Method and device for detecting environmental protection risk target in power transmission and transformation project construction period, electronic equipment and storage medium

Abstract

The invention provides an environmental protection risk target detection method and device for a power transmission and transformation project construction period, electronic equipment and a storage medium, and relates to the technical field of computer vision, and the method comprises the steps: obtaining an unmanned aerial vehicle aerial image of the power transmission and transformation project construction period; inputting the unmanned aerial vehicle aerial image into a pre-trained target detection model, and outputting an environmental protection risk target; wherein the target detection model is obtained through training based on an improved YOLOv11 convolutional neural network model which constructs a shallow feature detection layer and integrates an RCS-OSA convolution module. According to the method provided by the invention, detail information is effectively reserved and utilized through a superficial layer feature enhanced multi-scale fusion architecture, and the capturing capability of a tiny target is remarkably improved; the RCS-OSA module optimizes interaction and fusion of multi-scale features, reduces redundant parameters, and improves the multiplexing efficiency of local features.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and in particular to a power transmission and transformation project construction period environmental protection risk target detection method and device, electronic equipment and storage medium. BACKGROUND

[0002] Target detection technology has important application value in the field of intelligent monitoring of environmental protection risks in power transmission and transformation projects. With the popularization of unmanned aerial vehicle aerial photography technology, image analysis methods based on deep learning have become the core means for identifying environmental protection risk factors (such as material stacking and tower foundation disturbance) during the construction period. However, the power transmission and transformation project scene has the characteristics of large target size span, high proportion of small targets (<50x50 pixels), complex background interference (such as metal reflection, cloud and fog blocking), etc. Traditional target detection models are prone to problems such as shallow feature loss, inefficient multi-scale fusion, and light sensitivity, which can lead to high small target miss detection rate and false detection rate. Especially in backlight or low contrast scenes, tower foundation structure blur, building material reflection and other interference can significantly reduce the generalization ability of existing algorithms, making it difficult to meet the real-time monitoring needs of the project.

[0003] How to improve the accuracy of identifying small target objects in unmanned aerial vehicle aerial photography images during the construction period of power transmission and transformation projects is a technical problem that needs to be solved. SUMMARY

[0004] The present application provides a power transmission and transformation project construction period environmental protection risk target detection method, device, electronic equipment and storage medium to solve the defects in the prior art.

[0005] The present application provides a power transmission and transformation project construction period environmental protection risk target detection method, device, electronic equipment and storage medium to solve the defects in the prior art. Obtain an unmanned aerial vehicle aerial photography image during the construction period of a power transmission and transformation project; Input the unmanned aerial vehicle aerial photography image into a pre-trained target detection model and output an environmental protection risk target. The target detection model is an improved YOLOv11 convolutional neural network model based on the construction of a shallow feature detection layer and the integration of an RCS-OSA convolutional module.

[0006] According to the power transmission and transformation project construction period environmental protection risk target detection method provided by the present application, the unmanned aerial vehicle aerial photography image is input into the pre-trained target detection model, and the environmental protection risk target is output, which includes: Pretreat the unmanned aerial vehicle aerial photography image by a multi-scale Retinex image enhancement algorithm MSRCR with color restoration function to obtain a data enhanced image; Input the data enhanced image into the target detection model and output the environmental protection risk target.

[0007] The application provides a power transmission and transformation project construction period environmental protection risk target detection method, and a target detection model is constructed. The shallow feature detection layer of the second-level pyramid feature layer P2 is explicitly constructed in the backbone network of the YOLOv11 convolutional neural network model, and the shallow feature detection layer is used to perform twice convolution operation on an image with an input resolution of 640*640 to generate a shallow feature map with a resolution of 160*160. The C3K2 convolution module in the backbone network of the YOLOv11 convolutional neural network model is replaced by an RCS-OSA convolution module, and the C3K2 convolution module in the shallow feature path of the P2 layer in the neck network of the YOLOv11 convolutional neural network model is replaced by an RCS-OSA convolution module, and the RCS-OSA convolution module is used to realize feature multiplexing through double-path processing on input features. The target detection model is constructed based on the constructed shallow feature detection layer and the integrated RCS-OSA convolution module.

[0008] The double path includes a unit branch and an enhanced branch, the unit branch is used to keep the original input features for transmission and output first features, and the enhanced branch is used to extract deep features through stacked RCS modules and output second features. The RCS-OSA convolution module realizes feature multiplexing through double-path processing on input features, and the double-path processing includes: The input features are input into the unit branch and the enhanced branch through the RCS-OSA convolution module, and the first features and the second features are output. The first features and the second features are interactively fused through channel shuffling, and fused features are output.

[0009] The output layer of the target detection model is the second-level pyramid feature layer P2, the third-level pyramid feature layer P3, the fourth-level pyramid feature layer P4 and the fifth-level pyramid feature layer P5.

[0010] The application provides a power transmission and transformation project construction period environmental protection risk target detection method, and a target detection model is constructed. In the training process, the original loss function is replaced by a boundary box loss function WIoU-v3 loss function based on a dynamic non-monotonic focusing mechanism. The WIoU-v3 loss function is realized by the following formula: wherein, and denote the size of the smallest enclosing box, and are hyperparameters used to adjust the size of the box, denotes a small gradient gain assigned.

[0011] According to the power transmission and transformation project construction period environmental protection risk target detection method provided by the application, the unmanned aerial vehicle aerial image is input into the pre-trained target detection model, and after outputting the environmental protection risk target, the method further comprises: Comparing the environmental protection risk target output by the target detection model with the actually detected environmental protection risk target, an evaluation result of the target detection model is obtained; Based on the evaluation result of the target detection model, the parameters of the target detection model are adjusted to obtain an updated target detection model.

[0012] The application also provides a power transmission and transformation project construction period environmental protection risk target detection device, comprising the following modules: An acquisition module is configured to acquire an unmanned aerial vehicle aerial image of a power transmission and transformation project construction period. A detection module is configured to input the unmanned aerial vehicle aerial image into a pre-trained target detection model to output an environmental protection risk target. The target detection model is an improved YOLOv11 convolutional neural network model based on the construction of a shallow feature detection layer and the integration of an RCS-OSA convolutional module.

[0013] The application also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the power transmission and transformation project construction period environmental protection risk target detection method as described above.

[0014] The application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executable on a processor to implement the power transmission and transformation project construction period environmental protection risk target detection method as described above.

[0015] The application also provides a computer program product comprising a computer program, wherein the computer program is executable on a processor to implement the power transmission and transformation project construction period environmental protection risk target detection method as described above.

[0016] The application provides a power transmission and transformation project construction period environmental protection risk target detection method and device, electronic equipment and storage medium, which obtains a UAV aerial image of a power transmission and transformation project construction period; inputs the UAV aerial image into a pre-trained target detection model to output an environmental protection risk target; wherein the target detection model is an improved YOLOv11 convolutional neural network model based on construction of a shallow feature detection layer and integration of an RCS-OSA convolution module and obtained through training. Therefore, the application effectively retains and utilizes detailed information through a shallow feature enhanced multi-scale fusion architecture, significantly improves the capture ability of micro targets, optimizes the interaction and fusion of multi-scale features through an RCS-OSA module, reduces redundant parameters, and improves the reuse efficiency of local features. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0018] Figure 1 It is a flowchart of the power transmission and transformation project construction period environmental protection risk target detection method provided by the application.

[0019] Figure 2 It is a network architecture diagram of the improved YOLOv11 convolutional neural network model provided by the application.

[0020] Figure 3 It is a structure diagram of the RCS module provided by the application.

[0021] Figure 4 It is a structure diagram of the RCS-OSA module provided by the application.

[0022] Figure 5 It is an improved strategy diagram of the YOLOv11 convolutional neural network model provided by the application.

[0023] Figure 6 It is a complete flowchart of the power transmission and transformation project construction period environmental protection risk target detection method provided by the application.

[0024] Figure 7 It is a visualization result diagram on the data type of material stacking annotation data provided by the application.

[0025] Figure 8 It is a visualization result diagram on the public data set visdrone2019 provided by the application.

[0026] Figure 9 Figure 1 is a structural schematic diagram of the environmental risk target detection device for power transmission and transformation engineering construction period provided by the present application.

[0027] Figure 10 Figure 1 is a structural schematic diagram of the environmental risk target detection device for power transmission and transformation engineering construction period provided by the present application. DETAILED DESCRIPTION

[0028] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0029] Before the method of the present application is described in detail, the terms involved in the present application will be explained and described schematically.

[0030] 1) You Look Only Once Version 11 (YOLOv11): the 11th iteration version of the target detection algorithm model.

[0031] 2) Shallow Feature Detection Layer (SFDL): shallow feature detection layer, which refers to the feature map obtained by the main network early through less down-sampling. The closer to the original image space scale, the shallower, and the farther, the deeper.

[0032] 3) The RepVGG / RepConv ShuffleNet (RCS) and a OneShot-Aggregation of the RCS (RCS-OSA): RCS (Reparameterized Convolution based on channel Shuffle): Reparameterized Convolution based on channel Shuffle; OSA (One-Shot Aggregation): One-Shot Aggregation of multiple feature cascades.

[0033] 4) Wise-IoU (WIoU): boundary box loss based on dynamic non-monotonic focusing mechanism.

[0034] 5) The Multi-Scale Retinex with Color Restoration (MSRCR): Multi-Scale Retinex image enhancement algorithm with color restoration function.

[0035] 6)Bounding box regression (BBR): bounding box regression.

[0036] The application will be described below in conjunction with Figures 1-10 A power transmission and transformation project construction period environmental protection risk target detection method and device, electronic equipment and storage medium are described.

[0037] It should be noted that the current small target detection scheme for unmanned aerial vehicle aerial image is to add a shallow feature detection layer in the network structure of YOLOv11 model. For example, by adding a shallow feature extraction network in the P1 layer of the YOLOv5 backbone network, more details of small target objects can be better preserved. In the YOLOv8 backbone network, an additional shallow feature detection layer and a head network are designed, which can alleviate the problem of insufficient small target recognition accuracy caused by the dramatic change of the scale of unmanned aerial vehicle aerial image, and use the lightweight module MobileNetv3 as the feature extraction network, which significantly reduces the parameter redundancy and computational complexity. Research on adding a high-resolution shallow feature detection layer in the YOLOv10 backbone network to obtain more detailed information of small target objects, and using a double-layer routing attention algorithm (Bi-Level Routing Attention) in the backbone, which effectively reduces the complex environmental interference of unmanned aerial vehicle images and improves the attention of the model to key information. In the YOLOv11 backbone network, five times of downsampling generate five multi-scale feature maps, which provide the network with multi-level features from shallow to deep and from high resolution to low resolution. The application selects P2, P3, P4 and P5 as the output layer in the final detection, which takes into account the detection needs of targets of different scales, while retaining deep global semantic information. The newly added P2 output layer and shallow detection branch significantly enhance the ability to capture and locate the details of low-pixel environmental protection risk targets, making the overall detection more comprehensive and accurate. This improvement improves the detection performance of small targets in the recognition of unmanned aerial vehicle aerial data in the power transmission and transformation project construction period.

[0038] However, when the shallow feature detection layer is added in the YOLOv11 target detection model, the backbone network needs to retain high-resolution feature maps, resulting in an increase in network computation and a decrease in inference speed, which conflicts with the real-time requirements of engineering construction scenes. The RCSOSA structure reparameterizes the convolution module to solve the problem of parameter redundancy and reduced inference speed using a lighter network structure. For example, research on the YOLOv8 network architecture integrates the RCSOSA reparameterized convolution module to replace the traditional C2f module, which improves the feature extraction capability of the network architecture while maintaining computational efficiency. Based on the improved YOLOv8n model, an additional small target detection layer is introduced to extract high-resolution shallow features from unmanned aerial vehicle power inspection data for multi-scale feature extraction and fusion. By integrating the RCSOSA reparameterized convolution module into the shallow layer of the YOLOv11 backbone and neck network, the model's detection accuracy and speed are effectively improved, meeting the real-time inspection and recognition requirements of transmission and transformation engineering construction. Compared with the original Yolov11 model, the proposed method provides multiple receptive field diversity features and enhances the network's sensitivity to different scales, which helps improve the model's detection capability for targets of different sizes. At the same time, the network's efficiency in handling dense connections is improved.

[0039] Secondly, the small pixel ratio in unmanned aerial vehicle image data makes it difficult to distinguish targets from background noise, so choosing the right loss function to improve the detection accuracy of these targets is also crucial. Wise-IoU (WIoU) introduces a dynamic focusing IoU variant as an alternative to the basic IoU to improve the Bounding box regression (BBR) loss. Based on the YOLOv9s network framework, the WioU-v3 loss function is used to replace the original CIoU loss function, and the WIoU mechanism can dynamically adjust the loss weight according to the size of the detected object, further improving the positioning ability of small targets. Based on the YOLOv11 network, the WIoU is used as an alternative to the basic IoU, which can significantly overcome the challenges of uneven distribution or large scale changes in the detection of metal roof tiles in the ground obtained by unmanned aerial vehicles. The previous YOLOv11 did not set the corresponding loss function, and the addition of the WIoU mechanism can overcome the problem of false detection and missed detection caused by different target scales in the aerial data of transmission and transformation engineering construction scenes.

[0040] In the prior art: (1) Insufficient utilization of shallow features leads to small target missed detection problem.

[0041] Small targets are small in size, mainly reflected in shallow features. Traditional target detection networks usually focus more on deep semantic features. Shallow features lack sufficient semantic information and are often underestimated or not fully utilized in the network, which can easily cause small targets to be ignored or missed during detection.

[0042] (2) Model parameter redundancy.

[0043] The addition of a shallow feature detection layer requires the preservation of higher resolution feature maps, and the complex network structure leads to parameter redundancy and high computational complexity. Model parameter redundancy increases memory and computational burden, reduces inference speed, and affects real-time detection performance.

[0044] (3) Generalization defects caused by unbalanced punishment of low-quality samples.

[0045] During training, the IoU loss function of the YOLOv11 algorithm may over-punish low-quality samples with more noise or inaccurate labeling, and may not optimize high-quality samples enough. This unbalanced punishment mechanism can cause the model to pay excessive attention to low-quality samples that can produce large gradients, resulting in poor performance of the model when generalizing, and the model cannot adapt well to the target detection task in the power transmission construction scene.

[0046] (4) Feature degradation problem under complex light interference.

[0047] The details of the target in the backlight scene of the unmanned aerial vehicle image data during the power transmission project construction period are lost, and the cloud cover and metal building material reflection cause the feature extraction layer to fail to capture stable features. This will reduce the network's ability to distinguish targets, and the detection accuracy of existing algorithms in the multi-scale heterogeneous construction scene will decrease significantly Based on this, the present application provides a power transmission project construction period environmental risk target detection method to solve at least one of the above problems.

[0048] Figure 1 The present application provides a flowchart of the power transmission project construction period environmental risk target detection method, as shown in Figure 1 The method comprises the following steps: Step 100, acquiring an unmanned aerial vehicle aerial image during the power transmission project construction period.

[0049] Specifically, during the construction process, the data obtained by the unmanned aerial vehicle due to the complex ground environment, uneven light, etc. can cause the boundary of the image target to be unclear or the features to be not obvious, which becomes the main obstacle for identifying environmental risk factors. In the data preprocessing stage, the MSRCR data enhancement module is applied to improve the overall brightness of the data and enrich the detailed features, providing better data support for subsequent network model target recognition. The MSRCR algorithm can be referred to the following formula: wherein, is the expression of MSRCR algorithm, representing the repaired image filtered through the first channel. is a color restoration factor to eliminate local color distortion, and its expression can be referred to the following formula: wherein, represents a gain constant, represents a controlled nonlinear factor.

[0050] Step 200, input the unmanned aerial vehicle aerial image into the pre-trained target detection model, and output the environmental protection risk target; The target detection model is obtained by training an improved YOLOv11 convolutional neural network model based on constructing a shallow feature detection layer and integrating an RCS-OSA convolutional module.

[0051] Figure 2 is the network architecture diagram of the improved YOLOv11 convolutional neural network model provided by the application, and the following will be described in combination with Figure 2 , the target detection model based on the improved YOLOv11 convolutional neural network model provided by the embodiment of the application is specifically described.

[0052] The construction process of the target detection model comprises: A shallow feature detection layer of a pyramid second-level feature layer P2 layer is explicitly constructed in the backbone network of the YOLOv11 convolutional neural network model, wherein the shallow feature detection layer is used to perform twice convolution operation on an input image with a resolution of 640*640 to generate a shallow feature map with a resolution of 160*160. The C3K2 convolutional module in the backbone network of the YOLOv11 convolutional neural network model is replaced by an RCS-OSA convolutional module, and the C3K2 convolutional module in the shallow feature path of the neck network of the YOLOv11 convolutional neural network model for processing the P2 layer features is replaced by an RCS-OSA convolutional module, wherein the RCS-OSA convolutional module is used to realize feature multiplexing through double-path processing on input features. Based on the constructed shallow feature detection layer and the integrated RCS-OSA convolutional module, the target detection model is constructed.

[0053] In one embodiment, on the basis of the YOLOv11 network, a shallow feature detection layer with high resolution at the P2 scale level is introduced to solve the problem of insufficient detection of small target object detail information in the original model. Unlike some research schemes that increase the shallow feature detection layer in the neck network of the YOLO model by upsampling operation, the present application introduces a feature map with a resolution of 160x160 from the neck network after two convolution operations on the input image with a resolution of 640x640 in the backbone network of YOLOv11. At this time, the high-resolution shallow feature map obtained after two convolution operations has more detailed information of small targets, and the loss of small target object detail information in the original image will not occur when upsampling operation is performed in the neck network. The improved model uses detectors on four scales P1, P2, P3 and P4 to enhance the stability of the model when processing targets of different scales.

[0054] Figure 3 is a structural schematic diagram of the RCS module provided by the present application, Figure 4 is a structural schematic diagram of the RCS-OSA module provided by the present application, Figure 5 is an improved strategy schematic diagram of the YOLOv11 convolutional neural network model provided by the present application, referring to Figure 3 、 Figure 4 and Figure 5 , the C3K2 block in the backbone network of the YOLOv11 target detection model is completely replaced by the RCSOSA block, and the C3K2 block in the newly added shallow feature extraction part of the neck network is replaced, realizing the collaborative optimization of calculation efficiency and feature expression ability. In the architecture design of the RCSOSA composite module, the feature reuse is realized through a double-path processing mechanism. The module divides the feature flow into two parallel branches: the unit branch directly transmits the original feature, and the enhanced branch performs deep feature extraction through the stacked RCS module. After cross-layer connection processing, the features of the two paths are fused through channel mixing. This double-path collaborative mechanism is integrated into the YOLOv11 target detection framework, effectively balancing the contradiction between feature reuse efficiency and calculation complexity through a single aggregation strategy, and only implementing feature fusion at the final feature layer, significantly reducing memory consumption and calculation redundancy. This design paradigm not only improves the small target detection accuracy, but also ensures the deployment feasibility of the model on mobile terminal devices.

[0055] Further, the training process of the target detection model comprises: In the training process, the original loss function is replaced by a boundary box loss function WIoU-v3 loss function based on a dynamic non-monotonic focusing mechanism; The WIoU-v3 loss function is realized by the following formula: wherein, and denote the size of the minimum bounding box, magnifying the common quality anchor box . The dynamic adjustment mechanism with as a parameter ensures that when , , and are hyperparameters for adjusting size, denotes a small gradient gain assigned.

[0056] Specifically, in the YOLOv11 target detection model, the WIoU-v3 loss function is adopted to replace the original loss function, so as to guide the model to the positioning of the common quality box, and the attention to the poor quality box is improved.

[0057] The above is a step-by-step description of the power transmission project construction period environmental risk target detection method provided by the present application. From the above step description, it can be seen that according to the power transmission project construction period environmental risk target detection method provided by the present application, the unmanned aerial vehicle aerial image of the power transmission project construction period is obtained; the unmanned aerial vehicle aerial image is input into the pre-trained target detection model, and the environmental risk target is output; wherein the target detection model is: an improved YOLOv11 convolutional neural network model based on constructing a shallow feature detection layer and integrating an RCS-OSA convolutional module is obtained after training. Therefore, it can be seen that the present application effectively retains and utilizes detailed information through the multi-scale fusion architecture of shallow feature enhancement, significantly improves the capture ability of micro targets; the RCS-OSA module optimizes the interaction and fusion of multi-scale features, reduces redundant parameters, and improves the reuse efficiency of local features.

[0058] Based on the above embodiment, in the present embodiment, after step 200 inputs the unmanned aerial vehicle aerial image into the pre-trained target detection model and outputs the environmental risk target, the method further comprises: Comparing the environmental risk target output by the target detection model with the actual detected environmental risk target to obtain the evaluation result of the target detection model; Based on the evaluation result of the target detection model, the parameters of the target detection model are adjusted to obtain an updated target detection model.

[0059] The environmental protection risk target detection method for power transmission and transformation project construction period provided by the embodiment improves the robustness of the feedback iteration improvement model under extreme environment, optimizes the weight distribution of the P2 layer shallow feature based on the evaluation result, and relieves the small target missed detection caused by insufficient utilization of shallow features.

[0060] Figure 6 is the complete flowchart of the environmental protection risk target detection method for power transmission and transformation project construction period provided by the application, and the following will be combined with Figure 6 to make a complete description of the environmental protection risk target detection method for power transmission and transformation project construction period provided by the application.

[0061] Step 1, acquiring the unmanned aerial vehicle aerial image of the power transmission and transformation project construction period.

[0062] Step 2, data enhancement is performed through MSRCR.

[0063] Step 3, data annotation is performed by using labelmg image annotation tool.

[0064] Step 4, the data set after annotation is divided.

[0065] Step 5, the data is input into the improved YOLOv11 target detection model for training.

[0066] Step 6, the model is evaluated and analyzed, the parameters are adjusted, and the feedback is fed back to the training module.

[0067] The application aims to explore a small target detection model in a multi-scale heterogeneous construction scene based on a YOLOv11 target detection algorithm, and to establish an efficient identification model of environmental protection risk factors in the power transmission and transformation project construction period based on the same. In the application, in view of the low accuracy of identifying small target objects in the unmanned aerial vehicle aerial image in the power transmission and transformation project construction period, a solution method of adding a shallow feature detection layer (shallow feature detection layer, SFDL), integrating an RCS-OSA reparameterization convolution module and introducing a Wise-IoUv3 loss of dynamic non-monotonic focusing mechanism is proposed.

[0068] First, for the environmental protection risk targets with unclear characteristics such as material stacking during the construction of power transmission towers, tower foundation disturbance, and small pixel ratio, a shallow feature detection layer is newly added in the main network structure of YOLOv11, the original skeleton structure is modified, and the small target object rich detail information is obtained to enhance the model's ability to identify small targets in the unmanned aerial vehicle aerial image of the construction site. The specific implementation is to newly add a high-resolution detector at the P2 layer in the main network of YOLOv11, and two-step convolution operation is performed on the input image with a resolution of 640x640 to obtain a shallow feature map with a resolution of 160x160. The above operation helps to retain more high-resolution spatial details. The improved YOLOv11 network model generates five different scale feature maps, namely 320x320 (P1), 160x160 (P2), 80x80 (P3), 40x40 (P4) and 20x20 (P5) in the main network through five convolution operations, and the model finally uses the P2, P3, P4 and P5 scales as the output layer.

[0069] Secondly, in view of the problems of parameter redundancy and insufficient multi-scale feature interaction of the high-precision target detection model in the power transmission engineering scene, an improved method is proposed to replace the C3K2 convolution module in the main network of YOLOv11 by integrating the RCS-OSA structure to re-parameterize the convolution module. The RCS-OSA module is proposed by combining RCS and One-Shot Aggregation (OSA), and the module realizes function reuse by repeatedly stacking RCS modules. Three feature cascades are retained on the OSA path, which aggregates multiple features in one iteration, thereby reducing the computational burden of the network and improving the computational efficiency, and realizing high-precision fast reasoning. Under the premise of ensuring real-time inference speed, it is expected to alleviate the ambiguity problem of small target feature expression caused by low resolution and weak texture in the power transmission engineering construction scene.

[0070] Thirdly, in order to overcome the challenge that the target scale changes greatly in the environmental protection image data obtained by the unmanned aerial vehicle, and the small target object is difficult to distinguish from the background noise, the Wise-IoU (WIoU) loss function is selected as the substitute of the basic IoU loss function. Since the labeled data contains some low-quality samples, the calculation of the aspect ratio will increase the punishment effect on low-quality samples, leading to the decline of the model generalization performance. The WIoU-v3 dynamically adjusts the gradient gain of the bounding box by introducing a dynamic focal coefficient, optimizes the weighting of small targets, and improves the detection accuracy of small target objects in the unmanned aerial vehicle data.

[0071] Finally, in order to solve the problems of blurred tower foundation structure in backlight scene, loss of small target details in shadow area, and detection performance degradation caused by cloud and mist obstruction and metal building material reflection in the unmanned aerial vehicle image data obtained on the engineering construction site, the MSRCR algorithm is used as the core module for data preprocessing, and the image enhancement effect of the algorithm can well solve the above problems. The unmanned aerial vehicle aerial data of the power transmission and transformation project construction period of the State Grid Corporation of China is divided into three environmental protection risk scenarios, namely, the project construction preparation period, the project construction period and the ecological recovery period, and the MSRCR algorithm is used for preprocessing, and the data is divided into training set, validation set and test set (the ratio is 8:1:1). With the three types of risk factor data of power transmission tower, tower foundation and material stacking, the improved YOLOv11 target detection model is trained. The proposal of the technology provides a new idea for the identification of related environmental protection risk scenarios in the project construction period.

[0072] Table 1 shows the specific improvements of the present application and the effects brought by the improvements. Referring to Table 1, with the continuous iteration and update of the YOLO target detection algorithm model, compared with the previous version, the YOLOv11 algorithm has higher flexibility and speed, and is more suitable for real-time identification and monitoring of environmental protection risks in the power transmission and transformation project construction period, and provides a new technical solution for exploring a target detection model that is more suitable for engineering environmental protection risk perception.

[0073] Table 1

[0074] Based on the above embodiment, in the present embodiment, the effect of target detection of the environmental protection risk target detection method for the power transmission and transformation project construction period provided by the present application is described.

[0075] The experimental results on the data type of material stacking annotation data can be seen from Table 2.

[0076] Table 2

[0077] Figure 7 is the visualization result schematic diagram provided by the present application on the data type of material stacking annotation data. The visualization result can be seen from Figure 7 .

[0078] The ablation experiment results on the data type of material stacking annotation data can be seen from Table 3.

[0079] Table 3

[0080] The experimental results on the public data set visdrone2019 can be seen from Table 4.

[0081] Table 4

[0082] Figure 8 is the visualization result schematic diagram provided by the present application on the public data set visdrone2019, and the visualization result can be seen in Figure 8 .

[0083] The environmental protection risk target detection method of the power transmission and transformation project construction period provided by the present application is as follows: (1) Breakthrough in detection accuracy.

[0084] The present application effectively retains and utilizes detailed information through a multi-scale fusion architecture enhanced by shallow features, significantly improving the ability to capture small targets; the RCS-OSA module optimizes the interaction and fusion of multi-scale features, reduces redundant parameters, and improves the reuse efficiency of local features; plus the dynamic gradient adjusted WIoU-v3 loss function, which can balance the gradient distribution of low-quality samples, ensuring more accurate boundary box regression in complex situations.

[0085] (2) Balance of engineering practicability.

[0086] The present application effectively reduces the computational complexity by adopting a lightweight RCS-OSA module and optimizing the multi-scale feature fusion strategy, ensuring the real-time inference capability of the model on unmanned aerial vehicles or embedded devices. While ensuring high accuracy, the parameter quantity and computational overhead of the model are also reasonably controlled, making it suitable for popularization and application in environmental protection risk identification of power transmission and transformation engineering construction.

[0087] (3) Scene generalization ability.

[0088] The present application uses MSRCR preprocessing technology to perform image enhancement processing on unmanned aerial vehicle images under complex lighting environments, so that the model can obtain stable feature expression under different shooting conditions and adapt to complex lighting environments in power transmission and transformation construction sites; the dynamic gradient adjusted WIoU-v3 loss function can provide more reasonable gradient feedback under low-quality samples, thereby improving the generalization performance of the model.

[0089] The environmental protection risk target detection device of the power transmission and transformation project construction period provided by the present application is described as follows, and the environmental protection risk target detection device described in the following description can be mutually corresponding and referred to the environmental protection risk target detection method described in the foregoing.

[0090] Figure 9 is the structural schematic diagram of the environmental protection risk target detection device of the power transmission and transformation project construction period provided by the present application, as shown in Figure 9 The environmental protection risk target detection device of the power transmission and transformation project construction period provided by the present application comprises: The acquisition module 901 is configured to acquire a UAV aerial image of a power transmission and transformation project construction period. The detection module 902 is configured to input the UAV aerial image into a pre-trained target detection model to output an environmental protection risk target. The target detection model is an improved YOLOv11 convolutional neural network model based on construction of a shallow feature detection layer and integration of an RCS-OSA convolutional module.

[0091] The environmental protection risk target detection device for the power transmission and transformation project construction period provided by the application acquires a UAV aerial image of a power transmission and transformation project construction period, inputs the UAV aerial image into a pre-trained target detection model to output an environmental protection risk target, wherein the target detection model is an improved YOLOv11 convolutional neural network model based on construction of a shallow feature detection layer and integration of an RCS-OSA convolutional module. Therefore, the application uses a multi-scale fusion architecture with enhanced shallow features to effectively retain and utilize detailed information and significantly improve the ability to capture small targets. The RCS-OSA module optimizes the interaction and fusion of multi-scale features, reduces redundant parameters, and improves the reuse efficiency of local features.

[0092] Based on the above embodiment, in the present embodiment, the detection module 902 is specifically configured to: The UAV aerial image is preprocessed by a multi-scale Retinex image enhancement algorithm MSRCR with a color restoration function to obtain a data enhanced image. The data enhanced image is input into the target detection model to output the environmental protection risk target.

[0093] Based on the above embodiment, in the present embodiment, the device further comprises a construction module, which is specifically configured to: A shallow feature detection layer of a pyramid 2nd feature layer P2 layer is explicitly constructed in a backbone network of the YOLOv11 convolutional neural network model, wherein the shallow feature detection layer is configured to perform twice convolution operation on an image with an input resolution of 640*640 to generate a shallow feature map with a resolution of 160*160. The C3K2 convolutional module in the backbone network of the YOLOv11 convolutional neural network model is replaced by an RCS-OSA convolutional module, and the C3K2 convolutional module in the shallow feature path of the P2 layer feature in the neck network of the YOLOv11 convolutional neural network model is replaced by an RCS-OSA convolutional module, wherein the RCS-OSA convolutional module is configured to realize feature reuse through double-path processing on input features. Based on the constructed shallow feature detection layer and the integrated RCS-OSA convolution module, the target detection model is constructed.

[0094] Based on the above embodiment, in this embodiment, the double path includes a unit branch and an enhanced branch, the unit branch is used to keep the original input feature for transmission and output the first feature, and the enhanced branch is used to extract deep features through the stacked RCS module and output the second feature. The device further includes a feature multiplexing module, specifically configured to: The input feature is input into the unit branch and the enhanced branch through the RCS-OSA convolution module, and the first feature and the second feature are output. The first feature and the second feature are interactively fused through channel shuffling, and the fused feature is output.

[0095] Based on the above embodiment, in this embodiment, the output layer of the target detection model is the pyramid 2nd feature layer P2 layer, the pyramid 3rd feature layer P3 layer, the pyramid 4th feature layer P4 layer, and the pyramid 5th feature layer P5 layer.

[0096] Based on the above embodiment, in this embodiment, the device further includes a training module, specifically configured to: In the training process, the original loss function is replaced by a boundary box loss Wiou-v3 loss function based on a dynamic non-monotonic focusing mechanism; The Wiou-v3 loss function is realized by the following formula: wherein, and represent the size of the minimum closed box, and is a hyperparameter for adjusting the size of , and represents a small gradient gain assigned.

[0097] Based on the above embodiment, in this embodiment, the device further includes an adjusting module, specifically configured to: After inputting the unmanned aerial vehicle aerial image into the pre-trained target detection model and outputting the environmental protection risk target, the environmental protection risk target output by the target detection model is compared with the actually detected environmental protection risk target, and the evaluation result of the target detection model is obtained. Adjust parameters of the target detection model based on the evaluation result of the target detection model to obtain an updated target detection model.

[0098] Figure 10 An example of a schematic diagram of a physical structure of an electronic device is shown in Figure 10 The electronic device can be a robot or other electronic device, which can include a processor 1010, a communications interface 1020, a memory 1030, and a communications bus 1040, wherein the processor 1010, the communications interface 1020, and the memory 1030 can communicate with each other through the communications bus 1040. The processor 1010 can invoke logical instructions in the memory 1030 to execute the environmental risk target detection method for power transmission and transformation project construction period, including: Obtaining an unmanned aerial vehicle aerial image of the power transmission and transformation project construction period; Inputting the unmanned aerial vehicle aerial image into a pre-trained target detection model to output an environmental risk target; The target detection model is an improved YOLOv11 convolutional neural network model based on constructing a shallow feature detection layer and integrating an RCS-OSA convolutional module.

[0099] In addition, the logical instructions in the memory 1030 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various storage medium that can store program codes.

[0100] On the other hand, the present application also provides a computer program product, which includes a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the environmental risk target detection method for power transmission and transformation project construction period provided by the above-mentioned methods, including: Obtaining an unmanned aerial vehicle aerial image of the power transmission and transformation project construction period; input the unmanned aerial vehicle aerial image into a pre-trained target detection model, and output an environmental protection risk target; The target detection model is obtained by training an improved YOLOv11 convolutional neural network model based on construction of a shallow feature detection layer and integration of an RCS-OSA convolutional module.

[0101] In another aspect, the present application also provides a non-transitory computer readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the environmental protection risk target detection method for a power transmission and transformation project construction period provided by each method. obtaining an unmanned aerial vehicle aerial image of the power transmission and transformation project construction period; inputting the unmanned aerial vehicle aerial image into a pre-trained target detection model, and outputting an environmental protection risk target; The target detection model is obtained by training an improved YOLOv11 convolutional neural network model based on construction of a shallow feature detection layer and integration of an RCS-OSA convolutional module.

[0102] The device embodiments described above are only schematic, wherein the units shown as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment. Those skilled in the art can understand and implement without creative labor.

[0103] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary general hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0104] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting environmental risk targets during the construction period of a power transmission and transformation project, characterized in that: include: Obtain drone aerial images during the construction period of power transmission and transformation projects; Inputting the drone aerial image into a pre-trained target detection model to output an environmental risk target; The target detection model is obtained by training an improved YOLOv11 convolutional neural network model that constructs a shallow feature detection layer and integrates an RCS-OSA convolution module.

2. The environmental risk target detection method during the construction period of a power transmission and transformation project according to claim 1 is characterized in that: The step of inputting the drone aerial image into a pre-trained target detection model and outputting an environmental risk target includes: The UAV aerial image is preprocessed using the multi-scale Retinex image enhancement algorithm MSRCR with color restoration function to obtain a data enhanced image; The data-enhanced image is input into the target detection model, and the environmental risk target is output.

3. The environmental risk target detection method during the construction period of a power transmission and transformation project according to claim 1 is characterized in that: The process of building the target detection model includes: A shallow feature detection layer of the second-level feature layer P2 of the pyramid is explicitly constructed in the backbone network of the YOLOv11 convolutional neural network model; wherein the shallow feature detection layer is used to: perform two convolution operations on an image with an input resolution of 640×640 to generate a shallow feature map with a resolution of 160×160; The C3K2 convolution module in the backbone network of the YOLOv11 convolutional neural network model is replaced with an RCS-OSA convolution module, and the C3K2 convolution module in the shallow feature path that processes P2 layer features in the neck network of the YOLOv11 convolutional neural network model is replaced with an RCS-OSA convolution module; wherein the RCS-OSA convolution module is used to: realize feature reuse of input features through dual-path processing; Based on the constructed shallow feature detection layer and the integrated RCS-OSA convolution module, the target detection model is constructed.

4. The method for detecting environmental risk targets during the construction period of a power transmission and transformation project according to claim 3, characterized in that: The dual path includes: a unit branch and an enhancement branch, wherein the unit branch is used to maintain the original input feature for transmission and output a first feature; the enhancement branch is used to perform deep feature extraction through a stacked RCS module and output a second feature; The RCS-OSA convolution module implements feature reuse on input features through dual-path processing, including: Inputting the input features into the unit branch and the enhancement branch respectively through the RCS-OSA convolution module, and outputting the first feature and the second feature; Interactive fusion processing is performed on the first feature and the second feature through channel shuffling, and a fused feature is output.

5. The method for detecting environmental risk targets during the construction period of a power transmission and transformation project according to claim 3, characterized in that: The output layers of the target detection model are the second-level feature layer P2 of the pyramid, the third-level feature layer P3 of the pyramid, the fourth-level feature layer P4 of the pyramid, and the fifth-level feature layer P5 of the pyramid.

6. The method for detecting environmental risk targets during the construction period of a power transmission and transformation project according to claim 3, characterized in that: The training process of the target detection model includes: During training, the original loss function is replaced by the bounding box loss WIoU-v3 loss function based on the dynamic non-monotonic focusing mechanism; The WIoU-v3 loss function is implemented by the following formula: in, and Indicates the size of the minimum enclosing box. and Is used to adjust The size hyperparameter, represents a small gradient gain for the distribution.

7. The method for detecting environmental risk targets during the construction period of a power transmission and transformation project according to claim 1, characterized in that: After inputting the drone aerial image into a pre-trained target detection model and outputting an environmental risk target, the method further includes: Comparing the environmental risk target output by the target detection model with the environmental risk target actually detected to obtain an evaluation result of the target detection model; Based on the evaluation result of the target detection model, the parameters of the target detection model are adjusted to obtain an updated target detection model.

8. A device for detecting environmental risk targets during the construction period of a power transmission and transformation project, characterized in that: include: The acquisition module is used to obtain drone aerial images during the construction period of the power transmission and transformation project; A detection module, configured to input the drone aerial image into a pre-trained target detection model and output an environmental risk target; The target detection model is obtained by training an improved YOLOv11 convolutional neural network model that constructs a shallow feature detection layer and integrates an RCS-OSA convolution module.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for detecting environmental risk targets during the construction period of a power transmission and transformation project as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for detecting environmental risk targets during the construction period of a power transmission and transformation project as claimed in any one of claims 1 to 7 is implemented.

Images