning framework to write code and design a scale-aware attention mechanism;

[0085] The attention learning process is sparsely processed using deformable convolution. Parameters such as the number of sparse sampling positions, the self-learning spatial offset, and the position self-learning importance scalar are set to build a spatial position perception module.

[0086] For task perception, we write code logic to dynamically switch the feature channel on and off. By setting thresholds and control functions, we can achieve flexible allocation of feature channel resources.

[0087] 3.2) Implement the DCNV3 module's deep crossover method, self-masking operation, and tri-BCE loss weight calculation method in the code. Fusion the Dynamic Head and DCNV3 into the DyHead-D3 detection head. Adjust the parameters and replace the original detection head. Integrate and debug the entire model. Pass the relevant parameters and detection results of the optimized detection head model to the model construction and optimization module.

[0088] 4. Receive the results of the feature fusion network module and the detection head module, and perform model construction and optimization operations.

[0089] 4.1): Model building operation, based on the YOLOv8n algorithm, build the YOLOv8-EED model (such as Figure 3 As shown in Figure 2, an EMA parallel attention module is introduced at the end of the YOLOv8 backbone network to dynamically enhance the channel and spatial weights of the target area;

[0090] The EC2f module is used in the neck network of YOLOv8 for feature fusion to reduce the loss of position information and enhance the ability to capture subtle features of small targets;

[0091] A dynamic deformable convolutional detection head is used in the head network of YOLOv8 to handle target overlap and occlusion issues;

[0092] The improved YOLOv8-EED model includes a backbone feature extraction unit, a neck feature fusion unit, and a detection unit connected in sequence. The neck feature fusion unit is equipped with multiple small target feature enhancement modules (UCEs). The small target feature enhancement module (UCE) includes an upsampling module (Upsample), a concatenation module (Concat), and an EC2f module connected in sequence, which restores the feature fusion part to the original Concat module.

[0093] The working process of the above small target feature enhancement module UCE is as follows: the shallow large feature map F 浅大 With the deep feature map F 深 The fusion process is as follows: first, F 深 Perform upsampling operation Upsample, the upsampling multiple is M, and use bilinear interpolation to get F 深上采样 ;

[0094] Then F 浅大 With F 深上采样 Through the concatenation operation Concat fusion, F 融合2 =F 浅大 ,F 深上采样 ;

[0095] After the EC2f convolution operation, the enhanced feature map is obtained:

[0096]

[0097] Among them, i and j are the coordinates of the output feature map, K EC2f is the convolution kernel parameter, β is the bias term, and m and n are the position offsets of the convolution kernel sliding.

[0098] In addition, the detection unit is equipped with multiple DyHead-D3 detection heads. The DyHead-D3 detection heads are equipped with a scale-aware attention module, a spatial position perception module, and a task perception module. The higher-resolution and more detailed shallow feature maps output by the UCE are used as the input of the DyHead-D3 detection head, which can enhance the detection sensitivity and accuracy of extremely small targets.

[0099] like Figure 4 As shown in Figure 2, in terms of scale perception, the scale-aware attention mechanism is used to weight features of different scales. Assume that the dimension of the feature tensor F is L×S×C, and the scale-aware attention function π LActing on F, it captures and processes information related to targets of different scales. The calculation process is: first, the linear function f(x) is approximated by 1×1 convolution, and the activation function σ(x) uses high-sigmoid. The output of the scale-aware attention module is: F 尺度 =σ(f(F))·π L (F), where (w l is the scale-related weight, F l for different scale features).

[0100] Spatial perception: Design a spatial position perception module and use deformable convolution to perform sparse processing on the attention learning process. Let the number of sparse sampling positions be K and the spatial offset of self-learning be p k +Δp k , for the initial position Δp k Adjust. Position p k The self-learning importance scalar is Δm k , extracted from the middle layer of feature F. After the spatial position perception attention module is operated, the features are sparse, focusing on the foreground target, improving the detection accuracy and efficiency. The output F of the spatial position perception module 空间 The calculation formula is:

[0101]

[0102] Task perception: Introduce an intelligent control mechanism to dynamically switch the open and closed states of feature channels according to different task requirements. Let the task requirement vector be T and the feature channel set be C 集合 , the channel control function is control function 1, then the characteristic channel C after control 调控 For: C 调控 = Control function 1 (C_{set}), T, where control function 1 is based on threshold judgment. If the value of an element in the task requirement vector T is greater than the set threshold θ, the corresponding channel is opened, otherwise it is closed.

[0103] DCNV3 module: DCNV3 introduces a deep cross-talk method. As the number of network layers n increases, the feature interaction order increases exponentially, which can be expressed as: Feature interaction order = 2 n At the same time, a self-masking operation is proposed to reduce the number of model parameters and filter noise. In the fusion layer, the tri-bce loss weight calculation method is used to provide supervision signals for different interaction methods. Assume that the loss weights of different interaction methods are w1, w2, and w3 respectively, and the tri-bce calculation method is:

[0104]

[0105] Fusion of Dynamic Head and DCNV3: The DCNV2 module in Dynamic Head is improved, and DCNV3 is used as a new module to fuse with Dynamic Head to form the DyHead-D3 detection head. In memory-sensitive scenarios, DCNV3's memory access mode is simple and direct, and its stability is good. When processing tasks that require clear probability distribution characteristics, its softmax normalization operation can make the results more interpretable.

[0106] 4.2): Model optimization operation. During the model training process, different convolution kernel sizes and step sizes are tried. The appropriate number of training rounds is determined based on the model convergence and training time. Early stopping is used to avoid overtraining. The optimized YOLOv8-EED model is passed to the verification module.

[0107] 5. The verification module performs experimental verification operations on the optimized YOLOv8-EED model.

[0108] 5.1) Ablation experiment: train the original model, the model with the UCE module added, the model with the Dyhead-D3 detector head replaced, and the model with both modules added. Train in the same environment and evaluate the average precision, recall, and other indicators of each model on the validation set.

[0109] 5.2) Model comparison experiments were conducted, testing the YOLOv8-EED model with mainstream object detection models such as Faster-RCNN, SSD, and YOLOv5-n, as well as other improved models in the same field, on the same dataset. The detection accuracy, recall rate, F1 value, average precision, and frame rate of each model were recorded.

[0110] 5.3) Robustness experiment: Add different levels of Gaussian noise and salt and pepper noise to the dataset and adjust the brightness. Use the YOLOv8-EED model to detect on the processed dataset, record the detection accuracy under different noise conditions, and analyze the experimental results to evaluate the model performance.

[0111] In this embodiment, experiments are conducted using a dataset of small target detection in infrared images of power equipment, and the effectiveness of the model is verified through ablation experiments, model comparison experiments, and robustness experiments.

[0112] Ablation experiment: Compare the experimental results after removing or modifying specific network structures, such as adding a small target enhancement (UCE) module, replacing the detection head with Dyhead-D3, etc., to analyze the impact of each improvement on the small target detection performance. Let the original model be M0 and the model after adding the UCE module be M UCE , the model after replacing the detection head with Dyhead-D3 is M Dyhead-D3 , add both models as M两者 By calculating indicators such as mean average precision (mAP50) and recall rate (Recall), the performance changes of the model are evaluated. The calculation formula is:

[0113]

[0114] (AP i is the average precision of each category, N is the number of categories),

[0115]

[0116] Model comparison experiment: Compared with mainstream target detection models such as Faster-RCNN, SSD, YOLOv5-n and other improved models in the same field, the model performance is analyzed in terms of detection accuracy (Precision), recall rate (Recall), F1 value, mean average precision (mAP50) and frame rate (FPS), and the model performance is evaluated.

[0117]

[0118] Robustness experiment: Gaussian noise and salt and pepper noise were added to the dataset, and the brightness was adjusted to simulate real-world interference factors. The accuracy of the model in identifying small targets under complex conditions was evaluated to verify the robustness of the model.

[0119] Specific experimental results such as Figure 5 and Figures 6a to 6c As shown, in order to systematically evaluate the robust performance of the experimental model constructed in this chapter in a complex environment, a multi-dimensional noise injection and interference simulation experiment is performed on the data set. This embodiment uses Gaussian noise (mean range 0-0.1, variance interval 0.01-0.05) to simulate circuit thermal noise and sensor signal interference, and uses salt and pepper noise (density parameter 0.01-0.05) to simulate the random bit error phenomenon during data transmission. At the same time, a dynamic brightness adjustment mechanism is introduced to simulate the significant changes in lighting conditions in actual detection scenes. The above interference processing strictly follows the ISO 15469 image quality assessment standard to ensure the repeatability of experimental conditions and the comparability of results. As Figure 5 As shown ( Figure 5 (a) corresponds to Gaussian noise σ = 0.001; Figure 5 (b) corresponds to Gaussian noise σ = 0.01; Figure 5 (c) corresponds to salt and pepper noise p = 0.01; Figure 5 (d) corresponds to salt and pepper noise p = 0.1; Figure 5 (e) corresponds to brightness L = 0.5; Figure 5(f) corresponds to brightness L = 1.5). The image sample after noise processing clearly shows the coupling changes between the target features and background interference. All processing parameters and operation procedures are strictly consistent with the previous chapter. By constructing a controllable interference environment, a scientific basis is provided for quantitatively analyzing the detection accuracy attenuation law and robustness performance of the model under degraded image conditions.

[0120] When conducting robustness experiments, Figures 6a to 6c (The horizontal axis in the figure is the noise value, the vertical axis is the accuracy, and the data includes a high-voltage 230V switch (Chave_H_230kV), a vertical 230V switch (Chave_V_230kV), a circuit breaker (DISJUNTOR), a lightning arrester with a rated voltage of 230 kilovolts (Pararraio_230kV), a lightning arrester with a rated voltage of 69 kilovolts (Pararraio_69kV), a switch current transformer (TC), and a coupling capacitor (TP)) It can be concluded that under conditions of constantly changing noise, the model has a relatively small loss in accuracy in identifying small targets, which indicates that the YOLOv8-EED model performs well in actual scenarios when facing weather, temperature changes, and rainy days.

[0121] 6. Result determination: Perform result determination operations based on the experimental results of the verification module.

[0122] 6.1): Indicator judgment operation, judging whether the YOLOv8-EED model meets the set performance indicator requirements in terms of small target detection accuracy, recall rate, robustness, etc.;

[0123] 6.2): ​​Result processing operation: If the requirements are met, the model is considered valid and can be applied to the actual small target detection task of substation power equipment; if not, it is necessary to readjust the model parameters or improve the model structure, and conduct experimental verification again until satisfactory performance indicators are achieved.

[0124] In summary, this solution adds a small target enhancement layer, concatenating and fusing shallow, detail-rich large feature maps with upsampled deep semantic feature maps to construct a feature extraction structure. Furthermore, a deformable convolutional detection head is used to add a branch to feature fusion to enhance information flow. The small target enhancement layer is used to implement a feature fusion network that fuses large features, enhancing the ability to capture subtle features of small infrared targets, reducing position information loss, and improving the accuracy of small target detection. Based on a dynamic deformable convolutional detection head, the model's detection capabilities for small targets in complex working conditions are comprehensively improved from three dimensions: scale perception, spatial perception, and task perception, effectively addressing target overlap and occlusion.

[0125] The improved YOLOv8-EED model proposed in this solution improves detection speed and reduces detection latency while maintaining detection accuracy. It also exhibits good adaptability in environments with limited hardware resources. A series of experiments have verified that the proposed method performs well in small target detection accuracy, recall rate, and other indicators, and is highly robust, better meeting the actual needs of power equipment detection.

Claims

1. A small target detection method for infrared images of power equipment, characterized in that: The following steps are involved: S1. Collect infrared image data samples of power equipment, perform image preprocessing, and construct a data set; S2. Use the dataset to train the improved YOLOv8-EED model and obtain the target detection model by optimizing and adjusting the model parameters. S3. Pre-process the current infrared image of the power equipment and input it into the target detection model to output the corresponding detection result.

2. The small target detection method for infrared images of power equipment according to claim 1 is characterized in that: The image preprocessing in step S1 includes image screening, cropping, normalization and denoising operations. The image screening specifically screens the infrared image to remove blurry, damaged or irrelevant images of the power equipment. The image cropping is specifically to crop the filtered image, remove irrelevant parts of the image edge, and only retain the area where the power equipment is located; The image normalization specifically maps the pixel values ​​of the cropped image to the interval [0, 1]; The image denoising specifically involves using a filtering algorithm to remove noise from an image.

3. The small target detection method for infrared images of power equipment according to claim 1 is characterized in that: The improved YOLOv8-EED model specifically introduces an EMA parallel attention module at the end of the YOLOv8 backbone network to dynamically enhance the channel and spatial weights of the target area; The EC2f module is used in the neck network of YOLOv8 for feature fusion to reduce the loss of position information and enhance the ability to capture subtle features of small targets; A dynamic deformable convolutional detection head is used in the head network of YOLOv8 to deal with target overlap and occlusion problems.

4. The small target detection method for infrared images of power equipment according to claim 3 is characterized in that: The improved YOLOv8-EED model includes a backbone feature extraction unit, a neck feature fusion unit, and a detection unit connected in sequence. The neck feature fusion unit is provided with a plurality of small target feature enhancement modules UCE. The small target feature enhancement module UCE includes an upsampling module Upsample, a concatenation module Concat, and an EC2f module connected in sequence. The detection unit is provided with a plurality of DyHead-D3 detection heads, and the DyHead-D3 detection heads are provided with a scale perception attention module, a spatial position perception module and a task perception module.

5. The small target detection method for infrared images of power equipment according to claim 4 is characterized in that: The working process of the small target feature enhancement module UCE is as follows: 浅大 With the deep feature map F 深 The fusion process is as follows: first, F 深 Perform upsampling operation Upsample, the upsampling multiple is M, and use bilinear interpolation to get F 深上采样 ; Then F 浅大 With F 深上采样 Through the concatenation operation Concat fusion, F 融合2 =F 浅大 ,F 深上采样 ; After the EC2f convolution operation, the enhanced feature map is obtained: Among them, i and j are the coordinates of the output feature map, K EC2f is the convolution kernel parameter, β is the bias term, and m and n are the position offsets of the convolution kernel sliding.

6. The small target detection method for infrared images of power equipment according to claim 4 is characterized in that: The scale-aware attention module specifically uses the scale-aware attention mechanism to weight features of different scales. Assuming the dimension of the feature tensor F is L×S×C, the scale-aware attention function is applied to F to capture and process information related to targets of different scales. The calculation process is as follows: first, the linear function f(x) is approximated by 1×1 convolution, and the activation function σ(x) uses high-sigmoid. The output of the scale-aware attention module is: F 尺度 =σ(f(F))·π L (F) Among them, π L is the scale-aware attention function, w l is the scale-related weight, F l Features of different scales.

7. The small target detection method for infrared images of power equipment according to claim 4 is characterized in that: The spatial position perception module specifically uses deformable convolution to perform sparse processing on the attention learning process. Assume that the number of sparse sampling positions is K and the spatial offset of self-learning is p k +Δp k , for the initial position Δp k Adjust, position p k The self-learning importance scalar is Δm k , extracted from the middle layer of feature F, and after being operated by the spatial position perception module, the output is:

8. The small target detection method for infrared images of power equipment according to claim 4 is characterized in that: The task perception module specifically introduces an intelligent control mechanism to dynamically switch the open and closed states of feature channels according to different task requirements. Let the task requirement vector be T and the feature channel set be C. 集合 , using the channel control function based on the threshold judgment, if the value of an element in the task demand vector T is greater than the set threshold θ, then C 集合 The corresponding channel is opened, otherwise it is closed.

9. The small target detection method for infrared images of power equipment according to claim 4, characterized in that: The step S2 comprises the following steps: S21. Use the dataset to construct and train the feature fusion network: build the feature fusion network architecture and determine the weight coefficient for the fusion of large feature maps and backbone feature maps; According to the size and characteristics of the small target of the power equipment, the convolution kernel size, number and step size are set, and the multiple and method of the upsampling operation are set; Set the convolution kernel parameters and bias terms for the EC2f operation; The preprocessed dataset is divided into a training set and a validation set. The Adam optimizer is selected and the initial learning rate is set. During the training process, the model parameters of the feature fusion network architecture are adjusted according to the validation set indicators. After multiple rounds of iterative training, the model learns the feature representation of small infrared targets and obtains a trained feature fusion network. S22. Build and optimize a dynamic deformable convolutional detection head using the dataset: Use Python combined with a deep learning framework to write code and design a scale-aware attention mechanism; The attention learning process is sparsely processed using deformable convolution. The number of sparse sampling positions, the self-learning spatial offset, and the position self-learning importance scalar parameters are set to build a spatial position perception module. For task perception, we write code logic to dynamically switch the feature channel on and off. By setting thresholds and control functions, we can achieve flexible allocation of feature channel resources. Implement the deep cross-talk method, self-masking operation, and tri-BCE loss weight calculation method of the DCNV3 module in the code, fuse the Dynamic Head and DCNV3 into the DyHead-D3 detection head, adjust the parameters and replace the original detection head, integrate and debug the entire model, and obtain the optimized DyHead-D3 detection head model; S23. Based on the YOLOv8n algorithm, combined with the trained feature fusion network and the optimized DyHead-D3 detection head model, the YOLOv8-EED model was constructed. During the YOLOv8-EED model training process, different convolution kernel sizes and step sizes were tried. The number of training rounds was determined based on the model convergence and training time. The early stopping method was used to avoid overtraining. Finally, the target detection model was trained.

10. A small target detection system for infrared images of power equipment, characterized in that: include: Image preprocessing module, used to perform image preprocessing operations on infrared images of power equipment, including image screening, cropping, normalization and denoising operations; The large feature fusion module is used to splice and fuse the shallow large feature map in the infrared image with the upsampled deep semantic feature map to obtain enhanced features; Dynamic deformable convolution module, used to perform scale-aware, spatial-aware, and task-aware operations on infrared images to obtain subtle features; The model building and optimization module is used to combine the large feature fusion module and the dynamic deformable convolution module to train the target detection model using the dataset; The verification module is used to receive the target detection results output by the target detection model and evaluate and provide feedback on the detection accuracy and robustness performance.

Images