Target detection model based on YOLOv11 model, infrared power detection method, medium and equipment

Abstract

The embodiment of the invention discloses a target detection model based on a YOLOv11 model, an infrared power detection method, a medium and equipment, and the model comprises a feature extraction module which carries out the feature extraction of an infrared power image, and comprises an HRTConv module and an HRT-C3K2 module, each HRTConv module divides the features of the input infrared power image into two parts according to channels, one part is subjected to standard convolution operation to extract standard features, the other part is subjected to lightweight convolution operation to extract lightweight features, and the standard features and the lightweight features are spliced to obtain first features; the HRT-C3K2 module performs feature extraction on the first feature again to obtain a second feature; fusing a preset number of second features of the module to obtain multi-scale features; the identification module identifies the information of the power equipment according to the multi-scale features. According to the invention, the HRTConv module and the HRT-C3K2 module are adopted, so that the detection precision is greatly improved while the light weight of the model is maintained.

Description

Technical Field

[0001] The present invention relates to the field of infrared power equipment detection technology, and in particular to a target detection model based on a YOLOv11 model, an infrared power detection method, a medium, and a device. Background Art

[0002] In the field of target detection for infrared power equipment, traditional methods rely on manually designed feature extraction mechanisms. These mechanisms often perform poorly in complex backgrounds and low-resolution infrared images, making it difficult to achieve high-precision target recognition. The introduction of deep learning technology has brought significant progress. Deep learning-based target detection algorithms can automatically generate more effective feature representations, significantly improving detection accuracy. However, existing deep learning models are computationally intensive, making them difficult to run on resource-constrained devices. To address this, many researchers have sought to lightweight these models. Current lightweight improvements, such as depthwise separable convolution and grouped convolution, have reduced model complexity and computational overhead, improving their deployment efficiency on resource-constrained devices. However, these traditional lightweight architectures still suffer from several key bottlenecks: first, the static models lack adaptability to different scenarios, requiring network redesign to accommodate diverse computational requirements; second, lightweight architectures sacrifice the model's feature representation capabilities, resulting in reduced recognition accuracy.

[0003] To overcome these shortcomings, existing research has proposed a variety of lightweight convolutional modules, implemented by cascading standard convolutions and lightweight convolutions. These lightweight convolutional modules meet the deployment requirements of resource-constrained devices to a certain extent. However, existing solutions still struggle to achieve a dynamic balance between the flexibility of static architectures and the representational power of lightweight neural networks. Therefore, developing a method that can flexibly adjust the model's computational load, parameter count, and accuracy to ensure efficient operation in different application scenarios has become a key challenge in the current field of infrared power equipment target detection. Summary of the Invention

[0004] Based on this, it is necessary to propose a target detection model, infrared power detection method, medium and equipment based on the YOLOv11 model to address the above problems.

[0005] A target detection model based on the YOLOv11 model, the model comprising:

[0006] Feature extraction module, fusion module and recognition module.

[0007] The feature extraction module is used to extract features from infrared power images. The feature extraction module includes: a preset number of HRTConv modules and HRT-C3K2 modules.

[0008] Each HRTConv module divides the features of the input infrared power image into two parts according to the channel, one part performs a standard convolution operation to extract standard features, and the other part performs a lightweight convolution operation to extract lightweight features, and the standard features and lightweight features are spliced ​​to obtain the first features.

[0009] The lightweight convolution operation includes a 3×3 depth convolution operation, a channel recalibration operation, and a 1×1 convolution operation.

[0010] The HRT-C3K2 module extracts the first feature again to obtain the second feature.

[0011] A fusion module is used to fuse a preset number of the second features to obtain multi-scale features.

[0012] An identification module is used to identify information of the electric power equipment according to the multi-scale features.

[0013] Among them, the HRT-BottleNeck module is constructed by the HRTConv module, and the HRT-BottleNeck module includes the first HRTConv module and the second HRTConv module. The first HRTConv module reduces the number of channels of the first feature by half, and the second HRTConv module restores the number of channels to be consistent with the number of channels of the first feature; the HRT-BottleNeck module is used to construct the HRT-C3K2 module.

[0014] The HRT-C3K2 module is constructed according to different values ​​of C3k, wherein C3 represents three convolutional layers and k represents a convolution kernel. The C3K2 module in the YOLOv11 model includes a BottleNeck module and a C3 module. When C3k is True, the BottleNeck module in the C3K2 module is replaced by the HRT-BottleNeck to obtain the HRT-C3K2 module. When C3k is false, the BottleNeck module in the C3 module is replaced by the HRT-BottleNeck module to construct an HRT-C3 module. The C3 module in the C3K2 module is replaced by the HRT-C3 module to obtain the HRT-C3K2 module.

[0015] The HRTConv module includes a standard convolution module, a lightweight convolution module, a splicing module, and a channel shuffling module.

[0016] The standard convolution module is used to perform a standard convolution operation on a portion of the infrared power image to extract standard features.

[0017] The lightweight convolution module is used to perform a 3×3 depth convolution operation on another part of the infrared power image to extract intermediate features, perform a channel recalibration operation on the intermediate features, and then perform a 1×1 convolution operation on some channels to extract lightweight features.

[0018] Among them, the channel number allocation ratio of the standard convolution module and the lightweight convolution module is controlled by the scaling factor: if the scaling factor is 0, the standard convolution module is fully adopted; if the scaling factor is 1, the lightweight convolution module is fully adopted.

[0019] The splicing module is used to splice the standard features and the lightweight features.

[0020] The channel shuffling module is used to divide the spliced ​​standard features and lightweight features into two groups along the channel dimension, and perform a transposition and then a flattening operation on each group to obtain a first feature.

[0021] The fusion module is constructed by replacing the C3K2 module in the YOLOv11 model with the HRT-C3K2 module.

[0022] An infrared power detection method is applied to the target detection model based on the YOLOv11 model as described above, and the method includes:

[0023] Collect infrared power images containing various power equipment.

[0024] The image is input into the object detection model based on the YOLOv11 model, which includes:

[0025] Feature extraction module, fusion module and recognition module; the feature extraction module is used to extract features of infrared power images, and the feature extraction module includes: a preset number of HRTConv modules and HRT-C3K2 modules, each of the HRTConv modules divides the features of the input infrared power image into two parts according to the channel, one part performs a standard convolution operation to extract standard features, and the other part performs a lightweight convolution operation to extract lightweight features, and the standard features and lightweight features are spliced ​​to obtain a first feature; the lightweight convolution operation includes a 3×3 depth convolution operation, a channel recalibration operation and a 1×1 convolution operation; the HRT-C3K2 module performs feature extraction on the first feature again to obtain a second feature; the fusion module is used to fuse a preset number of the second features to obtain multi-scale features; the recognition module is used to identify information of power equipment based on the multi-scale features.

[0026] The feature extraction of the infrared power image is followed by feature splicing to obtain multi-level features, where the features at different levels represent targets of different sizes, specifically including:

[0027] A standard convolution operation is performed on a portion of the infrared power image to extract standard features.

[0028] A 3×3 depthwise convolution operation is performed on another part of the infrared power image to extract intermediate features. After channel recalibration, a 1×1 convolution operation is performed on the intermediate features to extract lightweight features.

[0029] The standard features and lightweight features are spliced ​​together.

[0030] Along the channel dimension, the spliced ​​standard features and lightweight features are divided into two groups, and each group is transposed and then flattened to obtain the first feature.

[0031] The method further comprises:

[0032] The infrared power image is used as a data set, and the target detection model based on the YOLOv11 model is trained using the data set.

[0033] A computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the steps of the above method.

[0034] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0035] The embodiments of the present invention have the following beneficial effects:

[0036] The feature extraction module includes a preset number of HRTConv modules and HRT-C3K2 modules. Each HRTConv module divides the features of the input infrared power image into two parts according to the channel. One part performs a standard convolution operation to extract standard features, and the other part performs a lightweight convolution operation to extract lightweight features. The standard features and lightweight features are then spliced ​​together to obtain the first feature. Among them, the lightweight convolution operation significantly reduces the amount of calculation and the number of parameters compared to ordinary convolution, while the standard convolution operation provides stronger feature expression capabilities. The HRT-C3K2 module further optimizes the feature extraction process, extracts features from the first feature again, and obtains the second feature through multiple HRTConv modules and HRT-C3K2 modules. Second features of different resolutions are obtained, thereby improving the robustness and generalization ability of the model. Furthermore, a preset number of second features are fused through the fusion module to obtain multi-scale features. The recognition module is responsible for converting multi-scale features into category and positioning information. The present invention adopts the HRTConv module and the HRT-C3K2 module to significantly reduce the model complexity while maintaining the model detection accuracy. In addition, by performing channel division through the HRTConv module and adopting parallel standard convolution and lightweight convolution, the calculation ratio is adjustable, so that the model calculation amount can be quickly adjusted according to the actual application scenario, so that it can run efficiently in different application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0038] in:

[0039] Figure 1 A schematic diagram of the structure of an embodiment of a target detection model based on the YOLOv11 model provided by the present invention;

[0040] Figure 2 A schematic structural diagram of an embodiment of the HRTConv module provided by the present invention;

[0041] Figure 3 Schematic diagram of the YOLOv11 model structure before and after improvement. Figure 3 a is a schematic diagram of the structure of the YOLOv11 model. Figure 3 b is a schematic structural diagram of another embodiment of a target detection model based on the YOLOv11 model provided by the present invention;

[0042] Figure 4A schematic structural diagram of an embodiment of the HRT-BottleNeck module provided by the present invention;

[0043] Figure 5 A schematic structural diagram of an embodiment of the HRT-C3K2 module provided by the present invention;

[0044] Figure 6 A schematic structural diagram of another embodiment of the HRT-C3K2 module provided by the present invention;

[0045] Figure 7 A schematic flow chart of an embodiment of an infrared power detection method provided by the present invention;

[0046] Figure 8 A schematic flow chart of another embodiment of an infrared power detection method provided by the present invention;

[0047] Figure 9 A schematic structural diagram of an embodiment of the device provided by the present invention;

[0048] Figure 10 This is a schematic structural diagram of an embodiment of the medium provided by the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0050] like Figure 1 As shown, Figure 1 This is a schematic diagram of the structure of an embodiment of a target detection model based on the YOLOv11 model provided by the present invention. The target detection model based on the YOLOv11 model includes: a feature extraction module Backbone, a fusion module Neck, and a recognition module Head;

[0051] The Backbone feature extraction module is used to extract features from infrared power images. It includes a preset number of HRTConv and HRT-C3K2 modules. Each HRTConv module divides the input infrared power image features into two parts, one for standard convolution to extract standard features, and the other for lightweight convolution to extract lightweight features. The standard and lightweight features are then concatenated to produce the primary features.

[0052] For example, referring to Figure 2 , Figure 2This is a schematic diagram of the structure of an embodiment of the HRTConv module provided by the present invention. The HRTConv module includes a standard convolution module (Vanilla Conv), a lightweight convolution module (Lightweight Conv), a concatenation module (Concat), and a channel shuffle module (Shuffle). Furthermore, by adjusting the ratio of the number of channels in the standard convolution module (Vanilla Conv) and the lightweight convolution module (Lightweight Conv), the model's parameter count, computational complexity, and accuracy can be flexibly adjusted.

[0053] Specifically, the scaling factor α∈(0,1) is used to control the channel allocation of the standard convolution module Vanilla Conv and the lightweight convolution module Lightweight Conv, thereby flexibly adjusting the computational complexity and accuracy of the model and improving the flexibility of the model in different scenarios. The allocation method is shown in the following formula:

[0054]

[0055] in, is the number of channels of the standard convolution module, is the number of channels of the lightweight convolution module, C is the total number of output channels, and F vanilla is the standard feature, F lightwight Lightweight feature.

[0056] The HRTConv module uses a parallel heterogeneous dual-path structure to perform a standard convolution operation on a part of the infrared power image through the standard convolution module Vanilla Conv to extract the standard feature F vanilla Specifically, the number of channels of the standard convolution module Vanilla Conv is The extraction process is as follows:

[0057] F vanilla =Conv(X);

[0058] Among them, F vanilla is a standard feature, X is an infrared power image, and Conv is a standard convolution operation.

[0059] At the same time, the lightweight convolution module Lightweight Conv performs a 3×3 depth convolution operation on another part of the infrared power image to extract the intermediate features. After the channel recalibration operation is performed on the intermediate features, a 1×1 convolution operation is performed on some channels to extract the lightweight features F lightwight Specifically, the number of channels of the lightweight convolution module Lightweight Conv is: The extraction process is shown below:

[0060] F lightwight=CEDSConv(X);

[0061] Among them, F lightwight is a lightweight feature, X is an infrared power image, and CEDConv is a lightweight convolution operation.

[0062] Furthermore, the standard features and lightweight features are concatenated through the concatenation module Concat. Specifically, the features of the two roads are concatenated as shown below:

[0063] F fusion =Concat(F lightweight ,F vanilla );

[0064] Among them, F fusion is the fusion feature, Concat is the splicing operation, F vanilla is the standard feature, F lightwight Lightweight feature.

[0065] Furthermore, the channel shuffle module shuffles along the channel dimension to divide the concatenated standard features and lightweight features into two groups, and then transposes and flattens each group to obtain the first feature. Specifically, along the channel dimension, the channel is divided into two groups as shown in the following formula:

[0066]

[0067] in, is the first set of fusion features, is the second set of fusion features.

[0068] Furthermore, the output of the previous step is transposed and then flattened as shown below:

[0069] Y=Flatten(Transpose(F fusion ));

[0070] Among them, Y is the first feature, Flatten is the flattening operation, and Transpose is the device operation.

[0071] The HRT-C3K2 module extracts the first feature again to obtain the second feature.

[0072] For example, referring to Figure 3 , Figure 3 Schematic diagram of the YOLOv11 model structure before and after improvement. Figure 3 a is a schematic diagram of the structure of the YOLOv11 model. Figure 3b is a structural diagram of another embodiment of a target detection model based on the YOLOv11 model provided by the present invention. The target detection model based on the YOLOv11 model includes a feature extraction module Backbone, a fusion module Neck, and a recognition module Head; the feature extraction module Backbone is used to extract features from infrared power images, and the feature extraction module Backbone includes: 5 HRTConv modules and 4 HRT-C3K2 modules. By stacking the HRTConv modules and the HRT-C3K2 modules, multi-level semantic features (multi-level second features) are obtained - shallow semantic features are low-level but high-resolution, and deep semantic features are high-level but low-resolution. In the target detection task, the key to achieving precise positioning is to obtain semantic features that have both high-level semantic features and high-resolution.

[0073] Specifically, an infrared power equipment image (resolution 640×640×3) is input. First, an HRTConv module with a stride of 2 is used to perform preliminary feature extraction and downsampling, outputting shallow features (feature size 320×320×64). Here, the scale factor α of the HRTConv module is 0.75, that is, 48 ​​of the 64 output channels (the number of channels is the number of output feature maps) are obtained by processing the input feature map through lightweight convolution (accounting for 75% of the total output channels), and 16 are obtained by processing the input feature map through standard convolution (accounting for 25% of the total output channels). Then, the first feature of the first layer (resolution 160×160×128) is obtained through another HRTConv module with a stride of 2. The first feature is extracted again through an HRT-C3K2 module. While the resolution remains unchanged (160×160), the output channel is expanded to 256, and the second feature of the first layer is output (resolution is 160×160×256). Here, HRT-C3K2 also adopts α=0.75, that is, lightweight convolution is mainly used.

[0074] Furthermore, an HRTConv module with a step size of 2 and α = 0.75 is used to further reduce the resolution to 80 × 80, while keeping the number of channels unchanged. An HRT-C3K2 module with α = 0.75 is used to maintain the resolution unchanged and expand the output channels to 512, outputting the second layer of second features (resolution 80 × 80 × 512). The second layer of second features includes the contours of medium-sized objects, such as the circuit breaker body.

[0075] Furthermore, the second feature output of the second layer, with a resolution of 80×80×512, passes through an HRTConv module with α=0.75 and a stride of 2, and the resolution is reduced to 40×40, while the number of channels remains unchanged at 512. An HRT-C3K2 module with α=0.75 maintains the resolution and number of channels unchanged, making it the main lightweight path to reduce the amount of computation, and outputs the second feature of the third level (resolution of 40×40×512). The second feature of the third level contains high-level semantics, such as device category.

[0076] Furthermore, the second feature output at the third level, with a resolution of 40×40×512, is reduced to 20×20 through an HRTConv module with an α=0.75 and a stride of 2, and the number of channels is increased to 1024. An HRT-C3K2 module adjusts the scale factor α to 0.75, primarily using a lightweight path to maximize computational efficiency, and outputs the second feature at the fourth level (resolution 20×20×1024). This second feature contains high-level semantics and global context information.

[0077] Furthermore, the second feature of the fourth level is subjected to a 5×5 pooling operation in series in the SPPF (Spatial Pyramid Pooling Fast) module to generate a multi-scale pooled second feature (containing multi-scale context information).

[0078] Furthermore, the second feature after multi-scale pooling is processed with channel attention and spatial attention through C2PSA (Channel and PositionSpatialAttention). Specifically, channel weights are generated through global average pooling (GAP) and fully connected layers to strengthen the feature responses of important channels. At the same time, a spatial weight map is generated through convolution operations to highlight the spatial position information of the target area. The channel weights and spatial weights are fused by addition or multiplication to form the final feature enhancement weight. The second feature after multi-scale pooling is dynamically weighted by the feature enhancement weight to improve the positioning accuracy of the feature, and finally the optimized second feature after multi-scale pooling (resolution of 20×20×1024) is output.

[0079] It should be noted that the value of the scale factor α is ∈ [0, 1], and the channel number distribution ratio of the standard convolution module and the lightweight convolution module is controlled by the scale factor: if the scale factor is 0, the standard convolution module is fully adopted; if the scale factor is 1, the lightweight convolution module is fully adopted.

[0080] The fusion module Neck is used to fuse a preset number of second features to obtain multi-scale features.

[0081] For example, see Figure 3 By replacing the C3K2 module in the YOLOv11 model with the HRT-C3K2 module, a fusion module is constructed. The fusion module fuses the second features of the second layer at different resolutions, the second features of the third layer, and the optimized second features after multi-scale pooling to obtain multi-scale features. This integrates high-level semantic information from deep layers into shallow high-resolution features, ensuring that features at each scale contain high-level semantic information.

[0082] The identification module Head is used to identify the information of power equipment based on multi-scale features.

[0083] Exemplarily, the multi-scale features output by the fusion module Neck are converted into target category information and positioning information through the recognition module Head.

[0084] From the above description, it can be seen that the feature extraction module includes a preset number of HRTConv modules and HRT-C3K2 modules. Each HRTConv module divides the features of the input infrared power image into two parts according to the channel. One part performs a standard convolution operation to extract standard features, and the other part performs a lightweight convolution operation to extract lightweight features. The standard features and lightweight features are then spliced ​​together to obtain the first feature. Among them, the lightweight convolution operation significantly reduces the amount of calculation and the number of parameters compared to ordinary convolution, while the standard convolution operation provides stronger feature expression capabilities. The HRT-C3K2 module further optimizes the feature extraction process, extracts features from the first feature again, and obtains the second feature through multiple HRTConv modules and HRT-C3K2 modules. Second features of different resolutions are obtained, thereby improving the robustness and generalization ability of the model. Furthermore, the preset number of second features are fused through the fusion module to obtain multi-scale features. The recognition module is responsible for converting multi-scale features into category and positioning information. The present invention adopts the HRTConv module and the HRT-C3K2 module to significantly reduce the amount of calculation while maintaining detection accuracy. In addition, by performing channel division through the HRTConv module and adopting parallel standard convolution and lightweight convolution, the calculation ratio is adjustable, so that the model calculation amount can be quickly adjusted according to the actual application scenario, so that it can operate efficiently in different application scenarios.

[0085] like Figure 4 、 Figure 5 and Figure 6 As shown, Figure 4 This is a structural diagram of an embodiment of the HRT-BottleNeck module provided by the present invention. Figure 5 This is a structural diagram of an embodiment of the HRT-C3K2 module provided by the present invention. Figure 6This is a schematic diagram of the structure of another embodiment of the HRT-C3K2 module provided by the present invention. The HRT-BottleNeck module is constructed using the HRTConv module. The HRT-BottleNeck module includes a first HRTConv module and a second HRTConv module. The first HRTConv module halves the number of channels of the first feature, while the second HRTConv module restores the number of channels to the same as the first feature. The HRT-C3K2 module is constructed using the HRT-BottleNeck module.

[0086] The HRT-C3K2 module is constructed according to the different values ​​of C3k, where C3 is the 3 convolution layers and k is the convolution kernel. The C3K2 module in the YOLOv11 model includes the BottleNeck module and the C3 module. When C3k is True, Figure 6 As shown, the BottleNeck module in the C3K2 module is replaced by HRT-BottleNeck to obtain the HRT-C3K2 module. When C3k is false, as shown in Figure 7 As shown, the BottleNeck module in the C3 module is replaced by the HRT-BottleNeck module to construct the HRT-C3 module, and the C3 module in the C3K2 module is replaced by the HRT-C3 module to obtain the HRT-C3K2 module.

[0087] Exemplarily, the HRT-BottleNeck module is constructed by the HRTConv module. The HRT-BottleNeck module includes a first HRTConv module and a second HRTConv module. The first HRTConv module halves the number of channels of the first feature, and the second HRTConv module restores the number of channels to be consistent with the number of channels of the first feature. HRT-BottleNeck first uses the first HRTConv module to reduce the number of channels of the input first feature to half of the original number, and then uses the second HRTConv to restore it to the original number of channels. Compared with directly stacking two convolution modules with the same number of output channels, the number of parameters and the amount of calculation of the module can be greatly reduced. At the same time, the combination of the residual structure further increases the stability of the module during training. In addition, since HRT-BottleNeck adopts the HRTConv structure, the amount of calculation and the amount of parameters of HRT-BottleNeck can be adjusted by adjusting the scale coefficient of HRTConv. The first feature is extracted again by the HRT-C3K2 module to obtain the second feature.

[0088] Specifically, the HRT-C3K2 module is constructed based on the value of C3k, where C3 represents three convolutional layers and k represents the convolution kernel. In the YOLOv11 model, the C3K2 module consists of a BottleNeck module and a C3 module. When C3k is True, the BottleNeck module in the C3K2 module is replaced with HRT-BottleNeck, while the other modules remain unchanged, resulting in the HRT-C3K2 module.

[0089] When C3k is false, the BottleNeck module in the C3K2 module is replaced with HRT-BottleNeck, while other modules remain unchanged, thus completing the HRT-C3 construction. This module adopts the idea of ​​CSPnet, that is, convolution operations are only performed on half of the input channels. Experiments have shown that this structure can significantly reduce the amount of computation with a small loss of accuracy.

[0090] In the process of power equipment target monitoring, the following indicators are used to evaluate the performance of the target detection model based on the YOLOv11 model.

[0091] (1) Precision

[0092] Precision refers to the proportion of samples predicted by the model as positive that are actually positive. It is used to measure the accuracy of the model's predictions, that is, how many of all the instances predicted as positive samples are actually positive samples. The calculation formula for precision is:

[0093]

[0094] Among them, TP is the number of true positives and FP is the number of false positives.

[0095] (2) Recall

[0096] Recall rate refers to the proportion of instances that are correctly predicted as positive by the model among all instances that are actually positive samples. It is used to measure the coverage ability of the model, that is, how many of the samples that are actually positive are captured by the model. The calculation formula of recall rate is:

[0097]

[0098] Here, FN represents the number of false negatives.

[0099] (3) Mean Average Precision (mAP)

[0100] mAP is the average of the mean precision (AP) across all categories. AP is calculated for a single category and measures the balance between precision and recall for that category. The mAP calculation involves integrating the precision and recall at different thresholds to create an area under the curve, which is then averaged across all categories.

[0101] (4) mAP@0.5:0.95

[0102] mAP@0.5:0.95 is a variant of mAP that calculates the average precision over an IoU (Intersection over Union) threshold range of 0.5 to 0.95 (in steps of 0.05). This calculation method provides a more rigorous evaluation metric by considering the model's performance at different IoU thresholds, thereby more comprehensively reflecting the model's detection capabilities.

[0103] As shown in Table 1, which compares the algorithms, this model achieves accuracy close to that of Yolov8 (mAP@0.5 decreases by 0.3% and mAP@0.5:0.9 decreases by 1.8%) when the scaling factor a = 0.75, but with only 57% of the computational cost. Compared to YOLOv11, the accuracy is similar (mAP@0.5 is the same, mAP@0.5:0.9 decreases by 0.3%), with only 74.6% of the computational cost. The computational cost is close to that of Yolov5, but the mAP@0.5:0.9 increases by 5%.

[0104] Table 1 Algorithm comparison table

[0105]

[0106] Storage refers to the disk space occupied by the model file, and Complexity refers to the computational complexity.

[0107] Table 2 also shows a comparison of scaling factors. By adjusting the scaling factor a, the model can maintain a controllable accuracy drop (ΔmAP@0.5≤0.3%, ΔmAP@0.5:0.9≤1.6), significantly reducing the computational complexity (down to a maximum of 61.9% of Yolov11). Furthermore, when a=0.5, not only is the computational complexity significantly reduced, but detection accuracy is also improved.

[0108] Table 2 Scale factor comparison table

[0109]

[0110]

[0111] The present invention not only enables the model to have the ability to adjust the amount of calculation, but also reduces the amount of calculation when α=0.5 while improving the accuracy, which is very practical.

[0112] like Figure 7 As shown, Figure 7 This is a flow chart of an embodiment of an infrared power detection method provided by the present invention. An infrared power detection method is applied to a target detection model based on the YOLOv11 model, and the method includes:

[0113] S101: Collect infrared power images containing various power equipment.

[0114] For example, FLIR thermal imagers were used to capture infrared images of power equipment at multiple substations. These images form a dataset consisting of 3,930 images with a fixed resolution of 640×640 pixels to ensure that key details are fully preserved. The dataset covers seven types of power equipment: arresters, circuit breakers, bushings, current-transformer disconnectors, disconnectors, disconnector 2, and insulators. A total of 16,413 bounding boxes were annotated (in PASCAL VOC format), providing high-quality annotations for subsequent model training. This dataset provides comprehensive and reliable support for the training of object detection models, significantly improving their recognition accuracy and generalization capabilities for diverse power equipment.

[0115] To mitigate the potential overfitting risk caused by limited data and enhance the object detection model's adaptability to complex real-world scenarios, we employed multi-dimensional data augmentation techniques, including rotation, flipping, affine transformation, cropping, brightness adjustment, Gaussian noise injection, and Gaussian blurring, to simulate real-world imaging variations such as viewpoint change, camera shake, and sensor noise. This augmentation expanded the dataset to 5,895 images, significantly improving the model's robustness.

[0116] To ensure objectivity and generalizability of the model evaluation, a stratified random sampling strategy was used to partition the dataset into training, validation, and test sets, accounting for approximately 70% (4,126 images), 15% (884 images), and 15% (885 images), respectively. Class balancing was strictly enforced during the partitioning process to ensure that the distribution ratio of the seven device categories within each subset was consistent with that in the original dataset.

[0117] S102: Input the image into the target detection model based on the YOLOv11 model, which includes: a feature extraction module, a fusion module and a recognition module; the feature extraction module is used to extract features of the infrared power image, and the feature extraction module includes: a preset number of HRTConv modules and HRT-C3K2 modules. Each HRTConv module divides the features of the input infrared power image into two parts according to the channel, one part performs a standard convolution operation to extract standard features, and the other part performs a lightweight convolution operation to extract lightweight features, and the standard features and the lightweight features are spliced ​​to obtain the first feature; the lightweight convolution operation includes a 3×3 depth convolution operation, a channel recalibration operation and a 1×1 convolution operation; the HRT-C3K2 module extracts features from the first feature again to obtain a second feature; the fusion module is used to fuse a preset number of second features to obtain multi-scale features; the recognition module is used to identify information of power equipment based on the multi-scale features.

[0118] For example, referring to Figure 3 b. The object detection model based on the YOLOv11 model includes a feature extraction module, a fusion module, and a recognition module, Head. The feature extraction module, Backbone, is used to extract features from infrared power images. It includes a preset number of HRTConv and HRT-C3K2 modules. Each HRTConv module divides the input infrared power image features into two parts, one for standard convolution to extract standard features, and the other for lightweight convolution to extract lightweight features. The standard and lightweight features are then concatenated to obtain the first feature. Specifically, the HRTConv module includes a standard convolution module, a lightweight convolution module, a concatenation module, and a channel shuffle module. The standard convolution module performs standard convolution on one part of the infrared power image to extract standard features. The lightweight convolution module performs 3×3 depthwise convolution on the other part of the infrared power image to extract intermediate features. After performing channel recalibration on the intermediate features, 1×1 convolution is performed on some of the channels to extract lightweight features. The splicing module is used to splice the standard features and lightweight features. The channel shuffling module is used to divide the spliced ​​standard features and lightweight features into two groups along the channel dimension, and then transpose and flatten each group to obtain the first feature. The HRT-C3K2 module extracts the first feature again to obtain the second feature.

[0119] Furthermore, the second feature is pooled at multiple scales using SPPF (Spatial Pyramid Pooling Fast) to improve the model’s ability to detect small targets.

[0120] Furthermore, the second feature after multi-scale pooling is processed with channel attention and spatial attention through C2PSA (Channel and PositionSpatialAttention). Specifically, channel weights are generated through global average pooling (GAP) and fully connected layers to strengthen the feature responses of important channels. At the same time, a spatial weight map is generated through convolution operations to highlight the spatial position information of the target area. The channel weights and spatial weights are fused by addition or multiplication to form the final feature enhancement weight. The second feature after multi-scale pooling is dynamically weighted by the feature enhancement weight, thereby improving the positioning accuracy of the feature.

[0121] Furthermore, a fusion module is constructed by replacing the C3K2 module in the YOLOv11 model with the HRT-C3K2 module. The fusion module fuses a preset number of second features to obtain multi-scale features.

[0122] Finally, the multi-scale features output by the fusion module Neck are converted into target category information and positioning information through the recognition module Head.

[0123] like Figure 8 As shown, Figure 8 This is a flow chart of another embodiment of an infrared power detection method provided by the present invention. An infrared power detection method is applied to a target detection model based on the YOLOv11 model, and the method includes:

[0124] S201: Collect infrared power images containing various power equipment.

[0125] S202: Input the image into the target detection model based on the YOLOv11 model, which includes: a feature extraction module, a fusion module and an identification module; the feature extraction module is used to extract features of the infrared power image, and the feature extraction module includes: a preset number of HRTConv modules and HRT-C3K2 modules. Each HRTConv module divides the features of the input infrared power image into two parts according to the channel, one part performs a standard convolution operation to extract standard features, and the other part performs a lightweight convolution operation to extract lightweight features, and the standard features and the lightweight features are spliced ​​to obtain a first feature; the lightweight convolution operation includes a 3×3 depth convolution operation, a channel recalibration operation and a 1×1 convolution operation; the HRT-C3K2 module extracts features from the first feature again to obtain a second feature; the fusion module is used to fuse a preset number of second features to obtain multi-scale features; the identification module is used to identify information of power equipment based on multi-scale features.

[0126] It should be noted that steps S201-S202 are Figure 7 The implementation scenario shown has been discussed in detail and will not be repeated here.

[0127] S203: Using the infrared power image as a data set, a target detection model based on the YOLOv11 model is trained using the data set.

[0128] For example, it is necessary to train the target detection model based on the YOLOv11 model on the collected infrared power equipment data set in order to obtain the model parameters that can be used for infrared power equipment detection. The training phase is driven by massive data, and the back propagation algorithm is used to adjust the network parameters of the entire target detection model including the HTConv and HRT-C3K2 modules to minimize the difference between the predicted results and the true labels (loss function), and the optimizer is combined to achieve efficient convergence. After the training is completed, HTConv and HRT-C3K2 have automatically learned the characteristic information of the infrared power equipment. The algorithm proposed in this embodiment uses the pytorch framework, the batch size is set to 32, and a total of 350 generations are trained.

[0129] like Figure 9 As shown, Figure 9 The device 10 includes a memory 11 and a processor 12. The memory 11 stores a computer program, and the processor 12 executes the computer program when working to achieve the following. Figure 7 and Figure 8 The method shown.

[0130] The specific technical details of an infrared power detection method implemented when the above-mentioned device 10 executes a computer program have been discussed in detail in the above-mentioned method steps, so they will not be repeated here.

[0131] like Figure 10 As shown, Figure 10 The structure diagram of an embodiment of the medium provided by the present invention. The medium 20 stores at least one computer program 21, which is executed by the processor 12 to implement the following Figure 7 and Figure 8 In one embodiment, the storage medium 20 may be a memory chip, a hard disk, a mobile hard disk, a USB flash drive, an optical disk, or other readable and writable storage tools, or a server.

[0132] Additionally, the processes depicted in the accompanying figures do not necessarily have to be performed in the particular order shown, or sequential order, to achieve desired results. In certain implementations, multitasking and parallel processing are also possible or may be advantageous.

[0133] The various embodiments in this specification are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from the other embodiments. In particular, the device, apparatus, and non-volatile computer-readable storage medium embodiments are described briefly because they are generally similar to the method embodiments. For relevant portions, refer to the description of the method embodiments.

[0134] The apparatus, device, non-volatile computer-readable storage medium and method provided in the embodiments of this specification correspond to each other. Therefore, the apparatus, device, and non-volatile computer storage medium also have similar beneficial technical effects as the corresponding method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the corresponding apparatus, device, and non-volatile computer storage medium will not be repeated here.

[0135] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0136] For the convenience of description, when describing the above device, various units are divided into functions and described separately. Of course, when implementing this specification, the functions of each unit can be implemented in the same one or more software and / or hardware. It should be understood by those skilled in the art that this specification embodiment can be provided as a method, system, or computer program product. Therefore, this specification embodiment can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this specification embodiment can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0137] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0138] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0140] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0141] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0142] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0143] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0144] This specification may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0145] The various embodiments in this specification are described in a progressive manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiments are described briefly because they are generally similar to the method embodiments. For relevant parts, refer to the description of the method embodiments.

[0146] The above disclosure is merely a preferred embodiment of the present invention and certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A target detection model based on the YOLOv11 model, characterized in that: The model includes: Feature extraction module, fusion module and recognition module; The feature extraction module is used to extract features from infrared power images. The feature extraction module includes: a preset number of HRTConv modules and HRT-C3K2 modules. Each HRTConv module divides the features of the input infrared power image into two parts according to the channel, performs a standard convolution operation on one part to extract standard features, and performs a lightweight convolution operation on the other part to extract lightweight features, and then splices the standard features and the lightweight features to obtain a first feature; The lightweight convolution operation includes a 3×3 depth convolution operation, a channel recalibration operation, and a 1×1 convolution operation; The HRT-C3K2 module extracts the first feature again to obtain a second feature; a fusion module, configured to fuse a preset number of the second features to obtain a multi-scale feature; An identification module is used to identify information of the electric power equipment according to the multi-scale features.

2. The target detection model based on the YOLOv11 model according to claim 1, characterized in that An HRT-BottleNeck module is constructed using the HRTConv module. The HRT-BottleNeck module includes the first HRTConv module and the second HRTConv module. The first HRTConv module halves the number of channels of the first feature, and the second HRTConv module restores the number of channels to be consistent with the number of channels of the first feature. The HRT-C3K2 module is constructed using the HRT-BottleNeck module.

3. A target detection model based on the YOLOv11 model according to claim 2, characterized in that The HRT-C3K2 module is constructed according to different values ​​of C3k, where C3 represents three convolutional layers and k represents a convolution kernel. The C3K2 module in the YOLOv11 model includes a BottleNeck module and a C3 module. When C3k is True, the BottleNeck module in the C3K2 module is replaced by the HRT-BottleNeck to obtain the HRT-C3K2 module. When C3k is false, the BottleNeck module in the C3 module is replaced by the HRT-BottleNeck module to construct an HRT-C3 module. The C3 module in the C3K2 module is replaced by the HRT-C3 module to obtain the HRT-C3K2 module.

4. A target detection model based on the YOLOv11 model according to claim 1, characterized in that, The HRTConv module includes a standard convolution module, a lightweight convolution module, a splicing module, and a channel shuffling module. The standard convolution module is used to perform a standard convolution operation on a portion of the infrared power image to extract standard features; The lightweight convolution module is used to perform a 3×3 depth convolution operation on another part of the infrared power image to extract intermediate features, perform a channel recalibration operation on the intermediate features, and then perform a 1×1 convolution operation on some channels to extract lightweight features; A splicing module, used for splicing the standard features and the lightweight features; The channel shuffling module is used to divide the spliced ​​standard features and lightweight features into two groups along the channel dimension, and perform a transposition and then a flattening operation on each group to obtain a first feature.

5. The target detection model based on the YOLOv11 model according to claim 3, characterized in that The fusion module is constructed by replacing the C3K2 module in the YOLOv11 model with the HRT-C3K2 module.

6. The target detection model based on the YOLOv11 model according to claim 4, characterized in that The channel number allocation ratio of the standard convolution module and the lightweight convolution module is controlled by the scaling factor: if the scaling factor is 0, the standard convolution module is fully adopted; if the scaling factor is 1, the lightweight convolution module is fully adopted.

7. An infrared power detection method, the method being applied to the target detection model based on the YOLOv11 model according to any one of claims 1 to 6, characterized in that: The method comprises: Collect infrared power images containing various power equipment; The image is input into the object detection model based on the YOLOv11 model, which includes: Feature extraction module, fusion module and recognition module; the feature extraction module is used to extract features of infrared power images, and the feature extraction module includes: a preset number of HRTConv modules and HRT-C3K2 modules, each of the HRTConv modules divides the features of the input infrared power image into two parts according to the channel, one part performs a standard convolution operation to extract standard features, and the other part performs a lightweight convolution operation to extract lightweight features, and the standard features and lightweight features are spliced ​​to obtain a first feature; the lightweight convolution operation includes a 3×3 depth convolution operation, a channel recalibration operation and a 1×1 convolution operation; the HRT-C3K2 module performs feature extraction on the first feature again to obtain a second feature; the fusion module is used to fuse a preset number of the second features to obtain multi-scale features; the recognition module is used to identify information of power equipment based on the multi-scale features.

8. The infrared power detection method according to claim 7, characterized in that: After feature extraction is performed on the infrared power image, feature splicing is performed to obtain multi-level features. The features at different levels represent targets of different sizes, specifically including: Performing a standard convolution operation on a portion of the infrared power image to extract standard features; Performing a 3×3 depthwise convolution operation on another part of the infrared power image to extract intermediate features, performing a channel recalibration operation on the intermediate features, and then performing a 1×1 convolution operation on the intermediate features to extract lightweight features; splicing the standard features and the lightweight features; Along the channel dimension, the spliced ​​standard features and lightweight features are divided into two groups, and each group is transposed and then flattened to obtain the first feature.

9. The infrared power detection method according to claim 8, characterized in that: The method further comprises: The infrared power image is used as a data set, and the target detection model based on the YOLOv11 model is trained using the data set.

10. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 7 to 9.

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 7 to 9.

Images