A power equipment target detection method, system and device for intelligent inspection

Abstract

A power equipment target detection method for intelligent inspection, comprising: acquiring an inspection image; inputting the inspection image into an improved Yolov8 target detection model to obtain a power equipment target detection result; wherein the improved Yolov8 target detection model comprises a backbone network, a local non-attention feature enhancement module and a detection head; the backbone network is used for extracting multi-scale features of a to-be-detected image and fusing the multi-scale features to obtain a first feature map containing multi-scale feature information; the local non-attention feature enhancement module is used for performing feature enhancement on the first feature map to obtain a second feature map; and the detection head is used for outputting a detection result according to the second feature map. The design fuses a local non-attention feature enhancer and a YOLOv8 target detection algorithm, significantly improves the recognition ability for small equipment and equipment located in a complex background, and improves the sensitivity of the target detection algorithm to micro changes of power equipment.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method, system and device for detecting targets in power equipment for intelligent inspection. Background Technology

[0002] As data centers continue to expand in scale and increase in business complexity, ensuring the reliability and security of data center equipment has become increasingly important.

[0003] In the daily operation and maintenance of power facilities, inspection is crucial for ensuring the stable operation of data center equipment. However, current inspection work mainly relies on regular manual checks or automated monitoring systems based on preset rules. These methods are often inefficient due to their time-consuming nature, and are particularly inadequate in real-time response to subtle changes in equipment status, resulting in insufficient guarantee of efficiency and accuracy. The expansion of data center scale further highlights the limitations of these traditional inspection methods, urgently requiring a more efficient and intelligent solution.

[0004] Driven by this demand, the exploration and development of innovative intelligent inspection technologies can not only improve inspection efficiency and accuracy but also enable real-time monitoring and early warning of equipment status, providing a solid guarantee for the stable and reliable operation of data centers. Against this backdrop, YOLOv8, as a leading object detection algorithm, has received widespread attention for its high processing speed and accurate recognition capabilities. This algorithm can quickly and accurately identify multiple targets in complex images while reducing computational resource consumption, demonstrating its broad potential and practical value in multiple application fields. However, applying YOLOv8 to intelligent inspection of data center equipment still faces a series of challenges, including the detection of small targets in complex backgrounds, the limitations of feature representation capabilities, the trade-off between real-time performance and computational efficiency, and adaptability to dynamic environments. When subtle changes occur in facility status, these systems may miss the optimal time for handling due to failure to identify or misjudgment, posing a potential risk to the stable operation of the power system. To address these challenges, this invention proposes an improved YOLOv8 target detection algorithm that integrates a feature enhancer without attention (LAFFE) to optimize target detection performance in intelligent inspection systems and provide crucial support for subsequent equipment status analysis and fault diagnosis. Summary of the Invention

[0005] The purpose of this invention is to overcome the problem of insufficient real-time response to subtle changes in the state of power facilities during intelligent inspection in existing technologies, and to provide a power facility target detection method for intelligent inspection that is highly sensitive to subtle changes in power facilities and can accurately identify power facilities in complex backgrounds.

[0006] In a first aspect, embodiments of the present invention provide a method for detecting targets in power equipment for intelligent inspection, the detection method comprising:

[0007] Acquire inspection images;

[0008] The inspection images are input into the improved Yolov8 target detection model to obtain the target detection results of power equipment;

[0009] The improved Yolov8 object detection model includes:

[0010] The backbone network is used to extract multi-scale features from the inspection images and fuse the multi-scale features to obtain a first feature map containing multi-scale feature information.

[0011] A local non-attention feature enhancement module is used to enhance the features of the first feature map to obtain a second feature map;

[0012] The detection head is used to output the detection result based on the second feature map.

[0013] Feature enhancement of the first feature map includes:

[0014] The first feature map is divided into multiple image blocks, the relevance score of each image block is calculated, and attention enhancement is performed on each image block based on the relevance score.

[0015] Calculating the relevance score of the i-th image patch among the plurality of image patches includes:

[0016] Calculate the correlation value between the i-th image patch and every other image patch;

[0017] The correlation scores r of the i-th image patch are obtained by summing the correlation values ​​between the i-th image patch and each of the other image patches. i ;

[0018] Wherein, the correlation value r between the i-th image patch and the j-th image patch i,j Calculate according to the following formula, where i≠j;

[0019] r i,j =f(x) i ,x j )=φ(x i ) T ψ(x j ) T (1);

[0020] In the formula, φ(·) or ψ(·) is the embedding function, x i For the i-th image patch, x j Let j be the j-th image patch.

[0021] Attention enhancement for the i-th image patch among the plurality of image patches based on the relevance score includes:

[0022] Channel attention enhancement is performed on the i-th image patch to obtain the first image patch.

[0023]

[0024] In the formula, ⊙ represents the element-wise product, x i For the i-th image patch, γ c (·) represents the channel attention function, which includes a fully connected layer. The channel attention function assigns different channel attention weights to each feature channel of the image patch through adaptive learning.

[0025] For the first image block Spatial attention enhancement is performed to obtain the second image patch.

[0026]

[0027] In the formula, ⊙ represents the element-wise product, x i For the i-th image patch, δ s (·) represents the spatial attention function, which includes a convolutional layer and assigns different spatial attention weights to each location in the image patch through adaptive learning.

[0028] Based on the i-th image patch x i Second image block The relevance score r of the i-th image patch i Obtain the enhanced i-th image patch

[0029]

[0030] In the formula, λ is the learnable enhancement coefficient, and r i For the i-th image patch x i The relevance score.

[0031] The backbone network includes a backbone network, an FPN structure, and an SPP structure.

[0032] The extraction and fusion of multi-scale features from the inspection images includes:

[0033] Multi-scale features of inspection images are extracted using a backbone network;

[0034] The multi-scale features are fused using FPN to construct a multi-scale feature pyramid;

[0035] Spatial pooling of multi-scale feature pyramids is performed using the SPP structure to fuse features from different receptive fields.

[0036] The backbone network is either a ResNeXt-101 network or a CSPDarknetX network.

[0037] The detection results output based on the second feature map include:

[0038] Based on the second feature map, predict the location and category of each key point;

[0039] Generate prediction boxes and corresponding category information for target boxes based on the location and category of each key point;

[0040] The predicted bounding boxes are processed by NMS, and then filtered using a confidence threshold. The detection results are then output.

[0041] Secondly, embodiments of the present invention provide a power equipment target detection system for intelligent inspection, comprising:

[0042] The image acquisition unit is used to acquire inspection images;

[0043] The target detection unit is used to input the inspection image into the improved Yolov8 target detection model to obtain the target detection results of power equipment;

[0044] The improved Yolov8 object detection model includes:

[0045] The backbone network is used to extract multi-scale features from the image to be detected and fuse the multi-scale features to obtain a first feature map containing multi-scale feature information.

[0046] A local non-attention feature enhancement module is used to enhance the features of the first feature map to obtain a second feature map;

[0047] The detection head is used to output the detection results based on the second feature map.

[0048] Thirdly, embodiments of the present invention provide an electronic device, which includes a memory and a processor. The memory is used to store computer program code and transmit the computer program code to the processor.

[0049] The processor is configured to execute the power equipment target detection method for intelligent inspection as described above, according to the instructions in the computer program code.

[0050] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the power equipment target detection method for intelligent inspection described above.

[0051] This invention provides a method, system, electronic device, and storage medium for power equipment target detection in intelligent inspection. The method uses an improved YOLOv8 target detection model to detect targets in inspection images. By integrating a local attention-free feature enhancer algorithm into the existing YOLOv8 model, the model is optimized from two dimensions: feature enhancement and target detection. This improves the sensitivity and accuracy of the target detection model to detailed features of power facilities, thereby enhancing the efficiency and accuracy of intelligent inspection.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] 1. In the power facility target detection method for intelligent inspection of this invention, an improved Yolov8 target detection model is used for target detection. The improved Yolov8 target detection model includes a backbone network, a local non-attention feature enhancement model, and a detection head. The local non-attention feature enhancement module is used to enhance local features of the image, including attention enhancement and spatial attention enhancement, to more effectively capture subtle changes and fault signs in power facilities, providing richer and more discriminative feature representations, and improving the accuracy of target detection. Therefore, this design uses a local non-attention feature enhancement module to enhance local features of the image, thereby improving the accuracy of target detection.

[0054] 2. In the power facility target detection method for intelligent inspection of this invention, the local attention-free feature enhancement module divides the first feature map into multiple image blocks, calculates the correlation values ​​between the image blocks, and for any image block, sums the correlation values ​​between that image block and all other image blocks to obtain the correlation score of that image block. Based on the correlation score, channel attention enhancement and spatial attention enhancement are then applied. Through the above feature enhancement steps, the enhancement level of each image block can be dynamically adjusted using the attention-free mechanism. Furthermore, channel and spatial attention are used to refine the local features. The local attention-free feature enhancement module, through a global guidance + local enhancement approach, performs detailed processing of the feature map, more specifically improving the expressive power of local features and enhancing the model's ability to identify small targets in complex backgrounds. Therefore, in this design, the local attention-free feature enhancement module, through a global guidance + local enhancement approach, more specifically improves the expressive power of local features and enhances the model's ability to identify small targets in complex backgrounds.

[0055] 3. In the power facility target detection method for intelligent inspection of this invention, the local feature enhancement function of the local non-attention feature enhancement module is used to dynamically enhance the local features of the image. This can more effectively capture subtle changes and fault signs of power facilities, enabling the image to provide richer and more discriminative feature representations and make more accurate judgments. At the same time, the local non-attention feature enhancement module improves the model's adaptability to environmental changes, ensuring that the model can maintain high detection stability and accuracy even when factors such as lighting conditions and equipment location change. Therefore, in this design, the local non-attention feature enhancement module dynamically enhances the local features of the image, providing the image with richer and more discriminative feature representations, while improving the model's adaptability to environmental changes and enhancing the detection accuracy of the target detection model.

[0056] 4. In the power facility target detection method for intelligent inspection of this invention, the attention-free mechanism simplifies the feature processing process and reduces computational complexity, thereby helping the intelligent inspection system achieve efficient real-time operation. Therefore, this design reduces the computational complexity of target detection, has high detection efficiency, and helps the intelligent inspection system achieve efficient real-time operation.

[0057] 5. The power facility target detection system for intelligent inspection of the present invention includes: an image acquisition unit for acquiring inspection images; and a target detection unit for inputting the inspection images into an improved Yolov8 target detection model to obtain power facility target detection results. This system implements the steps of the power facility target detection method for intelligent inspection provided in the above technical solution. Therefore, this system simultaneously includes all the beneficial effects of the power facility target detection method for intelligent inspection provided in the above technical solution, which will not be repeated here. Attached Figure Description

[0058] Figure 1 This is a schematic flowchart of a power facility target detection method for intelligent inspection provided in an embodiment of the present invention.

[0059] Figure 2 This is a schematic flowchart of a method for feature enhancement of a first feature map provided in an embodiment of the present invention.

[0060] Figure 3 This is a schematic diagram of a process for enhancing attention to the i-th image block based on the relevance score, provided by an embodiment of the present invention.

[0061] Figure 4 This is a schematic diagram of a process for identifying a target in an image to be detected based on the second feature map, according to an embodiment of the present invention.

[0062] Figure 5This is a schematic diagram of the structure of an improved Yolov8 target detection model provided in an embodiment of the present invention.

[0063] Figure 6 This is a schematic diagram of a backbone network provided in an embodiment of the present invention.

[0064] Figure 7 This is a schematic diagram of a detection process for an improved Yolov8 target detection model provided in an embodiment of the present invention.

[0065] Figure 8 This is a schematic diagram of a power equipment target detection system for intelligent inspection provided in an embodiment of the present invention.

[0066] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0067] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] In existing technologies, data center inspections rely on regular manual checks or automated monitoring systems based on preset rules. However, these methods are often inefficient due to their time-consuming nature, and their real-time response to subtle changes in facility status is particularly inadequate. Furthermore, the increasing scale of data centers exacerbates the limitations of these traditional inspection methods, creating an urgent need for a more efficient and intelligent solution. Driven by this demand, the exploration and development of innovative intelligent inspection technologies can not only improve inspection efficiency and accuracy but also enable real-time monitoring and early warning of power facility status, providing a solid guarantee for the stable and reliable operation of data centers. Meanwhile, the rapid development of deep learning technology has brought theoretical and practical innovations to fields such as image recognition and sequence data analysis.

[0069] To improve the efficiency and accuracy of power equipment inspections, especially to meet the need for real-time response to subtle changes in equipment status, intelligent inspection methods based on machine vision and target detection technologies are gaining increasing attention. These methods can simulate the human visual system, automatically acquiring and processing images of power facilities and accurately identifying the facilities and their related components within those images. By analyzing facility status data in real time, intelligent inspection systems can respond rapidly to subtle changes in facility status, providing timely and accurate decision support for maintenance personnel.

[0070] To address the technical problems existing in related technologies, this invention provides a method for detecting targets in power facilities for intelligent inspection, such as... Figure 1 As shown, Figure 1This is a flowchart illustrating a power facility target detection method for intelligent inspection provided in an embodiment of the present invention. The method includes steps S1 to S2:

[0071] S1. Obtain inspection images.

[0072] S2. Input the inspection image into the improved Yolov8 target detection model to obtain the target detection results of power equipment.

[0073] Among them, power equipment refers to equipment related to power production and transmission and distribution. Power equipment can be power generation, transmission, transformation and distribution equipment, or auxiliary equipment such as servers, switches, monitoring equipment, UPS, and cooling equipment used in power system data centers.

[0074] YOLOv8 is an advanced single-stage object detection algorithm that achieves efficient and fast object detection with relatively low computational overhead. This invention improves upon the original YOLOv8 to enhance its performance in intelligent inspection scenarios. The improved YOLOv8 object detection model in this invention is a pre-trained object detection model based on existing technologies. Specifically, as... Figure 5 As shown, Figure 5 This is a schematic diagram of an improved Yolov8 target detection model provided in an embodiment of the present invention. The improved Yolov8 target detection model includes:

[0075] The backbone network is used to extract multi-scale features from the image to be detected and fuse the multi-scale features to obtain a first feature map containing multi-scale feature information.

[0076] A local non-attention feature enhancement module is used to enhance the features of the first feature map to obtain a second feature map;

[0077] The detection head is used to output the detection results based on the second feature map.

[0078] In this embodiment, as Figure 6 As shown, Figure 6 This is a schematic diagram of a backbone network provided in an embodiment of the present invention. For the improved Yolov8 target detection model, its backbone network includes a backbone network, an FPN structure, and an SPP structure.

[0079] The backbone network is used to extract multi-scale features from the image to be detected and to fuse these multi-scale features, specifically including:

[0080] Multi-scale features of inspection images are extracted using a backbone network;

[0081] The multi-scale features are fused using FPN to construct a multi-scale feature pyramid;

[0082] Spatial pooling of multi-scale feature pyramids is performed using the SPP structure to fuse features from different receptive fields, thereby enhancing the detection capability for targets of different sizes.

[0083] Specifically, the backbone network is either the ResNeXt-101 network or the CSPDarknetX network. In the improved YOLOv8 object detection model, the YOLOv8 backbone network is modified, employing a deeper convolutional neural network structure, such as ResNeXt-101. Compared to the original YOLOv8 backbone network, ResNeXt-101 learns richer feature representations with the same computational cost.

[0084] ResNeXt-101 is a deep convolutional neural network model that improves and extends ResNet. By introducing mechanisms such as group convolutions and residual connections, ResNeXt-101 enhances the model's expressive power and performance. Combined with the Feature Pyramid Network (FPN), the ResNeXt-101 network can process information from feature maps of different scales, thereby improving the performance of object detection and segmentation. When used as the backbone network of YOLOv8, the combination of ResNeXt-101 and the Feature Pyramid Network (FPN) enables YOLOv8 to better detect objects of different sizes.

[0085] The CSPDarknetX network is a deep convolutional neural network, where X represents the number of layers in the CSPDarknet network. The number of layers in the CSPDarknet network can be configured according to requirements.

[0086] In this embodiment, a Spatial Pyramid Pooling (SPP) module is introduced into the backbone network. The SPP module obtains feature maps with different receptive fields through pooling kernels of different sizes, and then splices them together to obtain a first feature map containing multi-scale feature information. The SPP module fuses features of different scales to enhance the detection capability of targets of different sizes.

[0087] In this embodiment, the local non-attention feature enhancement module is used to enhance the features of the first feature map to obtain the second feature map.

[0088] The Local Unattended Feature Enhancement module aims to enhance the YOLOv8 model's ability to represent detailed features of power facilities, improving its sensitivity to subtle changes and potential faults. The core idea of ​​Local Unattended Feature Enhancement is to dynamically enhance the representation of local features through an unattended mechanism, thereby obtaining richer and more discriminative feature representations.

[0089] Specifically, such as Figure 2 As shown, Figure 2 This is a schematic flowchart of a method for feature enhancement of a first feature map provided in an embodiment of the present invention. Feature enhancement of the first feature map includes steps S101 to S103:

[0090] S101. Divide the first feature map into multiple image blocks.

[0091] Specifically, in this embodiment, when dividing the image into blocks, let the input first feature map be X∈R C×H×W Where C, H, and W represent the number of channels, height, and width, respectively, the first feature map is divided into M×N image blocks of the same size, where the i-th image block...

[0092] S102. Calculate the relevance score for each image patch.

[0093] After dividing the image into blocks, the relevance score of each image block is calculated. Specifically, taking the i-th image block as an example, calculating the relevance score between the i-th image block and the other image blocks includes steps S1021 to S1022:

[0094] S1021. Calculate the correlation value between the i-th image block and each of the other image blocks;

[0095] Specifically, the i-th image patch x i With the j-th image patch x j The correlation value r between them i,j Calculate according to the following formula, where i≠j;

[0096] r i,j =f(x) i ,x j )=φ(x i ) T ψ(x j ) T (1);

[0097] In equation (1), φ(·) or ψ(·) is the embedding function, x i For the i-th image patch, x j Let j be the j-th image patch.

[0098] In this embodiment, the embedding function can be a convolutional or fully connected layer. The embedding function is used to extract image patch features, such as the ResNeXt-101 network. These networks can learn effective representations in image patches, and the similarity between image patches can be compared using these representations.

[0099] S1022. Sum the correlation values ​​between the i-th image patch and each of the other image patches to obtain the correlation score r corresponding to the i-th image patch. i The correlation score r corresponding to the i-th image patch i The calculation formula is as follows:

[0100] r i =∑ j r i,j (5).

[0101] Specifically, the local attention-free feature enhancement module first divides the input feature map into blocks, dividing it into multiple local regions. Then, it applies an attention-free mechanism to each local region and calculates its correlation score with all other regions.

[0102] S103. Perform channel attention enhancement and spatial attention enhancement on each image patch based on the correlation score.

[0103] In this embodiment, the relevance score reflects the similarity between the region and the global features. The higher the score, the more important the region is and the more it needs to be enhanced. For each local region, the local non-attention feature enhancement module dynamically adjusts the degree of feature enhancement based on its relevance score. High relevance regions will receive greater enhancement, while low relevance regions will receive less enhancement.

[0104] Specifically, such as Figure 3 As shown, Figure 3 This is a flowchart illustrating attention enhancement for the i-th image patch based on its relevance score. Taking the i-th image patch among a plurality of image patches as an example, attention enhancement for the i-th image patch among a plurality of image patches based on its relevance score includes the following steps S201 to S203:

[0105] S201. Perform channel attention enhancement on the i-th image block to obtain the first image block.

[0106]

[0107] In equation (2), ⊙ represents the element-wise product, x i For the i-th image patch, γ c (·) represents the channel attention function, which includes a fully connected layer. The channel attention function assigns different weights to each feature channel through adaptive learning.

[0108] Specifically, the channel attention function may include a fully connected layer. The parameters of the fully connected layer are obtained through adaptive learning. The fully connected layer transforms the feature phasor to obtain a weight phasor. Each value in the weight phasor corresponds to the weight of a channel of the image patch. The obtained weight vector is multiplied with the original image patch channel by channel to achieve channel attention enhancement and obtain the first image patch.

[0109] S202, For the first image block Spatial attention enhancement is performed to obtain the second image patch.

[0110]

[0111] In the formula, ⊙ represents the element-wise product, x i For the i-th image patch, δ s (·) represents the spatial attention function, which includes a convolutional layer. The spatial attention function uses adaptive learning to focus on specific spatial regions in the image while ignoring regions that are irrelevant to the target task or are redundant.

[0112] Specifically, the spatial attention function may include a convolutional layer, the parameters of which are obtained through adaptive learning. The spatial attention function generates a weight map of the same size as the feature map based on the input feature map. Each value in the weight map represents the importance or attention at the corresponding spatial location. By multiplying the weight map element-wise with the original first image patch, spatial attention can be enhanced and a second image patch can be obtained.

[0113] S203, Based on the i-th image block x i Second image block The relevance score r of the i-th image patch i Obtain the enhanced i-th image patch

[0114]

[0115] In the formula, λ is the learnable enhancement coefficient, and r i For the i-th image patch x i The corresponding relevance score.

[0116] In this embodiment, the specific operation of feature enhancement includes two parts: channel attention enhancement and spatial attention enhancement. Specifically, channel attention enhancement uses a fully connected layer to calculate the attention weight of each channel, and then re-weights the channel features according to the weight to highlight the role of important channel features; spatial attention enhancement uses convolution operations to calculate the attention weight of each position and re-encodes the spatial position information.

[0117] After channel attention enhancement and spatial attention enhancement operations, the local non-attention feature enhancement module outputs the enhanced second feature map, which is then fused into the feature extraction stage of the YOLOv8 model for object detection.

[0118] The YOLOv8 model introduces a local non-attention feature enhancement module. This module first divides the feature map into multiple local regions, then dynamically adjusts the enhancement level of each region using a non-attention mechanism, and further enhances the local features with fine-grained attention through channel and spatial attention. This "global guidance + local enhancement" approach can more effectively improve the expressive power of local features.

[0119] The local attention-free feature enhancement module enables the improved YOLOv8 model to better capture the detailed features of power facilities, improving its sensitivity to subtle changes and potential faults. At the same time, the attention-free mechanism simplifies the feature enhancement process, reduces computational complexity, and facilitates real-time detection. In addition, the strategy of dynamically adjusting the enhancement level also makes the model more adaptable to environmental changes, ensuring that the model can maintain high detection stability and accuracy even when factors such as lighting conditions and equipment location change.

[0120] In this embodiment, as Figure 4 As shown, Figure 4 This is a schematic flowchart of identifying a target in a detection image based on the second feature map. Identifying a target in a detection image based on the second feature map includes the following steps S401 to S403:

[0121] S301. Based on the second feature map, predict the location and category of each key point;

[0122] S302. Generate prediction boxes and category information corresponding to target boxes based on the location and category of each key point;

[0123] S303. Perform NMS processing on the predicted bounding boxes, filter the predicted bounding boxes in combination with the confidence threshold, and output the detection results.

[0124] The detection results include: the location of the target in the image to be detected, and the target's category information.

[0125] In this embodiment, the detection head is a key point detection head, such as... Figure 7 As shown, Figure 7This is a schematic diagram of the detection process of the improved Yolov8 object detection model provided in the embodiment of the present invention. The key point detection head adapts to key point detection and detects the coordinate information of the key point position in the image. After obtaining the key point information, the target box and corresponding category information are recovered based on the key point information. Then, NMS (non-maximum suppression) processing and threshold filtering processing are used to remove redundant prediction results, and finally the target box coordinates (Bbox Coord) and category coordinates (Class Coord) are output.

[0126] After the improved Yolov8 object detection model outputs the final detection results, the results are visualized and saved for subsequent analysis.

[0127] By integrating a local attention-free feature enhancer algorithm on top of YOLOv8, this invention optimizes the model from two dimensions: feature enhancement and target detection. This improves the model's sensitivity to and accuracy in recognizing detailed features of power facilities, thereby enhancing the efficiency and accuracy of intelligent inspection.

[0128] In terms of feature enhancement, this invention employs a local non-attention mechanism to meticulously process the feature map, enhancing the model's ability to identify small targets in complex backgrounds. Simultaneously, by utilizing the local feature enhancement function of the local non-attention feature enhancer, it more effectively captures subtle changes and fault signs in power facilities, providing richer and more discriminative feature representations to assist YOLOv8 in making more accurate judgments. Furthermore, the local non-attention feature enhancer simplifies the feature processing process and reduces computational complexity by adopting a non-attention mechanism, thereby contributing to the efficient real-time operation of the intelligent inspection system. Therefore, by dynamically enhancing local features, the model's adaptability to environmental changes is improved, ensuring that even when factors such as lighting conditions and equipment location change, the model maintains high detection stability and accuracy.

[0129] In terms of object detection, firstly, ResNeXt-101 is used as the backbone network. Compared with the original YOLOv8 backbone network, ResNeXt-101 learns richer feature representations with the same computational cost. Secondly, the detection head does not use the traditional anchor box mechanism, but instead uses key point representation. By predicting the position and category of the target key points, the target box is indirectly recovered. This can more flexibly adapt to targets of different shapes and sizes, and reduce the missed detection of difficult samples (such as dense or occluded targets).

[0130] In summary, this invention proposes a target detection method for power facilities in intelligent inspection. This method integrates a local non-attention feature enhancer and the YOLOv8 target detection algorithm, aiming to optimize the target detection performance in intelligent inspection systems, particularly in the monitoring and maintenance of power facilities in data centers. Compared to the traditional YOLOv8 algorithm, the proposed target detection method for power facilities in intelligent inspection significantly improves the accuracy of target detection, particularly for small devices and devices in complex backgrounds. Furthermore, by introducing a local non-attention feature enhancer, the algorithm's sensitivity to subtle changes in power facilities is improved, enabling more effective identification of potential fault signs and enhancing the timeliness and accuracy of early warnings.

[0131] Regarding real-time performance, although the local non-attention feature enhancer increases the computational steps of the model, the processing speed of the new model can still meet the needs of real-time monitoring by optimizing the detection model structure, ensuring the normal and efficient operation of the system. In terms of dynamic environmental adaptability, the new model maintains high detection stability and accuracy under changing environmental conditions such as illumination changes and equipment relocation, demonstrating its robustness in practical application environments.

[0132] Embodiments of the present invention also provide a power equipment target detection system for intelligent inspection. Figure 8 This is a schematic diagram of a power equipment target detection system for intelligent inspection provided in an embodiment of the present invention, as shown below. Figure 8 As shown, the power equipment target detection system for intelligent inspection provided in this embodiment of the invention includes:

[0133] The image acquisition unit is used to acquire inspection images;

[0134] The target detection unit is used to input the inspection image into the improved Yolov8 target detection model to obtain the target detection results of power facilities;

[0135] The improved Yolov8 object detection model includes:

[0136] The backbone network is used to extract multi-scale features from the image to be detected and fuse the multi-scale features to obtain a first feature map containing multi-scale feature information.

[0137] A local non-attention feature enhancement module is used to enhance the features of the first feature map to obtain a second feature map;

[0138] The detection head is used to output the detection results based on the second feature map.

[0139] In some embodiments, the feature enhancement of the first feature map includes:

[0140] The first feature map is divided into multiple image blocks, and the relevance score of each image block is calculated. Based on the relevance score, channel attention enhancement and spatial attention enhancement are performed on the corresponding image blocks.

[0141] Specifically, taking the i-th image patch as an example, calculating the correlation score between the i-th image patch and the remaining image patches includes:

[0142] Calculate the correlation value between the i-th image patch and every other image patch;

[0143] The correlation scores r of the i-th image patch are obtained by summing the correlation values ​​between the i-th image patch and each of the other image patches. i ;

[0144] Specifically, the correlation value r between the i-th image patch and the j-th image patch is calculated. i,j Calculate according to the following formula, where i≠j;

[0145] r i,j =f(x) i ,x j )=φ(x i ) T ψ(x j ) T (1);

[0146] In the formula, φ(·) or ψ(·) is the embedding function, x i For the i-th image patch, x j Let j be the j-th image patch.

[0147] Based on the correlation score, channel attention enhancement and spatial attention enhancement are performed on the i-th image block among the multiple image blocks, including:

[0148] Channel attention enhancement is performed on the i-th image block among the multiple image blocks to obtain the first image block.

[0149] In the formula, ⊙ represents the element-wise product, x i For the i-th image patch, γ c (·) represents the channel attention function, which includes a fully connected layer. The channel attention function assigns different channel attention weights to each feature channel of the image patch through adaptive learning.

[0150] For the first image block Spatial attention enhancement is performed to obtain the second image patch.

[0151]

[0152] In the formula, ⊙ represents the element-wise product, x i For the i-th image patch, δ s (·) represents the spatial attention function, which includes a convolutional layer and assigns different spatial attention weights to each location in the image patch through adaptive learning.

[0153] According to the second image block Enhanced image patches are obtained from correlation scores.

[0154]

[0155] In the formula, λ is the learnable enhancement coefficient, and r i For the i-th image patch x i The corresponding relevance score.

[0156] In some embodiments, the backbone network includes a backbone network, an FPN structure, and an SPP structure, wherein the backbone network is a ResNeXt-101 network or a CSPDarknetX network.

[0157] The extraction and fusion of multi-scale features from the image to be detected includes:

[0158] Multi-scale features of inspection images are extracted using a backbone network;

[0159] The multi-scale features are fused using an FPN structure to construct a multi-scale feature pyramid.

[0160] Spatial pooling of multi-scale feature pyramids is performed using the SPP structure to fuse features from different receptive fields.

[0161] In some embodiments, identifying a target in the image to be detected based on the second feature map includes:

[0162] Based on the second feature map, predict the location and category of each key point;

[0163] Generate prediction boxes and corresponding category information for target boxes based on the location and category of each key point;

[0164] The predicted bounding boxes are processed by NMS, and then filtered using a confidence threshold. The detection results are then output.

[0165] Embodiments of the present invention also provide an electronic device, such as... Figure 9 As shown, Figure 9This is a schematic diagram of an electronic device provided in an embodiment of the present invention. The electronic device includes a memory and a processor. The memory is used to store computer program code and transmit the computer program code to the processor. The processor is used to execute the above-described power equipment target detection method for intelligent inspection according to the instructions in the computer program code.

[0166] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described power equipment target detection method for intelligent inspection.

[0167] Generally, the computer instructions for implementing the method of the present invention can be carried on any combination of one or more computer-readable storage media. Non-transitory computer-readable storage media can include any computer-readable medium except for the signal itself, which is temporarily propagating.

[0168] Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples (not an exhaustive list) of computer-readable storage media include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with a system, apparatus, or device that executes instructions.

[0169] Computer program code for performing the operations of this invention can be written using one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. In particular, Python, suitable for neural network computation, and platform frameworks based on TensorFlow, PyTorch, etc., can be used. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of remote computers, the remote computer can be connected to the user's computer via various types of networks, including local area networks (LANs) or wide area networks (WANs), or via an internet service provider.

[0170] Although embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for detecting targets in power equipment for intelligent inspection, characterized in that: The detection method includes: Acquire inspection images; The inspection images are input into the improved Yolov8 target detection model to obtain the target detection results of power equipment; The improved Yolov8 object detection model includes: The backbone network is used to extract multi-scale features from the inspection images and fuse the multi-scale features to obtain a first feature map containing multi-scale feature information. A local non-attention feature enhancement module is used to enhance the features of the first feature map to obtain a second feature map; The detection head is used to output the detection result based on the second feature map; Feature enhancement of the first feature map includes: The first feature map is divided into multiple image blocks, the relevance score of each image block is calculated, and attention enhancement is performed on each image block based on the relevance score. Calculate the first of the multiple image blocks The relevance score of each image patch includes: Calculate the first The correlation value between each image patch and every other image patch; For the Summing the correlation values ​​between the first image patch and every other image patch yields the first... Correlation score of each image patch ; Among them, the The image patch and the first Correlation values ​​between image patches Calculate using the following formula. ; （1）； In the formula, or For the embedding function, For the first Image blocks, For the first Image blocks; Based on the correlation score, the first of the multiple image patches is... Attention enhancement for each image patch includes: For the Channel attention enhancement is performed on each image patch to obtain the first image patch. : （2）； In the formula, Represents element-wise product. For the first Image blocks, The channel attention function includes a fully connected layer and assigns different channel attention weights to each feature channel of the image patch through adaptive learning. For the first image block Spatial attention enhancement is performed to obtain the second image patch. : （3）； In the formula, Represents element-wise product. For the first Image blocks, The spatial attention function includes a convolutional layer and assigns different spatial attention weights to each location in the image patch through adaptive learning. According to the Image blocks Second image block , No. Correlation score of each image patch The enhanced first Image blocks : （4）； In the formula, To enhance the coefficient, For the first Image blocks The relevance score.

2. The power equipment target detection method for intelligent inspection according to claim 1, characterized in that: The backbone network includes a backbone network, an FPN structure, and an SPP structure. The extraction and fusion of multi-scale features from the inspection images includes: Multi-scale features of inspection images are extracted using a backbone network; The multi-scale features are fused using an FPN structure to construct a multi-scale feature pyramid. Spatial pooling of multi-scale feature pyramids is performed using the SPP structure to fuse features from different receptive fields.

3. The power equipment target detection method for intelligent inspection according to claim 2, characterized in that: The backbone network is either a ResNeXt-101 network or a CSPDarknetX network.

4. The power equipment target detection method for intelligent inspection according to claim 1, characterized in that: The detection results output based on the second feature map include: Based on the second feature map, predict the location and category of each key point; Generate prediction boxes and corresponding category information for target boxes based on the location and category of each key point; The predicted bounding boxes are processed by NMS, and then filtered using a confidence threshold to output the detection results.

5. A power equipment target detection system for intelligent inspection, the system being applied to the method described in any one of claims 1-4, the system comprising: The image acquisition unit is used to acquire inspection images; The target detection unit is used to input the inspection image into the improved Yolov8 target detection model to obtain the target detection results of power equipment; The improved Yolov8 object detection model includes: The backbone network is used to extract multi-scale features from the image to be detected and fuse the multi-scale features to obtain a first feature map containing multi-scale feature information. A local non-attention feature enhancement module is used to enhance the features of the first feature map to obtain a second feature map; Feature enhancement of the first feature map includes: The first feature map is divided into multiple image blocks, the relevance score of each image block is calculated, and attention enhancement is performed on each image block based on the relevance score. Calculate the first of the multiple image blocks The relevance score of each image patch includes: Calculate the first The correlation value between each image patch and every other image patch; For the Summing the correlation values ​​between the first image patch and every other image patch yields the first... Correlation score of each image patch ; Among them, the The image patch and the first Correlation values ​​between image patches Calculate using the following formula. ; （1）； In the formula, or For the embedding function, For the first Image blocks, For the first Image blocks; Based on the correlation score, the first of the multiple image patches is... Attention enhancement for each image patch includes: For the Channel attention enhancement is performed on each image patch to obtain the first image patch. : （2）； In the formula, Represents element-wise product. For the first Image blocks, The channel attention function includes a fully connected layer and assigns different channel attention weights to each feature channel of the image patch through adaptive learning. For the first image block Spatial attention enhancement is performed to obtain the second image patch. : （3）； In the formula, Represents element-wise product. For the first Image blocks, The spatial attention function includes a convolutional layer and assigns different spatial attention weights to each location in the image patch through adaptive learning. According to the Image blocks Second image block , No. Correlation score of each image patch The enhanced version Image blocks : （4）； In the formula, To enhance the coefficient, For the first Image blocks The relevance score; The detection head is used to output the detection results based on the second feature map.

6. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory is used to store computer program code and transfer the computer program code to the processor; The processor is configured to execute, according to instructions in the computer program code, the power equipment target detection method for intelligent inspection as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the power equipment target detection method for intelligent inspection as described in any one of claims 1 to 4.

Images