A real-time target detection and recognition system for unmanned aerial vehicle images

Abstract

This invention discloses a real-time target detection and recognition system and method for UAV imagery. The system adopts an edge-cloud collaborative processing architecture, including an airborne processing unit and a cloud-based collaborative server. The airborne unit is equipped with an adaptive lightweight detection network to perform preliminary detection on real-time imagery and output results with uncertainty scores. An intelligent collaborative gateway filters high-uncertainty targets based on dynamic collaborative thresholds, extracts relevant image regions, and uploads them to the cloud. The cloud server deploys a high-precision model for verification and sends the results back to the airborne unit for fusion, outputting the final detection result. This system achieves a balance between detection accuracy and real-time performance under limited airborne resources through dynamic adjustment of the airborne network structure, bandwidth-based dynamic threshold decision-making, and online model optimization mechanisms based on cloud feedback, thereby improving the robustness and efficiency of UAV target detection in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of target detection and recognition technology, specifically to a real-time target detection and recognition system for UAV imagery. Background Technology

[0002] With the rapid development of drone technology, drone platforms, due to their flexibility and efficiency, have been widely applied in many fields such as border patrol, emergency rescue, traffic monitoring, and power line inspection. In railway infrastructure inspection, the structural health of bridges and tunnels directly affects traffic safety. Traditional manual inspection methods are inefficient, risky, and have limited coverage, necessitating an automated and intelligent defect identification system. One of the core tasks of these applications is the automated target detection and identification of massive amounts of image data collected by drones in real time. However, when this task is implemented in engineering, it faces the dual challenges of constraints from the airborne platform and complex visual scenarios. Existing technical solutions struggle to achieve a good balance between real-time performance, accuracy, and resource consumption.

[0003] First, the onboard computing devices of drones are severely limited in terms of weight, size, and power consumption, resulting in limited available computing resources. To meet the demands of real-time processing, the current mainstream solution is to deploy highly lightweight target detection models on the drone. While these models offer fast inference speeds, their simple structure and few parameters result in weak feature extraction and discrimination capabilities. In real-world scenarios such as small-scale targets, targets rotating arbitrarily, complex dynamic backgrounds, and drastic lighting changes, detection accuracy and recall rates drop significantly, leading to a large number of false negatives and false positives.

[0004] Secondly, to improve accuracy, another approach is to transmit the image data back to a cloud server with powerful computing capabilities via a wireless communication link for processing. However, this approach introduces new bottlenecks: on the one hand, continuous transmission of high-resolution video streams consumes a large amount of bandwidth, is costly, and is difficult to implement in areas with poor signal coverage; on the other hand, data transmission and cloud processing introduce considerable network latency, making the system unable to meet the needs of security, inspection, and other scenarios with extremely high real-time requirements, thus negating the rapid response advantage of drones.

[0005] In recent years, the "cloud-edge collaboration" architecture, combining edge computing and cloud computing, has been introduced into this field, aiming to combine the advantages of both. However, in some existing collaboration solutions, the collaboration strategies are often relatively simple and static. For example, some solutions fixate on placing some detection layers at the edge and others in the cloud, failing to adaptively adjust according to dynamic changes in image content and network status; others upload all low-confidence detection results, resulting in significant bandwidth pressure in complex scenarios, and cloud computing power being consumed by a large number of "simple" or "explicit" negative samples, leading to low efficiency. Furthermore, existing solutions generally lack an effective mechanism to quantify the "confidence" or "uncertainty" of the onboard model regarding its own detection results, and to use this as a precise basis for intelligent collaboration decisions, resulting in an imprecise and inefficient collaboration process.

[0006] Therefore, a new real-time target detection and recognition system for UAV imagery has become a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] This application provides a real-time target detection and recognition system for UAV imagery, including: An airborne processing unit, mounted on an unmanned aerial vehicle (UAV) platform; and, The cloud-based collaborative server is connected to the airborne processing unit via a wireless communication link, forming an edge-cloud collaborative processing architecture; The airborne processing unit includes: The image acquisition module is used to acquire raw image sequences through an optical sensor, perform real-time preprocessing on the raw image sequences, and output a standardized real-time video stream. An adaptive lightweight detection network, connected to the image acquisition module, is used to perform target detection on each frame of the real-time video stream and generate a preliminary detection result set containing the predicted bounding box, class label and associated uncertainty score of each target. The intelligent collaborative gateway is connected to the output of the adaptive lightweight detection network and the wireless communication link, respectively. The intelligent collaborative gateway is configured to perform collaborative decision-making and control, including: Receive the preliminary detection result set; For each target in the result set, its associated uncertainty score is compared with a cooperative threshold. When the uncertainty score of a target reaches or exceeds the collaborative threshold, it is determined that the target needs to be reviewed in the cloud. Then, the image region containing the target is extracted from the corresponding original frame image to form an image data packet to be reviewed. The image data packet to be reviewed is uploaded to the cloud collaborative server via the wireless communication link. The cloud-based collaborative server is equipped with a high-precision detection model with a higher parameter quantity and calculation accuracy than the adaptive lightweight detection network in the airborne processing unit. This model is used to perform high-confidence target verification on the received image data packets to be verified, generate verification results, and send the verification results back to the airborne processing unit via the wireless communication link. The airborne processing unit is also used to fuse the received verification result with the locally generated preliminary detection result set and output the final detection result.

[0009] Optionally, the structure of the adaptive lightweight detection network is dynamically adjusted in response to the real-time computing resource status or preset task mode of the UAV platform; The adaptive lightweight detection network is a lightweight multi-scale rotating target detection network, which includes a backbone network, a neck network and a detection head connected in sequence. The backbone network is used to enhance the preservation of local image details by reorganizing spatial information to the channel dimension during downsampling, thereby improving the feature extraction effect for small targets. The neck network is used to adaptively weighted fuse feature maps at different levels. The adaptive weighting includes block-level weighting of the feature map as a whole and channel-level weighting of the feature channels to optimize the representation of multi-scale features. The detection head is used to enhance the feature map's ability to perceive the scene before prediction by fusing global contextual attention with local large-field information extracted by dilated convolution.

[0010] Optionally, the downsampling and information recombination process in the backbone network is as follows: the input feature map is divided into grids and extracted into pixels in the spatial dimension by a preset step size, and the extracted pixel sub-images are interleaved and spliced ​​in the channel dimension. Finally, the number of channels is adjusted and the output is processed through a convolutional layer. The adaptive weighted fusion process in the neck network is as follows: for the input multi-layer feature map, firstly, the overall importance of each feature map is weighted by a learnable block-level weight tensor, and then the weighted fused feature is refined and adjusted by a learnable channel-level weight tensor before output. The feature perception enhancement process in the detection head is as follows: global average pooling and fully connected operations are performed in parallel on the input feature map to obtain the global context vector. At the same time, dilated convolution with adjustable dilation rate is performed to expand the receptive field. Finally, the obtained global information, local large field of view information and the original features are weighted and fused together before output.

[0011] Optionally, the dynamic adjustment specifically involves the onboard processing unit monitoring the remaining battery percentage and central processing unit load rate of the UAV platform in real time. When the remaining battery power is lower than the first threshold or the processor load rate is higher than the second threshold, the adaptive lightweight detection network is controlled to switch to a simplified inference path. The simplified inference path is achieved by skipping at least one feature processing module in the backbone network or uniformly reducing the number of channels in all feature maps in the neck network. When the remaining battery power and processor load both return to normal, control the network to switch back to the full inference path.

[0012] Optionally, in the intelligent collaborative gateway, the collaborative threshold is a dynamic threshold; The intelligent collaborative gateway is used to monitor the available uplink bandwidth of the wireless communication link in real time; The intelligent collaborative gateway is also configured to dynamically adjust the collaborative threshold based on the available uplink bandwidth value through a built-in mapping relationship. Specifically, when the available bandwidth is higher than a preset high bandwidth threshold, the collaborative threshold is lowered to increase the cloud review ratio; when the available bandwidth is lower than a preset low bandwidth threshold, the collaborative threshold is raised to reduce the amount of data uploaded.

[0013] Optionally, when the intelligent collaborative gateway captures an image region, it uses the center point of the predicted bounding box of the target as a reference and expands the bounding box size in all directions by a preset ratio of pixel distance to form an image data packet containing the complete target and necessary background context. Before uploading the image data packet to be reviewed, the intelligent collaborative gateway selects different compression strategies based on its associated uncertainty score: when the uncertainty score is in the first interval, lossy compression is used; when the uncertainty score is in the higher second interval, lossless compression is used.

[0014] Optionally, the cloud-based collaborative server is further configured to: store the review results and corresponding preliminary detection results issued in each iteration, and analyze the differences between the two in terms of target category and location; When statistics reveal a systematic deviation between the initial detection results of the adaptive lightweight detection network and the cloud-based verification results for a specific category or scenario, corresponding incremental update parameters for the model are generated. The incremental update parameters of the model are sent to the airborne processing unit, which then applies them to the adaptive lightweight detection network to achieve online optimization.

[0015] Optionally, the uncertainty score is calculated based on the category prediction probability distribution entropy, bounding box coordinate prediction variance, or a weighted sum of the two output by the adaptive lightweight detection network. The category prediction probability distribution entropy is obtained by calculating the Shannon entropy of the category probability distribution output by the model; the bounding box coordinate prediction variance is obtained by calculating the variance of the bounding box coordinates predicted by the model multiple times for the same target.

[0016] This application also provides a method for real-time target detection and recognition of UAV imagery, including: Airborne preliminary detection and screening steps: Based on the images collected in real time by the UAV, a preliminary detection result containing uncertainty scores is generated, and the targets and their associated image data that require high-precision verification are intelligently screened according to the uncertainty scores; Cloud-based high-precision verification steps: Receive and process the associated image data, and generate verification results through a high-precision detection model; The result fusion and output steps are as follows: The verification results are fused with the preliminary detection results to output the final detection results.

[0017] Optionally, the airborne preliminary detection and screening steps specifically include: An adaptive lightweight detection network is used to process real-time video stream frames to generate a preliminary detection result set containing target bounding boxes, categories, and uncertainty scores. The uncertainty score is compared with the collaborative threshold to filter out targets that need to be reviewed in the cloud, wherein the collaborative threshold is a dynamic threshold that is dynamically adjusted according to the real-time available bandwidth of the wireless communication link; Extract the image region containing the target from the corresponding frame image to form a data packet to be reviewed; Upload the data packet to be reviewed to the cloud collaboration server; The cloud-based high-precision verification step also includes: generating model update parameters based on historical verification data and sending them to the airborne processing unit for online optimization of the adaptive lightweight detection network.

[0018] Optionally, the system is suitable for the automatic identification of surface defects in railway bridges and tunnels, the targets including one or more of cracks, spalling, corrosion, and water seepage.

[0019] Optionally, the cloud-based collaborative server is also used to generate structural health trend reports and maintenance recommendations based on the results of previous defect identifications.

[0020] The beneficial effects of this application are as follows: 1. This invention introduces "uncertainty score" as the core decision-making criterion, enabling the intelligent collaborative gateway to accurately filter out "difficult" targets with low confidence and high error rates in the airborne network, and upload only partial image data of such targets to the cloud for high-precision verification. Compared with traditional fixed-division or full-data-upload schemes, this content-based dynamic filtering mechanism significantly reduces bandwidth consumption and cloud load while ensuring that key and difficult points receive the highest quality assessment. Thus, it achieves a balance between low latency and high detection accuracy end-to-end within limited airborne computing power and communication resources. 2. The adaptive lightweight detection network of this invention can dynamically adjust its inference path based on the remaining battery power and real-time processor load of the UAV platform. It achieves optimal performance when resources are abundant and ensures basic real-time performance when resources are scarce, thus flexibly responding to the computing power requirements of different mission stages and flight states. Simultaneously, its integrated multi-scale rotating target detection structure is specifically optimized for the challenge of detecting small and rotating targets from the UAV's perspective, enabling the airborne model itself to maintain superior basic detection capabilities while remaining lightweight. 3. This invention not only completes the collaboration of a single detection task, but also establishes a mechanism for continuous optimization. By analyzing the differences between historical review results and the initial airborne results, the cloud server can identify systematic deviations of the airborne network in specific categories or scenarios, and generate incremental update parameters accordingly, which are then sent to the airborne terminal. This allows the lightweight detection network deployed on the UAV to continuously absorb knowledge from the cloud during use, achieving targeted incremental learning and performance iteration, and enabling the entire system's recognition capability to continuously improve over time and with the accumulation of task scenarios. Attached Figure Description

[0021] Figure 1 This application provides a real-time target detection and recognition system for UAV imagery; Figure 2 The method for real-time target detection and recognition of UAV imagery provided in this application. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please refer to Figure 1 As shown, this application provides a real-time target detection and recognition system for UAV imagery, comprising: The airborne processing unit refers to the collection of computing devices mounted on the UAV platform, which is responsible for performing tasks such as image acquisition, preliminary target detection, and collaborative decision-making on the UAV. The onboard processing unit contains an image acquisition module. This module acquires raw image sequences generated during the UAV's flight by integrating optical sensors, such as visible light or infrared cameras. The acquired raw image sequences are then sent to a preprocessing unit for real-time preprocessing operations, such as noise reduction, color correction, distortion correction, or frame rate adjustment, to ensure image quality and format consistency. The preprocessed image data is output as a standardized real-time video stream, providing input for subsequent target detection tasks.

[0024] The real-time video stream output from the image acquisition module is transmitted to an adaptive lightweight detection network. This network, deployed on an onboard processing unit, primarily performs object detection on each frame of the video stream. During detection, the network identifies potential objects in the image and generates a series of information for each identified object, including a predicted bounding box, a label indicating the object's category, and an associated uncertainty score. This uncertainty score quantifies the network's confidence in its detection results. For example, the network can employ a lightweight convolutional neural network-based model, such as MobileNet or ShuffleNet, whose architecture is designed for fast inference with limited computational resources.

[0025] The initial detection result set generated by the adaptive lightweight detection network is transmitted to the intelligent cooperative gateway. This gateway is configured to receive these results and, for each target in the result set, compare its associated uncertainty score with a preset cooperative threshold. The cooperative threshold can be a fixed value, such as 0.7 or 0.8, used to define the model's confidence level in the detection results. When the uncertainty score of a target is higher than or equal to the cooperative threshold, it indicates that the airborne network's detection result for that target has high uncertainty and may require a higher level of verification.

[0026] When the intelligent collaborative gateway detects that the uncertainty score of a target reaches or exceeds the collaborative threshold through comparison, the gateway determines that the target needs to undergo cloud-based collaborative review. Subsequently, the gateway extracts an image region containing the target from the corresponding original frame image. For example, a rectangular region can be directly extracted based on the coordinates of the predicted bounding box. The extracted image region is packaged into an image data packet to be reviewed, ready for uploading.

[0027] After generating the image data packet to be reviewed, the intelligent collaborative gateway uploads the packet to the cloud-based collaborative server via a wireless communication link. This wireless communication link can be a cellular network (such as 4G / 5G), satellite communication, or a dedicated radio link. The uploaded data packet contains only the image of the target area to be reviewed, rather than the entire original video frame, thus effectively saving transmission bandwidth.

[0028] The cloud-based collaborative server can employ large-scale deep learning models such as ResNet and Transformer. After receiving the image data packet to be verified, the cloud-based collaborative server uses this high-precision detection model to perform high-confidence verification of the targets in the data packet. The verification process aims to correct potential issues such as missed detections, false detections, or inaccurate positioning in the airborne network, and to generate more accurate verification results.

[0029] After the high-precision detection model generates the verification results, the cloud-based collaborative server sends these results back to the onboard processing unit via a wireless communication link. This sending process ensures that the high-precision processing results from the cloud can be promptly transmitted back to the edge to guide subsequent decision-making or display.

[0030] After receiving the verification results from the cloud, the airborne processing unit merges them with the locally generated preliminary detection result set. The fusion strategy may include: for targets verified by the cloud, directly using the cloud results; for targets that did not trigger cloud verification, retaining the local preliminary detection results. After fusion, the system outputs the final detection result. This final detection result combines the real-time performance of the airborne unit with the high precision of the cloud, thereby improving the overall system performance.

[0031] It is worth noting that this application proposes an adaptive lightweight detection network whose structure dynamically adjusts in response to the real-time computing resource status or preset task mode of the UAV platform. This adaptive lightweight detection network is a lightweight multi-scale rotating target detection network, comprising a backbone network, a neck network, and a detection head connected in sequence. The backbone network enhances the preservation of local image details during downsampling by reorganizing spatial information to the channel dimension, thereby improving the feature extraction effect for small targets. The neck network performs adaptive weighted fusion of feature maps at different levels, including block-level weighting of the overall feature map and channel-level weighting of the feature channels, to optimize the representation of multi-scale features. The detection head enhances the scene perception capability of the feature map by fusing global contextual attention with local large-field-of-view information extracted by dilated convolution before prediction.

[0032] The adaptive lightweight detection network can dynamically adjust its internal structure based on the real-time computing resource status of the UAV platform or the preset task mode. This adjustment can include, but is not limited to: selecting different pre-trained model variants, adjusting the depth or width of network layers, enabling or disabling specific feature processing modules, and changing quantization precision. Through dynamic adjustment, the system can flexibly balance detection speed and accuracy when resources are limited or task requirements change, ensuring optimal performance under various operating conditions.

[0033] The backbone network is responsible for extracting multi-level features from the original image. During downsampling, the backbone network reorganizes spatial information into the channel dimension, concentrating pixel information that was originally scattered across the spatial dimension into the channel dimension, thereby effectively preserving local details of the image while reducing the spatial resolution of the feature map.

[0034] The neck network, located between the backbone network and the detection head, fuses feature maps from different levels output by the backbone network to integrate multi-scale information. This network employs an adaptive weighted fusion mechanism, specifically including block-level weighting of the overall feature map and channel-level weighting of the feature channels. Block-level weighting adjusts the overall feature representation by learning the importance of different feature map blocks; channel-level weighting further refines the representation by learning and adjusting the importance of each channel.

[0035] The detection head is responsible for generating the final target prediction based on the fused feature map. Before prediction, the detection head enhances the feature map's ability to perceive the scene by fusing global contextual attention with local large-view information extracted by dilated convolution. The global contextual attention mechanism can capture the overall semantic information of the image, helping the model understand the macroscopic environment in which the target is located; dilated convolution (or dilated convolution) can effectively expand the receptive field and obtain broader local region information without increasing the number of parameters and computational cost. By weightedly fusing global and local large-view information, the detection head can more comprehensively understand the image content, thereby improving the accuracy and robustness of target detection.

[0036] This application further proposes the following specific steps for the downsampling and information recombination process in the backbone network: the input feature map is divided into grids and pixels are extracted in the spatial dimension according to a preset stride, and the extracted pixel sub-images are interleaved and stitched together in the channel dimension. Finally, the number of channels is adjusted and the output is processed through a convolutional layer. The adaptive weighted fusion process in the neck network is as follows: for the input multi-layer feature map, the overall importance of each feature map is first weighted by a learnable block-level weight tensor, and then the weighted fused features are refined and adjusted by a learnable channel-level weight tensor before output. The feature perception enhancement process in the detection head is as follows: global average pooling and fully connected operations are performed in parallel on the input feature map to obtain the global context vector. At the same time, dilated convolution with adjustable dilation rate is performed to expand the receptive field. Finally, the obtained global information, local large-view information and the original features are weighted and fused before output.

[0037] Specifically, the downsampling and information recombination process in the backbone network aims to efficiently reduce the resolution of the feature map while preserving the maximum amount of local detail information in the image, which is particularly crucial for small target detection commonly seen in UAV imagery. The process first divides the input feature map into a grid with a preset stride in the spatial dimension and extracts pixels from each grid. For example, a 2x2 grid can be used, extracting all pixels or pixels at specific locations from each grid. Then, these pixel sub-images extracted from different grid locations or stride locations are staggered and stitched together in the channel dimension. This staggered stitching "folds" spatially adjacent or related pixel information to the channel dimension, thereby increasing the number of channels in the feature map while downsampling, allowing each channel to encode richer spatial context information. Finally, a convolutional layer processes the stitched feature map to adjust the number of channels to an appropriate dimension and further learns a more discriminative feature representation.

[0038] The adaptive weighted fusion process in the neck network focuses on optimizing the representation of multi-scale features to address the challenge of large target scale variations in UAV imagery. This process first weights the overall importance of multiple input feature maps at different levels by introducing a learnable block-level weight tensor. This block-level weight tensor can be a vector, with each element corresponding to an input feature map, automatically learned through network training to reflect the relative importance of each feature map in the overall fusion. After block-level weighting and initial fusion, to further refine the feature representation, this application introduces a learnable channel-level weight tensor. This channel-level weight tensor independently weights each channel of the fused features, allowing the network to perform more refined feature selection and enhancement along the channel dimension, for example, through the Squeeze-and-Excitation (SE) module or its variants. This enables the fused feature map to better highlight key information and suppress redundant information.

[0039] The feature-aware enhancement process in the detection head aims to improve its ability to understand scenes, thereby outputting more accurate prediction results. This process simultaneously acquires global contextual information and local large-view information through parallel processing. On one hand, by performing global average pooling on the input feature map, the spatial information of each channel is compressed into a single value representing the global feature of that channel. Subsequently, through a fully connected operation, these global features are transformed into a compact global context vector, which encodes the overall semantic information of the image. On the other hand, to capture local large-view information, this application employs dilated convolution with adjustable dilation rate. Dilated convolution can expand the receptive field of the convolutional kernel without increasing parameters or losing resolution. By adjusting the dilation rate, local environmental information can be captured at different scales. For example, dilated convolutions with different dilation rates can be processed in parallel, and the results are then fused. Finally, the acquired global information, local large-view information, and the original features are weighted and fused before output. This fusion of multi-source information enables the detection head to utilize both global semantics and local details for more accurate predictions, thereby enhancing its ability to perceive complex scenes.

[0040] In some embodiments described above in this application, an adaptive lightweight detection network is proposed for the airborne processing unit. This network can dynamically adjust in response to the real-time computing resource status or preset task mode of the UAV platform to optimize the representation of multi-scale features and enhance the scene perception capability. However, in practical applications, the computing resources of the UAV platform (such as power and processor load) are limited and dynamically changing. Without fine-grained management, continuous high-intensity computing may lead to resource exhaustion, affecting the stability and continuous operation capability of the system, and even causing task interruption.

[0041] The dynamic adjustment specifically involves the following: the onboard processing unit monitors the remaining battery percentage and central processing unit load rate of the UAV platform in real time; when the remaining battery is below a first threshold or the processor load rate is above a second threshold, the adaptive lightweight detection network is controlled to switch to a simplified inference path, which is achieved by skipping at least one feature processing module in the backbone network or uniformly reducing the number of channels in all feature maps in the neck network; when both the remaining battery and processor load return to normal, the network is controlled to switch back to the full inference path.

[0042] To achieve the aforementioned dynamic adjustments, the onboard processing unit is configured to monitor the critical resource status of the UAV platform in real time. Specifically, for monitoring the remaining battery percentage, the onboard processing unit can periodically obtain the current battery level information and calculate the percentage of remaining battery capacity by communicating with the UAV flight control system or battery management system (BMS). For monitoring the CPU load rate, the onboard processing unit can utilize the performance monitoring interfaces or APIs provided by its operating system, such as reading the ` / proc / loadavg` file in a Linux system or obtaining CPU utilization and average load through the `top` command, thereby assessing the processor's workload in real time. This monitoring data can be updated at a preset frequency, such as once per second, to ensure timely decision-making.

[0043] When the drone platform's resource status is detected to reach preset critical conditions—that is, when the remaining battery power is below a preset first threshold (e.g., 20%) or the CPU load rate is above a preset second threshold (e.g., 80%)—the onboard processing unit immediately controls the adaptive lightweight detection network to switch to a simplified inference path. This switching aims to reduce resource consumption by lowering the network's computational complexity. The simplified inference path can be implemented by skipping at least one feature processing module in the backbone network. For example, the backbone network typically consists of multiple stacked residual blocks or convolutional layers. In simplified mode, computationally intensive modules can be selectively disabled or skipped, especially those that contribute little to feature refinement or can be sacrificed under resource constraints, thereby directly reducing the computational load of forward propagation.

[0044] Conversely, when the remaining battery power and CPU load of the UAV platform return to normal levels—that is, when the remaining battery power is above the first threshold and the CPU load is below the second threshold—the onboard processing unit controls the adaptive lightweight detection network to switch back to the full inference path. This ensures that, when resources permit, the system can recover to its optimal detection performance to provide the highest accuracy in target detection results.

[0045] In a basic UAV imagery real-time target detection and recognition system, the onboard processing unit uploads target data packets whose uncertainty scores in the preliminary detection results reach or exceed a collaborative threshold to a cloud-based collaborative server for high-precision verification via an intelligent collaborative gateway. However, in practical applications, the available uplink bandwidth of the wireless communication link between the UAV platform and the cloud-based collaborative server may fluctuate drastically due to factors such as environment, distance, or network congestion. If the collaborative threshold is set to a fixed value, when bandwidth is ample, the system may not fully utilize the high-precision processing capabilities of the cloud, resulting in some targets whose accuracy could have been improved not being verified. Conversely, when bandwidth is limited, maintaining a low collaborative threshold may lead to a large amount of data being uploaded, causing network congestion, data transmission delays, or even data loss, severely impacting the system's real-time performance and overall performance.

[0046] In the intelligent collaborative gateway, the collaborative threshold is a dynamic threshold. This threshold is no longer a preset fixed value, but a parameter that can be adjusted in real time according to changes in system operating status or external environment. Its dynamism allows the system to adapt more flexibly to different working scenarios and resource conditions, thereby optimizing the efficiency and effectiveness of edge-cloud collaborative processing. The dynamic threshold can be a continuously changing value, or it can switch between several preset discrete values.

[0047] To enable dynamic adjustment of the coordination threshold, the intelligent coordination gateway monitors the available uplink bandwidth of the wireless communication link in real time. The intelligent coordination gateway continuously acquires the actual available uplink bandwidth information of the current wireless communication link through its built-in communication module or an interface with a communication module. This can be achieved by periodically sending probe packets, analyzing network-level traffic statistics, or receiving bandwidth reports provided by the underlying communication protocol stack. Real-time monitoring ensures that the system can promptly perceive changes in the network environment, providing accurate data for subsequent dynamic adjustments.

[0048] The intelligent collaborative gateway is further configured to dynamically adjust the collaborative threshold based on the available uplink bandwidth value through a built-in mapping relationship. The intelligent collaborative gateway internally presets or learns a mapping relationship that defines the correspondence or functional relationship between the available uplink bandwidth value and the collaborative threshold. When a change in available uplink bandwidth is detected, the intelligent collaborative gateway calculates and updates the current collaborative threshold based on this mapping relationship. This mapping relationship can be a simple piecewise function or a complex nonlinear relationship trained based on a machine learning model, aiming to balance the accuracy requirements of cloud verification with the load capacity of network transmission.

[0049] When the available bandwidth exceeds a preset high bandwidth threshold, the collaboration threshold is lowered to increase the proportion of cloud-based verification. This means that when the intelligent collaboration gateway detects that the available uplink bandwidth of the wireless communication link exceeds a preset high bandwidth threshold, it indicates that network resources are sufficient, and the system is capable of transmitting more data to the cloud for verification. Therefore, the intelligent collaboration gateway will lower the collaboration threshold. Lowering the collaboration threshold means that more targets with relatively low uncertainty scores but still above the new threshold will be identified as requiring cloud verification, thereby increasing the number of targets that need to be uploaded to the cloud for high-precision processing, and improving the overall accuracy and reliability of detection.

[0050] Conversely, when the available bandwidth is below a preset low bandwidth threshold, the coordination threshold is increased to reduce data upload volume. This means that when the intelligent coordination gateway detects that the available uplink bandwidth of the wireless communication link is below a preset low bandwidth threshold, it indicates that network resources are strained, and priority should be given to ensuring the transmission of critical data. Therefore, the intelligent coordination gateway will increase the coordination threshold. An increased coordination threshold means that only targets with very high uncertainty scores—that is, targets where the onboard processing unit has extremely low confidence in its detection results—will be determined to require cloud verification, thus significantly reducing the amount of data uploaded to the cloud. This helps alleviate network congestion, ensures timely transmission of core data under limited bandwidth, and maintains the system's real-time responsiveness.

[0051] In response, this application further proposes that when the intelligent collaborative gateway captures an image region, it does not simply capture the precise area defined by the predicted bounding box. Instead, it uses the center point of the target's predicted bounding box as a reference and expands the bounding box outwards by a preset ratio of pixel distance to form an image data packet containing the complete target and necessary background context. Specifically, the intelligent collaborative gateway first obtains the width and height of the target's predicted bounding box and calculates the pixel distance to be expanded in each of the four directions (up, down, left, and right) according to a preset expansion ratio (e.g., 0.1 or 0.2). Then, using the center point of the original bounding box as the new center, it generates a larger expanded bounding box and captures the image region covered by this expanded bounding box from the original frame image, thereby forming the image data packet to be verified. This expansion mechanism ensures that the image data packet to be verified not only contains the target itself but also the necessary background context information around the target. This is crucial for accurate verification by high-precision detection models in the cloud, especially in complex scenarios where the target is partially occluded or requires environmental information for assistance in judgment.

[0052] Furthermore, to optimize the bandwidth utilization efficiency of the wireless communication link, the intelligent collaborative gateway selects different compression strategies based on the associated uncertainty score before uploading the image data packet to be verified. The uncertainty score reflects the confidence level of the adaptive lightweight detection network within the airborne processing unit regarding the target detection result. When the uncertainty score is in the first interval, it indicates that the airborne detection network has a relatively high confidence level in the target. In this case, the intelligent collaborative gateway will use lossy compression technology to process the image data packet, such as using JPEG format and setting appropriate compression quality parameters, to significantly reduce the data volume with acceptable image quality loss. When the uncertainty score is in the higher second interval, it indicates that the airborne detection network has a lower confidence level in the target, and the necessity of cloud verification is higher. To ensure that the high-precision detection model in the cloud can obtain the most original and complete information for accurate verification, the intelligent collaborative gateway will use lossless compression technology to process the image data packet, such as using PNG or lossless WebP format, to ensure that the image data does not lose any detailed information during transmission. The first and second intervals can be preset with specific numerical ranges according to actual application requirements. For example, the first interval can be set to an uncertainty score between 0.2 and 0.5, while the second interval can be set to an uncertainty score between 0.5 and 0.8.

[0053] In real-time target detection and recognition systems based on UAV imagery, the onboard processing unit's adaptive lightweight detection network handles initial detection, while a cloud-based collaborative server performs high-precision verification. However, the onboard network may have inherent detection biases or limitations when facing specific target categories or complex scenarios. If these biases are not learned and corrected in a timely manner, the onboard network may continuously generate inaccurate initial detection results, leading to redundant consumption of cloud verification resources and affecting the overall system's detection efficiency and accuracy.

[0054] The cloud-based collaborative server is also configured to store the previously issued review results and corresponding preliminary detection results, and analyze the differences between the two in terms of target category and location. When statistics show that there is a systematic deviation between the preliminary detection results of the adaptive lightweight detection network and the cloud review results for a specific category or scenario, corresponding model incremental update parameters are generated. The model incremental update parameters are then sent to the airborne processing unit, which receives them and applies them to the adaptive lightweight detection network to achieve online optimization.

[0055] The cloud-based collaborative server is configured to continuously record each verification result sent from the cloud to the onboard processing unit, as well as the preliminary detection results generated by the onboard processing unit's adaptive lightweight detection network corresponding to these verification results. This storage mechanism can establish a historical dataset for subsequent analysis. By comparing this historical data, the cloud-based collaborative server can identify specific differences in the onboard network's target category recognition (e.g., misidentifying a class A target as a class B target) and target localization (e.g., a systematic offset between the predicted bounding box and the actual target location).

[0056] Based on continuous analysis of historical review results and preliminary detection results, when the cloud-based collaborative server, through statistical methods such as setting error thresholds, trend analysis, or machine learning algorithms, discovers a persistent, non-random, systematic deviation between the airborne adaptive lightweight detection network's preliminary detection results and the cloud-based high-precision review results when processing a specific target category (e.g., small vehicles, specific equipment models), or a specific scenario (e.g., low light, severe occlusion, dense targets), the cloud-based collaborative server triggers the generation of model update parameters. These incremental model update parameters aim to correct these systematic deviations in the airborne network.

[0057] This application further proposes the calculation of the uncertainty score, which is based on the category prediction probability distribution entropy, bounding box coordinate prediction variance, or a weighted sum of the two output by the adaptive lightweight detection network. Specifically, the category prediction probability distribution entropy is obtained by calculating the Shannon entropy of the category probability distribution output by the model; the bounding box coordinate prediction variance is obtained by calculating the variance of the bounding box coordinates predicted multiple times for the same target by the model.

[0058] Uncertainty score is a key metric for measuring the confidence level of an adaptive lightweight detection network in its initial detection results. Unlike a simple confidence score, it aims to more comprehensively capture the ambiguity or uncertainty that may exist in the model during the prediction process. The accuracy of this score directly affects the collaborative decision-making efficiency of the intelligent collaborative gateway and the reliability of the final detection results. Its calculation method can be based on various information from the model output to provide a richer confidence assessment.

[0059] Category prediction probability distribution entropy is a method used to quantify the uncertainty of adaptive lightweight detection networks in predicting the class of a target. When a model outputs a probability distribution for predicting the class of a target, the entropy value reflects the concentration of that probability distribution. If the probability distribution tends to be uniform, meaning the model gives similar probabilities for multiple classes, the entropy value is high, indicating greater uncertainty in the model's judgment of the target class; conversely, if the model gives extremely high probabilities for one class and extremely low probabilities for others, the entropy value is low, indicating greater certainty in the model's judgment of the target class. By calculating the Shannon entropy of the class probability distribution output by the model, the degree of hesitation of the model in classification tasks can be intuitively assessed.

[0060] Bounding box coordinate prediction variance is a method used to quantify the uncertainty of target localization in adaptive lightweight detection networks. It measures the dispersion of the model's coordinates when predicting the same target bounding box multiple times. A larger variance indicates that the model's predictions of target position and size are not stable enough or fluctuate significantly, i.e., high localization uncertainty; a smaller variance indicates that the model's localization predictions are more stable and accurate. To obtain this variance, various techniques can be used, such as introducing random perturbations during inference (e.g., test-time augmentation), using Monte Carlo Dropout or Bayesian deep learning methods, making multiple predictions of the same target, and then calculating the variance of these predictions.

[0061] By weighting and summing the category prediction probability distribution entropy and the bounding box coordinate prediction variance, a comprehensive uncertainty score can be obtained. This combination approach simultaneously considers the uncertainty of the model in both classification and localization dimensions, thus providing a more comprehensive and robust confidence assessment. The weighting coefficients can be adjusted according to the different emphases on classification accuracy and localization accuracy in the actual application scenario. For example, in scenarios with higher requirements for target category recognition, a greater weight can be given to the category prediction probability distribution entropy.

[0062] Please refer to Figure 2 As shown, this application proposes a real-time target detection and recognition method for UAV imagery, which includes an airborne preliminary detection and screening step, a cloud-based high-precision verification step, and a result fusion and output step.

[0063] The airborne preliminary detection and screening step aims to generate preliminary detection results containing uncertainty scores using images acquired in real time by the UAV, and intelligently screen out targets and their associated image data that require high-precision verification based on the uncertainty scores. Specifically, this step uses an adaptive lightweight detection network to process real-time video stream frames, generating a preliminary detection result set containing target bounding boxes, categories, and uncertainty scores. The adaptive lightweight detection network is a lightweight multi-scale rotating target detection network whose structure can be dynamically adjusted in response to the real-time computing resource status or preset task mode of the UAV platform. For example, when the remaining battery power is below a first threshold or the processor load rate is above a second threshold, the network can switch to a simplified inference path to adapt to limited airborne resources. The uncertainty score is used to quantify the confidence level of the preliminary detection results, and its calculation can be based on the category prediction probability distribution entropy, bounding box coordinate prediction variance, or a weighted sum of the two output by the adaptive lightweight detection network. Subsequently, the uncertainty score is compared with a collaborative threshold to screen out targets that require cloud verification. The collaboration threshold is not a fixed value, but a dynamic threshold that is dynamically adjusted based on the real-time available bandwidth of the wireless communication link. For example, when the available bandwidth is higher than a preset high bandwidth threshold, the collaboration threshold is lowered to increase the cloud review ratio; when the available bandwidth is lower than a preset low bandwidth threshold, the collaboration threshold is raised to reduce the amount of data uploaded. For the targets selected for cloud review, an image region containing the target is extracted from the corresponding frame image to form a data packet to be reviewed. When extracting the image region, the center point of the target's predicted bounding box is usually used as a reference, and the bounding box size is expanded outwards by a preset percentage of pixel distance to ensure that the complete target and necessary background context are included. Before uploading the data packet to be reviewed, different compression strategies are selected according to its associated uncertainty score. For example, when the uncertainty score is in the first interval, lossy compression is used; when the uncertainty score is in the higher second interval, lossless compression is used. Finally, the data packet to be reviewed is uploaded to the cloud collaboration server via the wireless communication link.

[0064] The cloud-based high-precision verification step aims to receive and process the associated image data, generating verification results through a high-precision detection model. The cloud-based collaborative server deploys a high-precision detection model with a higher parameter count and computational accuracy than the adaptive lightweight detection network within the onboard processing unit. This model can perform high-confidence target verification on the received image data packets to be verified, generating more accurate verification results. Furthermore, this step includes generating model update parameters based on historical verification data and sending them to the onboard processing unit for online optimization of the adaptive lightweight detection network. Specifically, the cloud-based collaborative server stores each previously sent verification result and the corresponding preliminary detection result, analyzing the differences between the two in target category and location. When statistics reveal a systematic deviation between the preliminary detection results of the adaptive lightweight detection network and the cloud-based verification results for a specific category or scene, corresponding incremental model update parameters are generated and sent to the onboard processing unit. The onboard processing unit receives these parameters and applies them to the adaptive lightweight detection network, thereby achieving online optimization.

[0065] The result fusion and output step aims to fuse the verification results with the preliminary detection results to output the final detection result. The airborne processing unit fuses the received cloud verification results with the locally generated preliminary detection result set to obtain the final, more reliable detection result.

[0066] This system is not only suitable for general target detection, but can also be specially configured for automatic identification of defects in railway bridges and tunnels.

[0067] In railway bridge and tunnel defect identification applications, the system's image acquisition module uses a high-definition optical sensor mounted on a drone or a structured light scanning device to acquire surface image sequences of bridge piers, beams, tunnel linings, and other components. The pre-trained adaptive lightweight detection network can identify various typical defect types, such as cracks, concrete spalling, steel corrosion, and water seepage. In the initial detection stage, the network not only outputs the defect's bounding box and category label but also an uncertainty score reflecting the severity of the defect or the detection confidence level.

[0068] Based on the uncertainty score and the current wireless bandwidth, the intelligent collaborative gateway dynamically decides whether to upload images of suspected defect areas to the cloud for high-precision verification. For example, for defects that are difficult to identify, such as fine cracks or early corrosion, the system can automatically trigger cloud verification, where a high-precision model in the cloud performs sub-pixel-level analysis to confirm the defect morphology and severity.

[0069] After completing defect verification, the cloud-based collaborative server can also combine historical inspection data to statistically analyze the defect development trend of the same bridge and tunnel structure and generate a structural health assessment report. Furthermore, the system supports an incremental model update mechanism to continuously optimize the airborne network's ability to identify new defects or those in specific environments, thereby improving the adaptability and accuracy of long-term inspections.

[0070] Through this system, railway operation and maintenance departments can achieve large-scale, high-frequency, and low-cost automatic inspection of bridges and tunnels.

[0071] In the automatic identification of defects in railway bridges and tunnels, the system is configured and operated as follows: (1) The image acquisition module acquires data for the surface of railway bridges and tunnels: The drone, equipped with a high-resolution optical sensor or infrared thermal imager, flies along a preset route to collect surface image sequences of structures such as bridge piers, beams, and tunnel linings. The preprocessing unit performs noise reduction, distortion correction, and contrast enhancement on the raw images, outputting a standardized video stream for subsequent processing by the detection network.

[0072] (2) An adaptive lightweight detection network performs preliminary identification of defect features: This pre-trained network is specifically designed to identify typical defect categories such as cracks, spalling, corrosion, and water seepage. For each image frame, the network outputs a predicted bounding box, category label, and uncertainty score for each defect. The uncertainty score reflects the network's confidence in its judgment of the current defect type; for example, it has a higher uncertainty score for fine cracks or faint corrosion.

[0073] (3) The intelligent collaborative gateway makes decisions based on the uncertainty of defect identification: The gateway receives the preliminary detection results and compares the uncertainty score of each defect with a dynamic collaborative threshold. This dynamic threshold is adjusted in real time based on the available uplink bandwidth of the current wireless communication link: in open areas with good signal, the threshold is lowered to increase the proportion of cloud-based verification; in signal-constrained areas such as tunnels, the threshold is raised to reduce the amount of data uploaded. For defects whose uncertainty reaches the threshold, the gateway extracts an image region containing the defect and its surrounding context from the original image to form a data packet to be verified.

[0074] (4) High-precision defect verification and analysis are performed using a cloud-based collaborative server: A high-precision detection model deployed in the cloud performs fine-grained identification and quantitative analysis on uploaded defect images, such as calculating crack width, estimating peeling area, and determining corrosion level. The verification results are then downloaded to the onboard unit and merged with the local preliminary results to output the final defect report.

[0075] (5) The system supports continuous learning and model optimization based on defect data: The cloud server stores historical defect identification data, analyzes the systematic deviations of the airborne network's identification results under specific structural, material, or lighting conditions, generates incremental update parameters, and sends them to the airborne network to achieve online model optimization and gradually improve the adaptability to new defects or complex scenarios.

[0076] The above description is merely a few embodiments of this application and is not intended to limit this application in any way. Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any changes or modifications made by those skilled in the art without departing from the scope of the technical solution of this application using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A real-time target detection and recognition system for UAV imagery, characterized in that, include: Airborne processing unit, mounted on the drone platform; as well as, The cloud-based collaborative server is connected to the airborne processing unit via a wireless communication link, forming an edge-cloud collaborative processing architecture; The airborne processing unit includes: The image acquisition module is used to acquire raw image sequences through an optical sensor, perform real-time preprocessing on the raw image sequences, and output a standardized real-time video stream. An adaptive lightweight detection network, connected to the image acquisition module, is used to perform target detection on each frame of the real-time video stream and generate a preliminary detection result set containing the predicted bounding box, class label and associated uncertainty score of each target. The intelligent collaborative gateway is connected to the output of the adaptive lightweight detection network and the wireless communication link, respectively. The intelligent collaborative gateway is configured to perform collaborative decision-making and control, including: Receive the preliminary detection result set; For each target in the result set, its associated uncertainty score is compared with a cooperative threshold. When the uncertainty score of a target reaches or exceeds the collaborative threshold, it is determined that the target needs to be reviewed in the cloud. Then, the image region containing the target is extracted from the corresponding original frame image to form an image data packet to be reviewed. The image data packet to be reviewed is uploaded to the cloud collaborative server via the wireless communication link. The cloud-based collaborative server is equipped with a high-precision detection model with a higher parameter quantity and calculation accuracy than the adaptive lightweight detection network in the airborne processing unit. This model is used to perform high-confidence target verification on the received image data packets to be verified, generate verification results, and send the verification results back to the airborne processing unit via the wireless communication link. The airborne processing unit is also used to fuse the received verification result with the locally generated preliminary detection result set and output the final detection result; The system is applicable to the automatic identification of surface defects in railway bridges and tunnels, including one or more of the following: cracks, spalling, corrosion, and water seepage.

2. The system according to claim 1, characterized in that, The structure of the adaptive lightweight detection network is dynamically adjusted in response to the real-time computing resource status or preset task mode of the UAV platform. The adaptive lightweight detection network is a lightweight multi-scale rotating target detection network, which includes a backbone network, a neck network and a detection head connected in sequence. The backbone network is used to enhance the preservation of local image details by reorganizing spatial information to the channel dimension during downsampling, thereby improving the feature extraction effect for small targets. The neck network is used to adaptively weighted fuse feature maps at different levels. The adaptive weighting includes block-level weighting of the feature map as a whole and channel-level weighting of the feature channels to optimize the representation of multi-scale features. The detection head is used to enhance the feature map's ability to perceive the scene before prediction by fusing global contextual attention with local large-field information extracted by dilated convolution.

3. The system according to claim 2, characterized in that, The downsampling and information recombination process in the backbone network is as follows: the input feature map is divided into grids and extracted pixels in the spatial dimension according to a preset step size, and the extracted pixel sub-images are interleaved and spliced ​​in the channel dimension. Finally, the number of channels is adjusted and the output is processed through a convolutional layer. The adaptive weighted fusion process in the neck network is as follows: for the input multi-layer feature map, firstly, the overall importance of each feature map is weighted by a learnable block-level weight tensor, and then the weighted fused feature is refined and adjusted by a learnable channel-level weight tensor before output. The feature perception enhancement process in the detection head is as follows: global average pooling and fully connected operations are performed in parallel on the input feature map to obtain the global context vector. At the same time, dilated convolution with adjustable dilation rate is performed to expand the receptive field. Finally, the obtained global information, local large field of view information and the original features are weighted and fused together before output.

4. The system according to claim 2, characterized in that, The dynamic adjustment specifically refers to the following: the onboard processing unit is used to monitor the remaining battery percentage and central processing unit load rate of the UAV platform in real time; When the remaining battery power is lower than the first threshold or the processor load rate is higher than the second threshold, the adaptive lightweight detection network is controlled to switch to a simplified inference path. The simplified inference path is achieved by skipping at least one feature processing module in the backbone network or uniformly reducing the number of channels in all feature maps in the neck network. When the remaining battery power and processor load both return to normal, control the network to switch back to the full inference path.

5. The system according to claim 1, characterized in that, In the intelligent collaborative gateway, the collaborative threshold is a dynamic threshold; The intelligent collaborative gateway is used to monitor the available uplink bandwidth of the wireless communication link in real time; The intelligent collaborative gateway is also configured to dynamically adjust the collaborative threshold based on the available uplink bandwidth value through a built-in mapping relationship. Specifically, when the available bandwidth is higher than a preset high bandwidth threshold, the collaborative threshold is lowered to increase the cloud review ratio; when the available bandwidth is lower than a preset low bandwidth threshold, the collaborative threshold is raised to reduce the amount of data uploaded.

6. The system according to claim 1, characterized in that, When the intelligent collaborative gateway captures an image region, it uses the center point of the predicted bounding box of the target as a reference and expands the bounding box size in all directions by a preset ratio of pixel distance to form an image data packet containing the complete target and necessary background context. Before uploading the image data packet to be reviewed, the intelligent collaborative gateway selects different compression strategies based on its associated uncertainty score: when the uncertainty score is in the first interval, lossy compression is used; when the uncertainty score is in the higher second interval, lossless compression is used.

7. The system according to claim 1, characterized in that, The cloud-based collaborative server is also configured to: store the review results and corresponding preliminary detection results issued in each iteration, and analyze the differences between the two in terms of target category and location; When statistics reveal a systematic deviation between the initial detection results of the adaptive lightweight detection network and the cloud-based verification results for a specific category or scenario, corresponding incremental update parameters for the model are generated. The incremental update parameters of the model are sent to the airborne processing unit, which then applies them to the adaptive lightweight detection network to achieve online optimization.

8. The system according to claim 1, characterized in that, The uncertainty score is calculated based on the category prediction probability distribution entropy, bounding box coordinate prediction variance, or a weighted sum of the two output by the adaptive lightweight detection network. The category prediction probability distribution entropy is obtained by calculating the Shannon entropy of the category probability distribution output by the model; the bounding box coordinate prediction variance is obtained by calculating the variance of the bounding box coordinates predicted by the model multiple times for the same target.

9. A real-time target detection and recognition method for UAV imagery based on the system described in any one of claims 1-8, characterized in that, include: Airborne preliminary detection and screening steps: Based on the images collected in real time by the UAV, a preliminary detection result containing uncertainty scores is generated, and the targets and their associated image data that require high-precision verification are intelligently screened according to the uncertainty scores; Cloud-based high-precision verification steps: Receive and process the associated image data, and generate verification results through a high-precision detection model; The result fusion and output steps are as follows: The verification results are fused with the preliminary detection results to output the final detection results.

10. The method according to claim 9, characterized in that, The airborne preliminary detection and screening steps specifically include: An adaptive lightweight detection network is used to process real-time video stream frames to generate a preliminary detection result set containing target bounding boxes, categories, and uncertainty scores. The uncertainty score is compared with the collaborative threshold to filter out targets that need to be reviewed in the cloud, wherein the collaborative threshold is a dynamic threshold that is dynamically adjusted according to the real-time available bandwidth of the wireless communication link. Extract the image region containing the target from the corresponding frame image to form a data packet to be reviewed; Upload the data packet to be reviewed to the cloud collaboration server; The cloud-based high-precision verification step also includes: generating model update parameters based on historical verification data and sending them to the airborne processing unit for online optimization of the adaptive lightweight detection network; The cloud-based collaborative server is also used to generate structural health trend reports and maintenance recommendations based on the results of previous defect identifications.

Images