An unmanned aerial vehicle cruising method and system based on a deep learning artificial intelligence image recognition algorithm

Abstract

The application discloses a kind of unmanned aerial vehicle cruise method and system based on deep learning artificial intelligence image recognition algorithm.The method is collected by unmanned aerial vehicle of improved YOLOv7-SwinT target recognition model in real time image data of inspection area, the model fuses the single-stage target detection architecture of YOLOv7 and the visual feature extraction network of Swin Transformer, and three-level recognition architecture is used to realize progressive target detection: primary abnormality detection based on light-weight CNN, intermediate accurate positioning combined with attention mechanism and advanced target classification of multi-sensor data fusion.System according to the recognition result combines the power of unmanned aerial vehicle, environmental condition and task priority, generates dynamic inspection path by adaptive path planning algorithm, and realizes multi-machine collaborative operation using intelligent task allocation algorithm.In the inspection process, sensor data is processed in real time by edge computing device, and intelligent energy management system is used to optimize charging scheduling.The application significantly improves the target recognition accuracy and cruise efficiency of unmanned aerial vehicle in complex environment, and is suitable for power inspection, security monitoring and other application scenarios.

Description

Technical Field

[0001] This invention belongs to the field of intelligent inspection technology for unmanned aerial vehicles (UAVs), specifically relating to a UAV patrol method and system based on deep learning artificial intelligence image recognition algorithms, which is particularly suitable for intelligent inspection of large infrastructure such as offshore wind farms, power facilities, and pipelines. Background Technology

[0002] With the rapid development of drone technology, its application in fields such as power line inspection, security monitoring, and agricultural plant protection is becoming increasingly widespread. However, existing drone patrol systems still face many technical challenges:

[0003] Insufficient target detection accuracy: Traditional UAV inspection systems mostly use target detection algorithms based on convolutional neural networks, such as YOLOv3 and Faster R-CNN. These algorithms have low detection accuracy for small targets in complex backgrounds, making it difficult to meet the requirements of high-precision inspection. Especially under harsh conditions such as long distances and low light, the detection performance drops significantly.

[0004] Multi-source data fusion is difficult: Existing systems usually rely on a single visual sensor and lack effective fusion of multimodal data such as infrared thermal imaging and lidar, resulting in insufficient robustness of target recognition in complex environments.

[0005] Real-time path planning is inefficient: Traditional path planning algorithms (such as A*, RRT, etc.) are difficult to dynamically adapt to environmental changes and task requirements, and cannot achieve optimal path planning under multiple constraints such as power consumption, environmental conditions and task priority.

[0006] Limited computing resources: The computing power of drone platforms is limited, and existing deep learning models have a large number of parameters and high computational complexity, making it difficult to achieve real-time processing on edge devices.

[0007] Inefficient energy management: Current drone energy management systems lack intelligent scheduling capabilities and cannot dynamically optimize charging paths based on mission progress and battery status, resulting in low inspection efficiency.

[0008] While existing research has attempted to improve target detection algorithms and optimize path planning strategies to address the aforementioned issues, a complete solution has yet to be developed. In particular, systematic innovations in areas such as multi-sensor data fusion, lightweight model design, and intelligent energy scheduling remain insufficient. Therefore, there is an urgent need to develop a drone patrol system that integrates efficient target recognition, intelligent path planning, and optimized energy management to meet the growing demands of industry applications. Summary of the Invention

[0009] In view of this, it is necessary to provide a drone patrol method and system based on deep learning artificial intelligence image recognition algorithms, aiming to solve the technical deficiencies of existing drone patrol systems in terms of target detection accuracy, multi-source data fusion, and path planning efficiency. This invention significantly improves the inspection efficiency and target recognition accuracy of drones in complex environments through an innovative three-level recognition architecture, multimodal data fusion technology, and adaptive path planning algorithm.

[0010] In a first aspect, embodiments of this application provide a method for drone navigation based on a deep learning artificial intelligence image recognition algorithm, the method comprising:

[0011] S1: Real-time image data of the inspection area is collected by a drone equipped with an improved YOLOv7-SwinT target recognition model. The improved YOLOv7-SwinT target recognition model integrates the single-stage target detection architecture of YOLOv7 and the visual feature extraction network of SwinTransformer, and adopts multi-sensor data fusion technology.

[0012] S2: Based on the target information identified in step S1, and combined with the drone's current battery level, environmental conditions, and task priority, a dynamic inspection path is generated through an adaptive path planning algorithm.

[0013] S3: Based on the inspection path generated in step S2, assign inspection tasks to multiple drones using an intelligent task allocation algorithm, and monitor the task execution status in real time.

[0014] S4: During the inspection process, sensor data is processed in real time through edge computing devices, and key data is transmitted to the control center.

[0015] S5: Based on the real-time status and mission progress of the UAV, optimize the scheduling of charging base stations and the energy allocation of the UAV through the intelligent energy management system; wherein, the improved YOLOv7-SwinT target recognition model in step S1 adopts a three-level recognition architecture, including primary anomaly detection based on lightweight CNN, intermediate precise localization combined with attention mechanism, and advanced target classification based on multi-sensor data fusion.

[0016] Optionally, in one implementation of the first aspect of the present invention, the improved YOLOv7-SwinT target recognition model specifically includes:

[0017] The YOLOv7 object detection framework includes: a backbone network constructed from depthwise separable convolutions, an adaptive anchor size selection module, and a multi-scale feature fusion pyramid structure.

[0018] The Swin Transformer feature extraction network includes a hierarchical window attention mechanism, a cross-window connection module, a local perception enhancement module, and a multi-sensor data fusion layer. It achieves feature-level fusion of visual images, infrared thermal imaging, and LiDAR point cloud data, feature weight allocation based on the attention mechanism, and spatiotemporally aligned multimodal data association.

[0019] Optionally, in one implementation of the first aspect of the present invention, the backbone network constructed by the depthwise separable convolutions includes:

[0020] A feature extraction network is used to extract multi-scale features from an input image, which includes at least one variable attention module to enhance the correlation between distant pixels and improve the feature extraction capability of small targets;

[0021] A feature fusion network is used to integrate the multi-scale features output by the feature extraction network, wherein at least one deformable convolutional module is included to expand the receptive field of the convolutional kernel and reduce the number of model parameters;

[0022] The detection head is used to detect targets based on the fused features and outputs the target's category, location, and confidence level.

[0023] The loss calculation module is used to calculate the model's loss function during training, including: localization loss, which calculates the matching degree between the predicted bounding box and the ground truth bounding box based on EIoU; target confidence loss, which measures the accuracy of the model's prediction of whether the target exists; and category prediction loss, which optimizes the classification result using label-smoothed cross-entropy loss.

[0024] The system employs Mosaic data augmentation, transfer learning, and fine-tuning strategies during the training phase to improve the model's detection accuracy for small targets in drone patrol scenarios.

[0025] Optionally, in one implementation of the first aspect of the present invention, the loss function of YOLOv7 is a multi-task loss function, consisting of three parts: localization loss L... CIoU Target confidence loss L obj And category prediction loss L cls ,

[0026] YOLOv7 uses an improved IoU loss function to calculate the localization bounding box, obtaining the regression error, and the localization loss L. CIoU The formula is as follows:

[0027]

[0028] IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box, ρ 2 (b,b gt) is the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box, l is the diagonal length of the minimum bounding rectangle, v is the aspect ratio consistency measure, and α is the weight coefficient.

[0029] Target confidence loss L obj The binary cross-entropy loss is used to measure whether the predicted bounding box contains the target, as shown in the following formula:

[0030]

[0031] S 2 B is the number of grid cells, and B is the number of anchor frames per grid cell. This is an indicator function; it is 1 if anchor frame j is responsible for detecting the target, and 0 otherwise. It is the confidence score of the prediction;

[0032] Category Prediction Loss L cls To measure the accuracy of the predicted category, a binary cross-entropy loss is used, as shown in the following formula:

[0033]

[0034] in: These are real category labels. This represents the predicted class probability, where classes is the total number of classes and c represents the number of classes.

[0035] The total loss function L of YOLOv7 total It is the weighted sum of the three loss functions mentioned above:

[0036] L total =λ coord L CIoU +λ obj L obj +λ cls L cls ,

[0037] Where, λ coord , λ obj , λ cls These are hyperparameters used to balance the weights of different tasks;

[0038] When the YOLOv7 framework includes an auxiliary head, its total loss function is L. total l :

[0039] L total l =L total +0.25L aux ,

[0040] Among them, L aux The loss function for the auxiliary head.

[0041] Optionally, in one implementation of the first aspect of the present invention, the adaptive anchor point size selection module includes:

[0042] K-means clustering units are used to perform cluster analysis on the width and height dimensions of all ground truth bounding boxes in the training dataset to generate an initial set of anchor point dimensions.

[0043] The Intersection over Union (IoU) optimization unit uses an improved IoU metric to calculate the matching degree between the predicted bounding box and the ground truth bounding box, and dynamically adjusts the distribution of anchor point sizes.

[0044] The scale-sensitive screening unit assigns a set of anchor points of a specific scale to each feature level based on the receptive field range of different feature layers in the target detection task.

[0045] The gradient backpropagation adaptation unit incorporates anchor size parameters into the network training process, optimizing the distribution of anchor size through end-to-end learning.

[0046] The gradient backpropagation adaptation unit automatically generates anchor point sizes that adapt to different detection scales by analyzing the geometric characteristics of the target dataset, thereby improving the model's detection accuracy for multi-scale targets.

[0047] Optionally, in one implementation of the first aspect of the present invention, the multi-scale feature fusion pyramid structure includes:

[0048] The bottom-up feature extraction pathway is used to extract multi-scale features of the input image layer by layer through a convolutional neural network to generate feature maps of different resolutions.

[0049] A top-down feature fusion pathway is used to fuse high-level semantic features with low-level high-resolution features step by step through upsampling operations.

[0050] The horizontal connection module uses a 1×1 convolution kernel to adjust the number of feature channels, enabling cross-scale connections between feature maps at different levels;

[0051] The feature enhancement unit introduces an attention mechanism during the feature fusion process to dynamically adjust the weight distribution of features at different scales;

[0052] A multi-level prediction output layer sets up detection heads on the fused feature maps at various scales to achieve simultaneous detection of targets of different sizes, including small, medium, and large.

[0053] The multi-scale feature fusion pyramid structure achieves complementary fusion of deep and shallow features through a bidirectional feature transfer path, thereby improving the model's detection performance for multi-scale targets.

[0054] Optionally, in one implementation of the first aspect of the present invention, based on the target information identified in step S1, and combined with the current battery level of the UAV, environmental conditions, and task priority, a dynamic inspection path is generated through an adaptive path planning algorithm, including:

[0055] S2.1: Based on the target information identified in step S1, establish an inspection target priority evaluation model. The evaluation model comprehensively considers: target anomaly score, based on the confidence score output by the deep learning model; target type weight, a pre-set priority coefficient for different types of targets; and target timeliness parameter, a time sensitivity factor calculated based on the target change rate.

[0056] S2.2: Construct a multi-constraint path planning model. Input parameters include:

[0057] The drone's current battery status and estimated flight time; environmental condition parameters, including wind speed, rainfall intensity, and visibility; task priority matrix, calculated in step S2.1; geofencing constraints;

[0058] S2.3: An improved genetic algorithm is used to solve for the optimal path, specifically including:

[0059] a) Initialize the population: Generate N feasible paths as initial solutions;

[0060] b) Fitness calculation: Construct a multi-objective fitness function based on path length, task coverage, and energy balance;

[0061] c) Selection operation: A combination of elite retention strategy and roulette wheel selection is adopted;

[0062] d) Crossover mutation: Design targeted mutation operators for inspection tasks;

[0063] e) Dynamic adjustment: Real-time monitoring of environmental changes triggers path replanning;

[0064] S2.4: Output the optimized inspection path and dynamically update it to the UAV patrol system.

[0065] Optionally, in one implementation of the first aspect of the present invention, the three-level recognition architecture includes primary anomaly detection based on lightweight CNN, intermediate precise localization combined with attention mechanism, and high-level target classification based on multi-sensor data fusion, comprising:

[0066] The primary anomaly detection module uses a lightweight CNN network to quickly extract features from the input image and achieves preliminary screening of anomaly regions through convolutional layers with low computational complexity.

[0067] The intermediate-level precise localization module, built on the Swing Transformer architecture, includes: a multi-head self-attention unit for capturing long-distance spatial dependencies; a cross-window feature interaction unit for global context modeling through a shifting window mechanism; and a dynamic receptive field adjustment unit for adaptively adjusting the attention range according to the target scale.

[0068] The advanced target classification module integrates a multi-source sensor data fusion mechanism, specifically including: a visible light and infrared feature alignment unit that uses a cross-modal feature mapping network; a spatiotemporal context fusion unit that integrates temporal sensor data through 3D convolution; and a multi-task decision layer that synchronously outputs target category, confidence level, and spatial location information.

[0069] A cascaded feature transfer path enables feature reuse and enhancement among three-level modules through residual connections and a feature pyramid structure.

[0070] The architecture employs a progressive recognition strategy to achieve high-precision detection and classification of multi-scale targets in UAV inspection scenarios under conditions of limited computing resources.

[0071] Secondly, embodiments of this application provide a drone patrol system based on a deep learning artificial intelligence image recognition algorithm, applied to the drone patrol method based on a deep learning artificial intelligence image recognition algorithm as described in the first aspect, the system comprising:

[0072] The image acquisition module is used to collect real-time image data of the inspection area by a UAV equipped with an improved YOLOv7-SwinT target recognition model. The improved YOLOv7-SwinT target recognition model integrates the single-stage target detection architecture of YOLOv7 and the visual feature extraction network of Swin Transformer, and adopts multi-sensor data fusion technology.

[0073] The path planning module is used to generate a dynamic inspection path based on the target information identified in step S1, combined with the drone's current battery level, environmental conditions, and task priority, through an adaptive path planning algorithm.

[0074] The task allocation module is used to allocate inspection tasks to multiple drones based on the inspection path generated in step S2 using an intelligent task allocation algorithm, and to monitor the task execution status in real time.

[0075] The data processing module is used to process sensor data in real time through edge computing devices during the inspection process and transmit key data to the control center.

[0076] The energy management module is used to optimize the scheduling of charging base stations and the energy allocation of drones through an intelligent energy management system based on the real-time status and mission progress of the drones. The improved YOLOv7-SwinT target recognition model in step S1 adopts a three-level recognition architecture, including primary anomaly detection based on lightweight CNN, intermediate precise localization combined with attention mechanism, and advanced target classification based on multi-sensor data fusion.

[0077] Thirdly, embodiments of this application provide an electronic device, including:

[0078] processor;

[0079] Memory used to store processor-executable instructions;

[0080] The processor is configured to implement the drone cruise method based on deep learning artificial intelligence image recognition algorithm as described in the first aspect when executing the instructions.

[0081] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the drone cruise method based on a deep learning artificial intelligence image recognition algorithm as described in the first aspect.

[0082] This invention discloses a drone patrol method and system based on a deep learning-based artificial intelligence image recognition algorithm. The method uses a drone equipped with an improved YOLOv7-SwinT target recognition model to collect real-time image data of the inspection area. This model integrates the single-stage target detection architecture of YOLOv7 and the visual feature extraction network of Swin Transformer, employing a three-level recognition architecture to achieve progressive target detection: primary anomaly detection based on lightweight CNN, intermediate precise localization combined with an attention mechanism, and advanced target classification through multi-sensor data fusion. Based on the recognition results, combined with the drone's battery level, environmental conditions, and task priority, the system generates a dynamic inspection path through an adaptive path planning algorithm and utilizes an intelligent task allocation algorithm to achieve multi-drone collaborative operation. During the inspection process, sensor data is processed in real-time through edge computing devices, and a smart energy management system optimizes charging scheduling. This invention significantly improves the target recognition accuracy and patrol efficiency of drones in complex environments and is suitable for applications such as power line inspection and security monitoring.

[0083] Beneficial effects:

[0084] (1) Significantly improved target detection performance. A three-level progressive recognition architecture (lightweight CNN initial screening + SwingTransformer precise localization + multi-sensor fusion classification) enhances the accuracy of small target detection; an improved CIoU loss function introduces aspect ratio consistency metrics, improving bounding box localization accuracy. Multimodal data fusion technology improves the robustness of recognition in complex environments.

[0085] (2) System response efficiency optimization: The combination of depthwise separable convolution and deformable convolution reduces the number of parameters in the model and improves inference speed. Edge computing architecture enables millisecond-level real-time data processing, reducing system response latency. Multi-machine collaborative task allocation algorithm improves overall inspection efficiency.

[0086] (3) Intelligent path planning: Improved genetic algorithm combined with multi-objective optimization model to shorten path planning time; dynamic replanning mechanism to improve environmental adaptability; real-time response to sudden situations such as wind speed changes; task priority evaluation model to improve the detection rate of key targets.

[0087] (4) Refined energy management. The intelligent charging scheduling system extends the effective operating time of UAVs, the energy consumption prediction algorithm enables precise power management and reduces ineffective flight distance; the multi-UAV collaborative charging strategy improves the overall energy utilization rate.

[0088] (5) Enhanced application adaptability, supporting multi-source data input from visible light / infrared / LiDAR, improving environmental adaptability; modular design allows for flexible configuration of detection models according to different task requirements; lightweight model is compatible with mainstream edge computing devices, reducing deployment costs. Attached Figure Description

[0089] Figure 1 This is a schematic flowchart of a drone cruise method based on a deep learning artificial intelligence image recognition algorithm provided in an embodiment of this application.

[0090] Figure 2 This is a structural diagram of the YOLOv7-SwinT model provided in an embodiment of this application.

[0091] Figure 3 This is a schematic diagram of a drone cruise system module based on a deep learning artificial intelligence image recognition algorithm, provided as an embodiment of this application.

[0092] Figure 4 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0093] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0094] It should be noted that, in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0095] It should be noted that in the embodiments of this application, the terms "first," "second," etc., are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order. Features specified as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0096] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0097] Example 1

[0098] Figure 1 This is a schematic flowchart of a drone cruise method based on a deep learning artificial intelligence image recognition algorithm provided in an embodiment of this application.

[0099] like Figure 1 As shown, a drone cruise method based on deep learning artificial intelligence image recognition algorithm includes:

[0100] S1: Real-time image data of the inspection area is collected by a UAV equipped with an improved YOLOv7-SwinT target recognition model. The improved YOLOv7-SwinT target recognition model integrates the single-stage target detection architecture of YOLOv7 and the visual feature extraction network of SwinTransformer, and adopts multi-sensor data fusion technology.

[0101] Image acquisition was performed using an improved YOLOv7-SwinT target recognition model. The improved YOLOv7-SwinT model integrates the single-stage target detection architecture of YOLOv7 with the visual feature extraction network of the Swin Transformer, and also employs multi-sensor data fusion technology (visual images, infrared thermal imaging, and LiDAR point cloud data).

[0102] Three-level recognition architecture: Primary anomaly detection: Based on a lightweight CNN, used for rapid identification of abnormal targets. Intermediate precise localization: Combines an attention mechanism to improve target localization accuracy. Advanced target classification: Achieves feature-level fusion of multi-modal data and spatiotemporally aligned multi-modal data association through multi-sensor data fusion.

[0103] Figure 2 This is a structural diagram of a YOLOv7-SwinT model provided in an embodiment of this application. Figure 2 As shown, specifically, the improved YOLOv7-SwinT target recognition model includes:

[0104] The YOLOv7 object detection framework includes: a backbone network constructed from depthwise separable convolutions, an adaptive anchor size selection module, and a multi-scale feature fusion pyramid structure.

[0105] In this embodiment, the YOLOv7 framework is mainly divided into three parts: the backbone network, constructed using depthwise separable convolutions to improve computational efficiency; the adaptive anchor size selection module, which optimizes anchor box matching and improves detection accuracy; and the multi-scale feature fusion pyramid structure, which enhances the model's ability to detect targets at different scales.

[0106] The Swin Transformer feature extraction network includes a hierarchical window attention mechanism, a cross-window connection module, a local perception enhancement module, and a multi-sensor data fusion layer. It achieves feature-level fusion of visual images, infrared thermal imaging, and LiDAR point cloud data, feature weight allocation based on the attention mechanism, and spatiotemporally aligned multimodal data association.

[0107] In the Swin Transformer feature extraction network, a hierarchical window attention mechanism is used to improve feature extraction capabilities. A cross-window connection module enhances feature fusion performance. A local perception enhancement module improves the model's ability to perceive local details. A multi-sensor data fusion layer achieves feature-level fusion of visual images, infrared thermal imaging, and LiDAR point cloud data, allocating feature weights based on an attention mechanism to achieve spatiotemporally aligned multimodal data association.

[0108] The improved YOLOv7-SwinT target recognition model in step S1 adopts a three-level recognition architecture, including primary anomaly detection based on lightweight CNN, intermediate precise localization combined with attention mechanism, and advanced target classification based on multi-sensor data fusion.

[0109] Specifically, the three-level recognition architecture includes primary anomaly detection based on lightweight CNN, intermediate precise localization combined with attention mechanism, and high-level object classification based on multi-sensor data fusion, comprising:

[0110] The primary anomaly detection module uses a lightweight CNN network to quickly extract features from the input image and achieves preliminary screening of anomaly regions through convolutional layers with low computational complexity.

[0111] The intermediate-level precise localization module, built on the Swing Transformer architecture, includes: a multi-head self-attention unit for capturing long-distance spatial dependencies; a cross-window feature interaction unit for global context modeling through a shifting window mechanism; and a dynamic receptive field adjustment unit for adaptively adjusting the attention range according to the target scale.

[0112] The advanced target classification module integrates a multi-source sensor data fusion mechanism, specifically including: a visible light and infrared feature alignment unit that uses a cross-modal feature mapping network; a spatiotemporal context fusion unit that integrates temporal sensor data through 3D convolution; and a multi-task decision layer that synchronously outputs target category, confidence level, and spatial location information.

[0113] A cascaded feature transfer path enables feature reuse and enhancement among three-level modules through residual connections and a feature pyramid structure.

[0114] The architecture employs a progressive recognition strategy to achieve high-precision detection and classification of multi-scale targets in UAV inspection scenarios under conditions of limited computing resources.

[0115] The backbone network constructed from depthwise separable convolutions includes:

[0116] A feature extraction network is used to extract multi-scale features from an input image, which includes at least one variable attention module to enhance the correlation between distant pixels and improve the feature extraction capability of small targets;

[0117] A feature fusion network is used to integrate the multi-scale features output by the feature extraction network, wherein at least one deformable convolutional module is included to expand the receptive field of the convolutional kernel and reduce the number of model parameters;

[0118] The detection head is used to detect targets based on the fused features and outputs the target's category, location, and confidence level.

[0119] The loss calculation module is used to calculate the model's loss function during training, including: localization loss, which calculates the matching degree between the predicted bounding box and the ground truth bounding box based on EIoU; target confidence loss, which measures the accuracy of the model's prediction of whether the target exists; and category prediction loss, which optimizes the classification result using label-smoothed cross-entropy loss.

[0120] The system employs Mosaic data augmentation, transfer learning, and fine-tuning strategies during the training phase to improve the model's detection accuracy for small targets in drone patrol scenarios.

[0121] The loss function of YOLOv7 is a multi-task loss function, consisting of three parts: localization loss L. CIoU Target confidence loss L obj And category prediction loss L cls ,

[0122] YOLOv7 uses an improved IoU loss function to calculate the localization bounding box, obtaining the regression error, and the localization loss L. CIoU The formula is as follows:

[0123]

[0124] IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box, ρ 2 (b,b gt ) is the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box, l is the diagonal length of the minimum bounding rectangle, v is the aspect ratio consistency measure, and α is the weight coefficient.

[0125] Target confidence loss L obj The binary cross-entropy loss is used to measure whether the predicted bounding box contains the target, as shown in the following formula:

[0126]

[0127] S 2 B is the number of grid cells, and B is the number of anchor frames per grid cell. This is an indicator function; it is 1 if anchor frame j is responsible for detecting the target, and 0 otherwise. It is the confidence score of the prediction.

[0128] Category Prediction Loss L cls To measure the accuracy of the predicted category, a binary cross-entropy loss is used, as shown in the following formula:

[0129]

[0130] in: These are real category labels. is the predicted class probability, classes is the total number of classes, and c represents the number of classes.

[0131] The total loss function L of YOLOv7 total It is the weighted sum of the three loss functions mentioned above:

[0132] L total =λ coord L CIoU +λ obj L obj +λ cls L cls ,

[0133] Where, λ coord , λ obj , λ cls These are hyperparameters used to balance the weights of different tasks;

[0134] When the YOLOv7 framework includes an auxiliary head, its total loss function is L. total l :

[0135] L total l =L total +0.25L aux ,

[0136] Among them, L aux The loss function for the auxiliary head.

[0137] The adaptive anchor point size selection module includes:

[0138] The K-means clustering unit is used to perform cluster analysis on the width and height dimensions of all ground truth bounding boxes in the training dataset to generate an initial set of anchor point dimensions; the K-means clustering unit generates an initial set of anchor point dimensions by performing cluster analysis on the width and height dimensions of all ground truth bounding boxes in the training dataset.

[0139] The Intersection over Union (IoU) optimization unit uses an improved IoU metric to calculate the matching degree between the predicted bounding box and the ground truth bounding box, and dynamically adjusts the distribution of anchor point sizes.

[0140] The scale-sensitive filtering unit assigns a set of anchor points of a specific scale to each feature level based on the receptive field range of different feature layers in the object detection task.

[0141] The gradient backpropagation adaptation unit incorporates anchor size parameters into the network training process and optimizes the distribution of anchor size through end-to-end learning.

[0142] The gradient backpropagation adaptation unit automatically generates anchor point sizes that adapt to different detection scales by analyzing the geometric characteristics of the target dataset, thereby improving the model's detection accuracy for multi-scale targets.

[0143] The multi-scale feature fusion pyramid structure includes:

[0144] The bottom-up feature extraction pathway is used to extract multi-scale features of the input image layer by layer through a convolutional neural network to generate feature maps of different resolutions.

[0145] The top-down feature fusion pathway is used to fuse high-level semantic features with low-level high-resolution features step by step through upsampling operations. The top-down feature fusion pathway fuses high-level semantic features with low-level high-resolution features step by step through upsampling operations.

[0146] The lateral connection module uses a 1×1 convolution kernel to adjust the number of feature channels, enabling cross-scale connections between feature maps at different levels. The lateral connection module adjusts the number of feature channels by using a 1×1 convolution kernel to achieve cross-scale connections between feature maps at different levels.

[0147] The feature enhancement unit introduces an attention mechanism during the feature fusion process to dynamically adjust the weight distribution of features at different scales.

[0148] The multi-level prediction output layer sets up detection heads on the fused feature maps at each scale to achieve simultaneous detection of targets of different sizes, including small, medium, and large.

[0149] The multi-scale feature fusion pyramid structure achieves complementary fusion of deep and shallow features through a bidirectional feature transfer path, thereby improving the model's detection performance for multi-scale targets.

[0150] This method achieves efficient and accurate target recognition by improving the YOLOv7-SwinT model, and combines adaptive path planning and intelligent task allocation algorithms to realize collaborative inspection by multiple UAVs. It is innovative in target recognition, path planning, task allocation, and energy management, and is suitable for UAV inspection tasks in complex environments.

[0151] S2: Based on the target information identified in step S1, and combined with the drone's current battery level, environmental conditions, and task priority, a dynamic inspection path is generated through an adaptive path planning algorithm.

[0152] The adaptive path planning algorithm combines target information with environmental conditions: based on the identified target information, combined with the drone's current battery level, environmental conditions (wind speed, rainfall intensity, visibility), and task priority, a dynamic inspection path is generated. A multi-constraint path planning model is constructed, where the input parameters can be the drone's current battery status, environmental condition parameters (wind speed, rainfall intensity, visibility), task priority matrix, and geofencing constraints. The output can be the optimized inspection path, dynamically updated to the drone's flight control system.

[0153] Specifically, S2: Based on the target information identified in step S1, and combined with the UAV's current battery level, environmental conditions, and task priority, a dynamic inspection path is generated through an adaptive path planning algorithm, including:

[0154] S2.1: Based on the target information identified in step S1, establish an inspection target priority evaluation model. This evaluation model comprehensively considers: target anomaly score (a confidence score output by a deep learning model); target type weight (pre-defined priority coefficients for different target types); and target timeliness parameter (a time sensitivity factor calculated based on the target's rate of change). The evaluation model comprehensively considers the target anomaly score, target type weight, and target timeliness parameter. The target anomaly score is based on the confidence score output by the deep learning model; the target type weight is a pre-defined priority coefficient for different target types; and the target timeliness parameter is a time sensitivity factor calculated based on the target's rate of change.

[0155] S2.2: Construct a multi-constraint path planning model. Input parameters include:

[0156] The input parameters include the drone's current battery status and estimated flight time; environmental condition parameters, including wind speed, rainfall intensity, and visibility; a task priority matrix, calculated in step S2.1; and geofencing constraints. These parameters collectively determine the feasibility and optimization direction of the inspection path.

[0157] S2.3: An improved genetic algorithm is used to solve for the optimal path, specifically including:

[0158] a) Initialize the population: Generate N feasible paths as initial solutions;

[0159] b) Fitness calculation: Construct a multi-objective fitness function based on path length, task coverage, and energy balance;

[0160] c) Selection operation: A combination of elite retention strategy and roulette wheel selection is adopted;

[0161] d) Crossover mutation: Design targeted mutation operators for inspection tasks;

[0162] e) Dynamic adjustment: Real-time monitoring of environmental changes triggers path replanning.

[0163] S2.4: Output the optimized inspection path and dynamically update it to the UAV cruise system. The optimized inspection path is dynamically updated to the UAV flight control system to ensure the UAV can execute tasks in real time.

[0164] S3: Based on the inspection path generated in step S2, inspection tasks are assigned to multiple drones using an intelligent task allocation algorithm, and the task execution status is monitored in real time.

[0165] Specifically, multi-drone task allocation: Based on the generated inspection path, tasks are assigned to multiple drones through an intelligent task allocation algorithm, and the task execution status is monitored in real time.

[0166] The selection of a multi-UAV task allocation algorithm is crucial for ensuring efficient task execution. Challenges faced by task allocation algorithms include computational complexity, task coupling, problem size, time constraints, and heterogeneity. To address these challenges, various algorithms can be chosen, such as auction-based algorithms, ant colony optimization (ACO), and particle swarm optimization (PSO).

[0167] In practical applications, task allocation typically involves the following steps: Task submission and observation: Follower UAVs submit observation information to the Leader UAV, which generates a task list and auctions tasks based on their execution costs. Task allocation and path planning: Based on the task allocation results, Follower UAVs determine the distance to the target. If the distance is less than a set value, missiles are launched directly; otherwise, Model Predictive Control (MPC) algorithms are used to plan the attack path. Dynamic adjustment and replanning: During task execution, if the UAV malfunctions or environmental conditions change, the system needs to adjust the task allocation and path planning in real time.

[0168] During mission execution, the system needs to monitor the drone's status and mission progress in real time. After each drone completes its inspection, it uploads the inspection data to the central server, generating raw inspection data. Subsequently, based on the raw inspection data, a data integration algorithm is used to process and integrate the inspection data from multiple drones, generating integrated inspection data. In addition, the system also generates inspection reports using a report generation method and uses a report review algorithm to ensure the accuracy and completeness of the report content.

[0169] The design of intelligent task allocation algorithms needs to consider multiple factors, including drone availability, endurance, task type, and geographical location. Dynamic task allocation can be achieved through market mechanisms, and conflict resolution can be achieved through consensus mechanisms. Furthermore, combining algorithms based on improved self-attention mechanisms and adaptive task planning can enable efficient task allocation in dynamic environments.

[0170] In multi-drone collaborative operations, task allocation and path planning need to consider the cooperation and communication between drones. After completing task allocation, the drones fly to the maintenance area, visit the assigned tasks in sequence, and autonomously return to their nests after completing the tasks. Collaborative operations among multiple drones are achieved through information exchange and task synchronization between them.

[0171] S4: During the inspection process, sensor data is processed in real time through edge computing devices, and key data is transmitted to the control center.

[0172] Specifically, edge computing and data transmission are used. Edge computing devices are used to process sensor data in real time during the inspection process and transmit key data to the control center.

[0173] In S4, edge computing devices are used to process sensor data in real time during inspections and transmit critical data to the control center. This process has several key characteristics: 1. Real-time data processing: Edge computing devices can perform preliminary processing of sensor data locally, thereby reducing the time for data transmission to the cloud and lowering latency. For example, in drone inspections, edge computing can reduce the time for data transmission to the cloud, thus improving data processing efficiency. 2. Data preprocessing and filtering: Edge computing devices preprocess and filter sensor data, uploading only critical information to the cloud, thereby reducing network bandwidth consumption and improving data transmission efficiency. 3. Local decision-making and response: In some cases, edge computing devices can make rapid decisions based on local data. Edge computing nodes can analyze device sensor data in real time, and once an anomaly is detected, a local alarm is immediately triggered, and the anomaly information is quickly pushed to the terminal devices of safety management personnel. 4. Data transmission to the control center: The critical data processed by the edge computing devices is transmitted to the control center for further analysis and processing. In remote substation inspections, sensor data is first preprocessed by terminal devices, and then the pre-processed data is transmitted to the data processing server within the substation for further processing and integration. 5. Improved System Performance: Edge computing can significantly improve the real-time performance and accuracy of the system. In intelligent operation and maintenance of power systems, the combined application of edge computing and sensor technology enables the data collected by sensors to be initially processed and analyzed locally, thereby reducing the requirement for data transmission to a remote center, significantly reducing latency, and enhancing the real-time performance of data processing.

[0174] The application of edge computing in the inspection process not only improves the efficiency of data processing, but also enhances the real-time performance and accuracy of the system, providing strong support for the efficient execution of inspection tasks.

[0175] S5: Based on the real-time status and mission progress of the drone, the intelligent energy management system optimizes the scheduling of charging base stations and the energy distribution of the drone.

[0176] Optimize charging scheduling and energy allocation, specifically by optimizing charging base station scheduling and drone energy allocation through an intelligent energy management system based on the real-time status and mission progress of the drone.

[0177] In S5, the intelligent energy management system optimizes the scheduling of charging stations and the energy allocation of drones by monitoring the drone's status and mission progress in real time. This process involves several key technologies and methods, as follows: 1. Real-time monitoring and data analysis: The intelligent energy management system monitors key parameters of the drone in real time, such as battery voltage, current, and temperature, through sensors and flight computers (e.g., Pixhawk Px4 and Raspberry Pi). This data provides precise support for energy management, ensuring the system can respond quickly based on the drone's real-time status. 2. Dynamic energy allocation: The system dynamically allocates energy to different mission components based on the drone's mission priority and current energy status. When the drone is performing an inspection mission, the system prioritizes the energy supply to critical sensors and communication modules while reducing the energy consumption of unnecessary equipment. Furthermore, the system can further reduce the drone's energy consumption by adjusting parameters such as flight altitude and speed. 3. Charging station scheduling: When the drone's battery level falls below a preset threshold (e.g., 20%), the system automatically triggers the charging process. The drone uses GPS navigation to locate the nearest charging station and uses image processing technology to identify available charging units. Subsequently, the drone autonomously lands at the charging station, completes charging, and continues its mission. This autonomous charging mechanism not only reduces the need for manual intervention but also improves mission continuity and efficiency. 4. Optimization Algorithms and Models: To achieve efficient energy management, the system employs various optimization algorithms, including a rule-based energy management system (EMS), an optimization-based energy management system (EMS), and an intelligent-based energy management system (EMS). For example, the optimization-based EMS can utilize dynamic programming and model predictive control to optimize the UAV's flight path and charging strategy, thereby minimizing energy consumption and extending flight time. 5. Intelligent Scheduling and Prediction: The intelligent energy management system also combines artificial intelligence and big data technologies to perform real-time analysis and prediction of the UAV's mission requirements, flight environment, and energy status. In this way, the system can plan the UAV's flight path and charging strategy in advance, ensuring sufficient energy support during critical mission phases. 6. Energy Recovery and Reuse: To further improve energy utilization efficiency, the system has also developed an intelligent energy recovery mechanism. For example, surplus energy generated by the UAV during flight can be used wirelessly to power other devices or charge the battery, thereby reducing energy waste.

[0178] The intelligent energy management system achieves efficient management of drone charging base station scheduling and energy distribution through various means such as real-time monitoring, dynamic allocation, optimization algorithms, and intelligent scheduling, thereby significantly improving the drone's endurance and mission execution efficiency.

[0179] Example 2

[0180] like Figure 3 As shown, this application provides a drone cruise system based on a deep learning artificial intelligence image recognition algorithm, which is applied to the drone cruise method based on a deep learning artificial intelligence image recognition algorithm as described in Embodiment 1, including: an image acquisition module 11, a path planning module 12, a task allocation module 13, a data processing module 14, and an energy management module 15.

[0181] It is understood that in this embodiment, the image acquisition module 11 is used to acquire real-time image data of the inspection area by a drone equipped with an improved YOLOv7-SwinT target recognition model. The improved YOLOv7-SwinT target recognition model integrates the single-stage target detection architecture of YOLOv7 and the visual feature extraction network of Swin Transformer, and adopts multi-sensor data fusion technology.

[0182] It is understood that in this embodiment, the path planning module 12 is used to generate a dynamic inspection path based on the target information identified in step S1, combined with the current battery level of the UAV, environmental conditions and task priority, through an adaptive path planning algorithm.

[0183] It is understood that in this embodiment, the task allocation module 13 is used to allocate inspection tasks to multiple drones according to the inspection path generated in step S2 through an intelligent task allocation algorithm, and to monitor the task execution status in real time.

[0184] It is understood that in this embodiment, the data processing module 14 is used to process sensor data in real time through edge computing devices during the inspection process and transmit key data to the control center.

[0185] It is understood that in this embodiment, the energy management module 15 is used to optimize the scheduling of charging base stations and the energy allocation of the drone based on the real-time status and mission progress of the drone through the intelligent energy management system; wherein, the improved YOLOv7-SwinT target recognition model in step S1 adopts a three-level recognition architecture, including primary anomaly detection based on lightweight CNN, intermediate precise localization combined with attention mechanism, and advanced target classification based on multi-sensor data fusion.

[0186] This invention achieves breakthroughs in detection accuracy, response speed, and energy efficiency through innovative algorithm architecture and system design. It is particularly suitable for application scenarios with stringent requirements for real-time performance and reliability, such as power line inspection, border monitoring, and disaster relief, and has significant technical advantages and application value.

[0187] Figure 4This is an electronic device provided in one embodiment of this application. For example... Figure 4 As shown, the electronic device includes at least the following components: processor 101 and memory 100, communication interface 103, and bus 102.

[0188] In this embodiment of the application, memory 100 is used to store executable instructions of processor 101, which, when configured to execute instructions, implements the method as described in the first aspect.

[0189] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0190] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these devices is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0191] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0192] It should be noted that the term "computer" as used here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, "computer-readable recording media" refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard drives built into a computer.

[0193] Furthermore, a "computer-readable recording medium" can include: a medium that dynamically stores a program for a short period of time, such as a communication line used when transmitting a program via a network such as the Internet or a communication line such as a telephone line; or a medium that stores a program for a fixed period of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining with programs already recorded in the computer.

[0194] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (device group) composed of multiple devices. Each device constituting the device group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a device group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0195] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for drone patrol based on deep learning artificial intelligence image recognition algorithm, characterized in that, The method includes: S1: Real-time image data of the inspection area is collected by a drone equipped with an improved YOLOv7-SwinT target recognition model. The improved YOLOv7-SwinT target recognition model integrates the single-stage target detection architecture of YOLOv7 and the visual feature extraction network of SwinTransformer, and adopts multi-sensor data fusion technology. The improved YOLOv7-SwinT target recognition model specifically includes: The YOLOv7 object detection framework includes: a backbone network constructed from depthwise separable convolutions, an adaptive anchor size selection module, and a multi-scale feature fusion pyramid structure. The Swin Transformer feature extraction network includes a hierarchical window attention mechanism, a cross-window connection module, a local perception enhancement module, and a multi-sensor data fusion layer. It achieves feature-level fusion of visual images, infrared thermal imaging, and lidar point cloud data, feature weight allocation based on the attention mechanism, and spatiotemporally aligned multimodal data association. S2: Based on the target information identified in step S1, and combined with the drone's current battery level, environmental conditions, and task priority, a dynamic inspection path is generated using an adaptive path planning algorithm, including: S2.1: Based on the target information identified in step S1, establish an inspection target priority evaluation model. The evaluation model comprehensively considers: target anomaly score, based on the confidence score output by the deep learning model; target type weight, a pre-set priority coefficient for different types of targets; and target timeliness parameter, a time sensitivity factor calculated based on the target change rate. S2.2: Construct a multi-constraint path planning model. Input parameters include: The drone's current battery status and estimated flight time; environmental condition parameters, including wind speed, rainfall intensity, and visibility; task priority matrix, calculated in step S2.1; geofencing constraints; S2.3: An improved genetic algorithm is used to solve for the optimal path, specifically including: a) Initialize the population: Generate N feasible paths as initial solutions; b) Fitness calculation: Construct a multi-objective fitness function based on path length, task coverage, and energy balance. c) Selection operation: A combination of elite retention strategy and roulette wheel selection is adopted; d) Crossover mutation: Design targeted mutation operators for inspection tasks; e) Dynamic adjustment: Real-time monitoring of environmental changes triggers path replanning; S2.4: Output the optimized inspection path and dynamically update it to the UAV patrol system; S3: Based on the inspection path generated in step S2, assign inspection tasks to multiple drones using an intelligent task allocation algorithm, and monitor the task execution status in real time. S4: During the inspection process, sensor data is processed in real time through edge computing devices, and key data is transmitted to the control center. S5: Based on the real-time status and mission progress of the UAV, optimize the scheduling of charging base stations and the energy allocation of the UAV through an intelligent energy management system; wherein, the improved YOLOv7-SwinT target recognition model in step S1 adopts a three-level recognition architecture, including primary anomaly detection based on lightweight CNN, intermediate precise localization combined with attention mechanism, and high-level target classification based on multi-sensor data fusion. The three-level recognition architecture, including primary anomaly detection based on lightweight CNN, intermediate precise localization combined with attention mechanism, and high-level target classification based on multi-sensor data fusion, includes: The primary anomaly detection module uses a lightweight CNN network to quickly extract features from the input image and achieves preliminary screening of anomaly regions through convolutional layers with low computational complexity. The intermediate-level precise localization module, built on the Swing Transformer architecture, includes: a multi-head self-attention unit for capturing long-distance spatial dependencies; a cross-window feature interaction unit for global context modeling through a shifting window mechanism; and a dynamic receptive field adjustment unit for adaptively adjusting the attention range according to the target scale. The advanced target classification module integrates a multi-source sensor data fusion mechanism, specifically including: a visible light and infrared feature alignment unit that uses a cross-modal feature mapping network; a spatiotemporal context fusion unit that integrates temporal sensor data through 3D convolution; and a multi-task decision layer that synchronously outputs target category, confidence level, and spatial location information. A cascaded feature transfer path enables feature reuse and enhancement among three-level modules through residual connections and a feature pyramid structure. The architecture employs a progressive recognition strategy to achieve high-precision detection and classification of multi-scale targets in UAV inspection scenarios under conditions of limited computing resources.

2. The drone cruise method based on a deep learning artificial intelligence image recognition algorithm according to claim 1, characterized in that, The backbone network constructed from depthwise separable convolutions includes: A feature extraction network is used to extract multi-scale features from an input image, which includes at least one variable attention module to enhance the correlation between distant pixels and improve the feature extraction capability of small targets; A feature fusion network is used to integrate the multi-scale features output by the feature extraction network, wherein at least one deformable convolutional module is included to expand the receptive field of the convolutional kernel and reduce the number of model parameters; The detection head is used to detect targets based on the fused features and outputs the target's category, location, and confidence level. The loss calculation module is used to calculate the model's loss function during training, including: localization loss, which calculates the matching degree between the predicted bounding box and the ground truth bounding box based on EIoU; target confidence loss, which measures the accuracy of the model's prediction of whether the target exists; and category prediction loss, which optimizes the classification result using label-smoothed cross-entropy loss. The system employs Mosaic data augmentation, transfer learning, and fine-tuning strategies during the training phase to improve the model's detection accuracy for small targets in drone patrol scenarios.

3. The drone cruise method based on a deep learning artificial intelligence image recognition algorithm according to claim 2, characterized in that, The loss function of YOLOv7 is a multi-task loss function, consisting of three parts: localization loss. Target confidence loss and category prediction loss , YOLOv7 uses an improved IoU loss function to calculate the localization bounding box, obtaining the regression error and localization loss. The formula is as follows: ， ， ， It is the intersection-union ratio (IUU) between the predicted bounding boxes and the ground truth bounding boxes. It is the center point of the prediction box. Center point of the real frame Euclidean distance, It is the length of the diagonal of the smallest bounding rectangle. It is the aspect ratio. Consistency measure; These are weighting coefficients; Target confidence loss The binary cross-entropy loss is used to measure whether the predicted bounding box contains the target, as shown in the following formula: ， It is the number of grid cells. It is a grid index. It is the number of anchor frames per grid. It is an indicator function, if the anchor box The value is 1 if the target is being detected, and 0 otherwise. It is the confidence score of the prediction; Category Prediction Loss To measure the accuracy of the predicted category, a binary cross-entropy loss is used, as shown in the following formula: ， in: These are real category labels. It is the predicted class probability. It is the total number of categories. Indicates category; YOLOv7 total loss function It is the weighted sum of the three loss functions mentioned above: ， in, These are hyperparameters used to balance the weights of different tasks; When the YOLOv7 framework includes an auxiliary header, its total loss function is: : ， in, The loss function for the auxiliary head.

4. The drone cruise method based on a deep learning artificial intelligence image recognition algorithm according to claim 1, characterized in that, The adaptive anchor point size selection module includes: K-means clustering units are used to perform cluster analysis on the width and height dimensions of all ground truth bounding boxes in the training dataset to generate an initial set of anchor point dimensions. The Intersection over Union (IoU) optimization unit uses an improved IoU metric to calculate the matching degree between the predicted bounding box and the ground truth bounding box, and dynamically adjusts the distribution of anchor point sizes. The scale-sensitive filtering unit assigns a set of anchor points of a specific scale to each feature level based on the receptive field range of different feature layers in the target detection task. The gradient backpropagation adaptation unit incorporates anchor size parameters into the network training process, optimizing the distribution of anchor size through end-to-end learning. The gradient backpropagation adaptation unit automatically generates anchor point sizes that adapt to different detection scales by analyzing the geometric characteristics of the target dataset, thereby improving the model's detection accuracy for multi-scale targets.

5. The drone cruise method based on a deep learning artificial intelligence image recognition algorithm according to claim 1, characterized in that, The multi-scale feature fusion pyramid structure includes: The bottom-up feature extraction pathway is used to extract multi-scale features of the input image layer by layer through a convolutional neural network to generate feature maps of different resolutions. A top-down feature fusion pathway is used to fuse high-level semantic features with low-level high-resolution features step by step through upsampling operations. The horizontal connection module uses a 1×1 convolution kernel to adjust the number of feature channels, enabling cross-scale connections between feature maps at different levels; The feature enhancement unit introduces an attention mechanism during the feature fusion process to dynamically adjust the weight distribution of features at different scales; A multi-level prediction output layer sets up detection heads on the fused feature maps at various scales to achieve simultaneous detection of targets of different sizes, including small, medium, and large. The multi-scale feature fusion pyramid structure achieves complementary fusion of deep and shallow features through a bidirectional feature transfer path, thereby improving the model's detection performance for multi-scale targets.

6. A drone patrol system based on a deep learning artificial intelligence image recognition algorithm, applied to the drone patrol method based on a deep learning artificial intelligence image recognition algorithm as described in any one of claims 1 to 5, characterized in that, The system includes: The image acquisition module is used to collect real-time image data of the inspection area by a UAV equipped with an improved YOLOv7-SwinT target recognition model. The improved YOLOv7-SwinT target recognition model integrates the single-stage target detection architecture of YOLOv7 and the visual feature extraction network of Swin Transformer, and adopts multi-sensor data fusion technology. The path planning module is used to generate a dynamic inspection path based on the target information identified in step S1, combined with the drone's current battery level, environmental conditions, and task priority, through an adaptive path planning algorithm. The task allocation module is used to allocate inspection tasks to multiple drones based on the inspection path generated in step S2 using an intelligent task allocation algorithm, and to monitor the task execution status in real time. The data processing module is used to process sensor data in real time through edge computing devices during the inspection process and transmit key data to the control center. The energy management module is used to optimize the scheduling of charging base stations and the energy allocation of drones through an intelligent energy management system based on the real-time status and mission progress of the drones. The improved YOLOv7-SwinT target recognition model in step S1 adopts a three-level recognition architecture, including primary anomaly detection based on lightweight CNN, intermediate precise localization combined with attention mechanism, and advanced target classification based on multi-sensor data fusion.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that instructs the device to perform the drone cruise method based on a deep learning artificial intelligence image recognition algorithm as described in any one of claims 1 to 5.

Images