Distributed intelligent identification processing system for unmanned aerial vehicle cluster collaborative aerial image

Abstract

The invention discloses a distributed intelligent identification processing system for unmanned aerial vehicle cluster collaborative aerial images, which belongs to the technical field of unmanned aerial vehicle image processing and specifically comprises the following steps: an unmanned aerial vehicle acquires aerial images according to a geographic division scheme and stores the aerial images in a local edge cache; splitting the aerial image data into double-layer tasks according to color and texture features, and uploading the double-layer tasks to a regional cache; each node obtains data from the three-level cache system, target identification is completed by using an intelligent collaborative identification algorithm after primary processing, and a result is stored in a global cache; the central node monitors the load weight of each node in real time, and allocates a new task by adopting a weighted minimum connection number algorithm; the center node aggregates the recognition results in the global cache according to grid IDs, eliminates splicing dislocation through a consistent Hash algorithm, generates an aerial image recognition report and outputs the aerial image recognition report; according to the invention, the problems of low processing efficiency and large cooperation error of a centralized architecture are effectively solved, and the efficiency and accuracy of unmanned aerial vehicle cluster aerial image processing are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of UAV image processing, and more specifically, it is a distributed intelligent recognition and processing system for collaborative aerial photography images of UAV swarms. Background Art

[0002] Traditional UAV aerial photography image processing usually adopts a centralized processing mode, that is, all UAVs transmit the collected image data back to the ground station or cloud server for unified analysis. However, this method has the following problems: it takes 30 minutes for a single server to process 20GB of images. When the task volume of the UAV swarm increases, a large amount of image data accumulates at a single node, resulting in a sharp decline in the system processing efficiency, and even system crashes, seriously affecting the timely completion of tasks; when multiple UAVs identify the same target, there is duplicate calculation, and misjudgment is prone to occur during the recognition process, affecting the accuracy of image recognition; in an electromagnetic interference environment, the data transmission error rate in the centralized architecture increases significantly, leading to problems such as misalignment of image stitching, and it cannot meet the requirements of high-quality image processing. Therefore, a new processing method is urgently needed to solve these problems. Summary of the Invention

[0003] Aiming at the deficiencies of the prior art, the present invention proposes a distributed intelligent recognition and processing system for collaborative aerial photography images of UAV swarms. The UAVs collect aerial photography images according to the geographical division plan and store them in the local edge cache, and then split the aerial photography image data into double-layer tasks according to color and texture features and upload them to the regional cache; each node obtains data from the three-level cache system, completes target recognition using the intelligent collaborative recognition algorithm after preliminary processing, and stores the results in the global cache; the central node monitors the load weights of each node in real time and uses the weighted least connection number algorithm to allocate new tasks; the central node aggregates the recognition results in the global cache according to the grid ID, eliminates stitching misalignment through the consistent hashing algorithm, and generates and outputs an aerial photography image recognition report; the present invention effectively solves the problems of low processing efficiency and large collaborative errors in the centralized architecture, and improves the efficiency and accuracy of UAV swarm aerial photography image processing.

[0004] To achieve the above object, the present invention provides the following technical solutions: A distributed intelligent recognition and processing system for collaborative aerial photography images of UAV swarms, including: a data acquisition module, a data cache module, a data recognition module, and a load balancing module; the data cache module includes a local edge cache unit, a regional cache unit, and a global cache unit; the load balancing module includes a load monitoring unit and a task allocation unit; Split the aerial photography image data collected by the data acquisition module in the local edge cache unit into double-layer tasks according to the split feature type, and upload the split subtasks and metadata to the regional cache unit; Each node in the UAV cluster obtains the aerial image data allocated to itself from the data cache module, and uses an improved intelligent collaborative recognition algorithm in the data recognition module to jointly perform target recognition on the aerial images, and stores the marked target recognition results in the global cache unit; the improved intelligent collaborative recognition algorithm is configured based on a deep convolutional neural network, an attention mechanism, and multi-scale feature fusion; The task allocation unit uses the weighted least connection number algorithm, combines the load weights of each node in the UAV cluster monitored in real time by the load monitoring unit, and allocates the newly arrived data processing tasks; After the new data processing task is completed, the central node aggregates all recognition results from the global cache according to the grid ID, forms an aerial image recognition report, and outputs the final result.

[0005] Specifically, the double-layer task includes: Split the color features of the aerial image data in the local edge cache unit, and allocate the hyperspectral data obtained after the color feature splitting to the nodes equipped with GPUs in the cluster; Split the texture features of the aerial image data in the local edge cache unit, and allocate the texture complex data obtained after the texture feature splitting to the nodes equipped with FPGAs; The metadata includes grid ID, feature type, and priority.

[0006] Specifically, the splitting of the color features of the aerial image data in the local edge cache unit and the allocation of the hyperspectral data obtained after the color feature splitting to the nodes equipped with GPUs in the cluster includes: A1: After the local edge cache unit receives the read instruction from the system, it extracts the aerial image data from the local edge cache and performs a preliminary check on the extracted aerial image data; If it is found that the extracted aerial image data has noise or color deviation, the Gaussian filtering algorithm and histogram equalization method are used to remove the noise and correct the color to obtain the preprocessed aerial image data; A2: Convert the preprocessed aerial image data from the original RGB color space to the CIELAB color space; A3: Split the CIELAB values in the CIELAB color space according to the preset color feature threshold, and separate the pixel data in the aerial image that meets the color features to obtain the split hyperspectral data; A4: The central node collects the information of the nodes equipped with GPUs in the cluster, calculates the comprehensive score of each node, and selects nodes according to the comprehensive score of each node; A5: Package the split hyperspectral data and transmit it to the nodes selected in A4 through the network.

[0007] Specifically, the method of splitting the texture features of the aerial image data in the local edge cache unit and allocating the texture complex data obtained after splitting according to the texture features to the nodes equipped with FPGAs includes: B1: After receiving the texture feature splitting instruction, the local edge cache unit reads the aerial image data from the local edge cache and performs preprocessing; B2: Calculate the spatial relationship between different gray levels in the aerial image through the gray-level co-occurrence matrix to obtain texture feature parameters; the texture feature parameters include contrast, correlation, energy, and homogeneity; B3: Evaluate the complexity of the texture through contrast calculation according to the obtained texture features; B4: Set a texture complexity threshold to divide the pixels in the aerial image into two categories: texture complex and texture simple; B5: According to the texture complexity evaluation result, split the aerial image data into texture complex data and texture simple data according to the preset texture complexity threshold, and separately extract the texture complex data to form a data subset to be allocated; B6: The central node collects information of the nodes equipped with FPGAs in the cluster; B7: According to the information of the nodes equipped with FPGAs, use the weighted round-robin algorithm to select FPGA nodes; B8: Package the data subset to be allocated and transmit it to the FPGA node selected in B7 through the network.

[0008] Specifically, each node in the UAV cluster obtains the aerial image data allocated to itself from the data cache module, and uses an improved intelligent collaborative recognition algorithm in the data recognition module to jointly perform target recognition on the aerial image, and stores the marked target recognition result in the global cache unit, including: C1: Each node in the UAV cluster locates in the distributed index system of the data cache module according to the parsed double-layer task allocation information; the specific process of the location includes: The node sends a query request to the index server, and the index server returns the physical address of the aerial image data in the cache module according to the task identifier and data features in the request; C2: The node obtains the aerial image data allocated to itself from the data cache module through the distributed file system according to the physical address returned by the index server and performs preprocessing; C3: Load the improved intelligent collaborative recognition algorithm model, and each node inputs the preprocessed aerial image into the improved intelligent collaborative recognition algorithm model for target recognition; the improved intelligent collaborative recognition algorithm model is configured based on a deep convolutional neural network, an attention mechanism, and multi-scale feature fusion; C4: After the target recognition is completed, each node marks the target recognition result; the marking is to draw a bounding box, annotate the target category and confidence on the original aerial image. C5: Each node encapsulates the marked target recognition result and stores the encapsulated target recognition result in the global cache at the storage location determined by the distributed index system of the double-layer task assignment information and the global cache through the network.

[0009] Specifically, the specific steps of C3 include: C3.1: Each node loads the improved intelligent collaborative recognition algorithm model in the local storage. C3.2: Each node inputs the preprocessed aerial image into the improved intelligent collaborative recognition algorithm model, performs convolution operations on the aerial image through the convolutional layer, and extracts multi-scale feature maps; in the convolution operation, the first 3 layers of the convolutional layer use separable convolution, and the last 2 layers of the convolutional layer use dilated convolution. C3.3: Perform weighted processing on the multi-scale feature maps through the attention mechanism to obtain weighted feature maps. C3.4: Input the weighted feature maps into the region proposal network, generate candidate regions containing the target through the sliding window, and perform classification and bounding box regression on the candidate regions to obtain the target recognition result; the target recognition result includes the target category, position coordinates, and confidence. C3.5: Each node encapsulates the recognition result into structured data, encrypts it, and sends it to 3 adjacent nodes within the communication radius. The adjacent nodes calculate the intersection over union of the received target recognition result and the local target recognition result of the adjacent nodes. If the intersection over union is greater than 0.5 and the confidence of the received target recognition result is greater than the confidence of the local target recognition result of the adjacent node, then replace the local target recognition result of the adjacent node with the received target recognition result. If the intersection over union is less than 0.3, then retain the received target recognition result as a new detection result. C3.6: Output the final target recognition result set.

[0010] Specifically, the construction process of the pre-trained improved intelligent collaborative recognition algorithm model in C3.1 includes: C3.11: Define the application scenario of the model and collect the data of the application scenario of the model. C3.12: Use the LabelImg annotation tool to accurately annotate the collected data of the application scenario of the model. C3.13: Load the model architecture based on the deep convolutional neural network and perform testing. C3.14: Introduce an attention mechanism into the well-tested deep convolutional neural network model architecture, and combine multi-scale feature fusion and separable convolution techniques to fuse feature maps of different layers; C3.15: Design the information interaction and communication methods between nodes, and formulate a collaboration strategy to form an improved intelligent collaborative recognition algorithm model; C3.16: Divide the preprocessed aerial image data into a training set, a validation set, and a test set, and select a cross-entropy loss function and a stochastic gradient descent optimizer according to the task type; C3.17: Use the training set to train the improved intelligent collaborative recognition algorithm model to obtain a well-trained improved intelligent collaborative recognition algorithm model.

[0011] Specifically, the specific steps of C3.3 include: C3.31: Organize the multi-scale feature maps obtained in C3.2 in the channel dimension, and perform global average pooling and global max pooling operations on each feature map to obtain global information feature vectors; The global average pooling operation is implemented by calculating the average value of all pixels in each channel; The global max pooling operation is implemented by extracting the maximum value of all pixels in each channel; The channel dimension is [batch_size, channels, height, width], where batch_size represents the number of samples processed at one time, channels represents the number of channels of the aerial image data, height represents the height of the image or feature map, and width represents the width of the aerial image; C3.32: Input the global information feature vectors into a shared multi-layer perceptron to obtain a channel attention weight matrix; C3.33: Perform channel average pooling and channel max pooling operations on the multi-scale feature maps in the channel dimension to generate a spatial attention weight matrix; The channel average pooling is implemented by calculating the average value of all channels at each position; The channel max pooling operation is implemented by calculating the maximum value of all channels at each position; C3.34: Multiply the channel attention weight matrix and the spatial attention weight matrix element by element to obtain an attention weight matrix; C3.35: Multiply the attention weight matrix and the multi-scale feature maps in C3.2 element by element to obtain weighted feature maps.

[0012] Specifically, the task allocation unit uses the weighted least connection number algorithm and combines the load weights of each node in the UAV cluster monitored in real time by the load monitoring unit to allocate the newly arrived data processing tasks, including: D1: Create an information record for each node in the UAV cluster; the information record includes the node identifier, the current connection number, and the load weight information; D2: Start the load monitoring unit to enable it to start monitoring the load conditions of each node in real time and update the load weights of each node; D3: After the task allocation unit receives a new data processing task, obtain the current connection number and load weight of each node from the load monitoring unit; D4: For each node, the task allocation unit calculates its priority using the weighted least connection number algorithm; the weighted least connection number algorithm is implemented by calculating the ratio of the current connection number of node i to the load weight of node i; D5: According to the calculated node priorities, select the node with the highest priority to process the new task, update the connection number of the selected node, and increment its connection number by 1; D6: The selected node starts to execute the task assigned to it. At the same time, the load monitoring unit continues to monitor the load conditions of each node in real time and dynamically updates the load weights of the nodes; D7: When the node completes the task, decrement its connection number by 1 and wait for the arrival of a new task.

[0013] Compared with the prior art, the beneficial effects of the present invention are: The present invention proposes a distributed intelligent recognition and processing system for collaborative aerial photography images of UAV clusters. In the data collection and processing stage, the geographical division and feature splitting strategy disperses the aerial photography tasks to each node, avoiding excessive processing pressure on a single node; the cooperation of local edge caching and regional caching realizes the efficient storage and distribution of data. The nodes equipped with GPUs and FPGAs process according to the image features, giving play to the hardware advantages and greatly shortening the processing time; the application of the intelligent collaborative recognition algorithm integrates the data of multiple nodes, effectively reducing the misjudgment rate of multi-aircraft repeated recognition, improving the accuracy of target recognition, and making the image recognition results more reliable. The present invention proposes a distributed intelligent recognition and processing system for collaborative aerial photography images of UAV clusters. The load balancing and result integration mechanism ensures the stability of the system and the processing quality. The central node dynamically allocates tasks using the weighted least connection number algorithm by monitoring the load weights of each node in real time, preventing single-point overload and ensuring the efficient and stable operation of the system; in the result processing link, the recognition results are aggregated according to the grid ID, and the consistent hashing algorithm is used to eliminate stitching misalignment, forming a complete and accurate aerial photography image recognition report, enhancing the adaptability of the system in complex environments, effectively solving the defects of the traditional centralized architecture, and providing a better solution for the processing of UAV cluster aerial photography images. Description of the Drawings

[0014] Figure 1 This is the principle flow chart of the distributed intelligent recognition and processing system for collaborative aerial photography images of the UAV cluster of the present invention; Figure 2 This is the architecture diagram of the distributed intelligent recognition and processing system for collaborative aerial photography images of the UAV cluster of the present invention; Figure 3 This is the flow chart for obtaining hyperspectral data of the distributed intelligent recognition and processing system for collaborative aerial photography images of the UAV cluster of the present invention; Figure 4 This is the target recognition flow chart of the distributed intelligent recognition and processing system for collaborative aerial photography images of the UAV cluster of the present invention. Detailed implementation manners

[0015] Embodiment 1 Please refer to Figure 1 and Figure 2 A kind of embodiment provided by the present invention: A distributed intelligent recognition and processing system for collaborative aerial photography images of a UAV cluster, including: A data acquisition module, a data cache module, a data recognition module, a load balancing module, and a report generation module; The data acquisition module collects aerial photography images of its respective grid areas through UAVs, and performs temporary storage and data splitting and uploading; The data cache module is used to build a three-level cache system to provide data storage and access services for each node of the UAV cluster, and ensure the efficient transfer of data at different processing stages; The data recognition module is used to perform preliminary processing and target recognition on the obtained aerial photography image data; The load balancing module is used to monitor the load weights of each node in real time and reasonably allocate new data processing tasks; The report generation module is used to integrate the recognition results of each node, eliminate stitching misalignment, and form and output a final report.

[0016] The data acquisition module includes: an image acquisition unit and a data splitting unit; The image acquisition unit is used to control the UAVs in the UAV cluster to collect aerial photography images of the grid areas they are responsible for according to the preset geographical division scheme; The data splitting unit is used to split the aerial photography image data in the local edge cache according to color and texture features to form a two-layer task, and upload the split subtasks and metadata including grid ID, feature type, and priority to the regional cache.

[0017] The data cache module includes: a local edge cache unit, a regional cache unit, and a global cache unit; The local edge cache unit is used to temporarily store the original aerial image data collected by the drones for the feature splitting and task allocation unit to read and process; The regional cache unit is used to store the split subtasks and metadata uploaded by the feature splitting and task allocation unit for each node to obtain corresponding data for processing according to the task allocation; The global cache unit is used to store the result data marked after target recognition by each node, providing data support for subsequent result aggregation and report generation.

[0018] The data recognition module includes: a data acquisition unit, a preliminary processing unit, and an intelligent collaborative recognition unit; The data acquisition unit is used to obtain the aerial image data allocated to itself from the regional cache in the three-level cache system; The preliminary processing unit is used to perform preliminary processing on the obtained aerial image data, such as data format conversion, noise reduction, and other operations; The intelligent collaborative recognition unit is used to use the intelligent collaborative recognition algorithm to interact with non-self nodes, jointly identify the targets in the aerial images, and store the marked target recognition results in the global cache.

[0019] The load balancing module includes: a load monitoring unit and a task allocation unit; The load monitoring unit monitors the load weights of each node in the drone cluster in real time through the central node, and calculates the load weights by comprehensively considering factors such as the processing capabilities of the nodes and the current task loads; The task allocation unit is used to use the weighted least connection number algorithm to allocate the newly arrived data processing tasks according to the load weight information of each node provided by the load monitoring unit, and preferentially allocate the tasks to the idle nodes with high weights to ensure system load balancing.

[0020] The report generation module includes: a result aggregation unit, a splicing processing unit, and a report generation unit; The result aggregation unit, after each node completes the task of aerial image recognition, the central node aggregates all the recognition results from the global cache according to the grid ID; The splicing processing unit uses the consistent hashing algorithm to process the aggregated recognition results, eliminate splicing misalignment, and ensure the integrity and accuracy of the image recognition results; The report generation unit is used to organize the recognition results after splicing processing, form an aerial image recognition report, and output the final result.

[0021] In summary, the overall implementation process of the distributed intelligent recognition processing system for collaborative aerial images of drone clusters includes: S1: The drones in the drone cluster conduct aerial image acquisition for their respective responsible grid areas according to a preset geographical division plan and store them in the local edge cache in real time; Furthermore, the specific steps of S1 include: (1) According to the geographical features, area size and aerial photography task requirements of the aerial photography area, a geographical division plan is formulated in advance, dividing the entire aerial photography area into N grid areas, and clarifying information such as the boundaries and coordinate ranges of each grid area; (2) According to the geographical division plan, each drone in the drone cluster is assigned its respective responsible grid area. Among them, the task information is transmitted to each drone through a wireless communication network. The task information includes the location information of the grid area and aerial photography parameters. The aerial photography parameters include, for example, flight altitude, speed, shooting angle, shooting time interval; (3) The drones fly to the airspace of their respective responsible grid areas according to the received task information and conduct image acquisition according to the preset aerial photography parameters. During the acquisition process, the cameras or other imaging devices on the drones continuously take pictures or record videos to obtain the aerial image data of the grid area; (4) After the drones acquire the aerial image data, they immediately store the data in the local edge cache device. It should be noted that the local edge cache can be a storage chip or a small storage device equipped on the drones to ensure that the data can be saved in a timely and safe manner for subsequent processing and transmission.

[0022] S2: Split the aerial image data in the local edge cache into double-layer tasks according to the split feature type, and upload the split subtasks and metadata to the regional cache; S3: Each node in the drone cluster obtains the aerial image data allocated to itself from the three-level cache system, conducts preliminary processing, and then jointly conducts target recognition on the aerial images using an intelligent collaborative recognition algorithm, and marks the target recognition results and stores them in the global cache; the three-level cache system includes local edge cache, regional cache, and global cache; S4: The central node monitors the load weights of each node in the drone cluster in real time and uses the weighted least connection number algorithm to allocate the newly arrived data processing tasks; the data processing tasks include the data to be collaboratively processed transmitted by the nodes and the processing tasks of newly acquired aerial images; S5: After each node completes the task of aerial image recognition, the central node aggregates all the recognition results from the global cache according to the grid ID, and uses the consistent hashing algorithm to eliminate stitching misalignment to form an aerial image recognition report and output the final result.

[0023] Furthermore, the specific steps of S5 include: After each node completes the aerial image recognition task, it sends a message indicating task completion to the central node. The message contains the node identifier and the grid ID information of the recognized grids. The central node receives and records the task completion information of each node to ensure that the recognition tasks of all nodes are completed. Based on the recorded task completion information of each node, the central node aggregates all the recognition results from the global cache according to the grid ID. Among them, the recognition results stored in the global cache usually include the category, location, and confidence information of the target, and are associated with the corresponding grid ID. The central node collects the recognition results corresponding to each grid ID together to form preliminary aggregated data. Define a 32-bit hash ring and map each storage node in the global cache to the hash ring through a hash function. In the present invention, the hash function uses the MD5 algorithm, and the MD5 algorithm is the prior art content in this field and is not the creative solution of this application, so it will not be elaborated here. The MD5 algorithm takes data of any length as input and generates a 128-bit hash value through operations. This hash value is usually represented by 32-bit hexadecimal digits. Its operation process mainly includes steps of grouping, initialization, and iterative processing, using logical operations and shift operations, and finally obtaining a hash value of a fixed length through multiple transformations of the input data.

[0024] For each grid ID, use the same hash function to map it to the hash ring. Then, starting from this mapping point, search clockwise for the first storage node encountered on the hash ring. This node is the ownership node of the recognition result of this grid ID. The recognition results of the same grid ID collected by different nodes may be spliced and misaligned. Through the consistent hashing algorithm, ensure that the recognition results of the same grid ID are always mapped to the same node, reduce the problem of inconsistent results caused by node changes, and at the same time, compare and merge the collected results to eliminate duplicate and contradictory information. Among them, the consistent hashing algorithm is the prior art content in this field and is not the creative solution of this application, so it will not be elaborated here. The central node sorts and statistically analyzes the recognition results processed by the consistent hashing algorithm according to the target category. Generate an aerial image recognition report based on the statistical results. Among them, the report content includes the total number of recognized targets, the quantity and distribution of various targets, and the recognition accuracy rate. The central node outputs the generated aerial image recognition report in JSON format for users to view and analyze.

[0025] Embodiment 2 Please refer to Figure 3 , in this embodiment, the double-layer task includes: Performing color feature splitting on the aerial image data in the local edge cache unit, and allocating the hyperspectral data obtained after color feature splitting to the nodes with GPUs in the cluster; Performing texture feature splitting on the aerial image data in the local edge cache unit, and allocating the texture complex data obtained after splitting according to the texture features to the nodes with FPGAs; The metadata includes grid ID, feature type, and priority.

[0026] The performing color feature splitting on the aerial image data in the local edge cache unit, and allocating the hyperspectral data obtained after color feature splitting to the nodes with GPUs in the cluster includes: A1: After the local edge cache unit receives the read instruction from the system, it extracts the aerial image data from the local edge cache and performs a preliminary check on the extracted aerial image data; If it is found that the extracted aerial image data has noise or color deviation, the Gaussian filtering algorithm and the histogram equalization method are used to remove the noise and correct the color, and the preprocessed aerial image data is obtained; Further, the specific steps of A1 include: (1) The local edge cache unit listens for instructions from the system at all times. When it receives the read instruction, the local edge cache unit starts to prepare to extract the aerial image data from the local edge cache; (2) According to the requirements of the read instruction, the local edge cache unit locates the position where the aerial image data is stored and extracts the aerial image data from the local edge cache; (3) Perform a preliminary check on the extracted aerial image data, mainly by calculating the mean value of the aerial image data to check for noise and color deviation. Among them, the noise is manifested as random pixel points in the aerial image, and the color deviation causes the image color to not match the actual scene; If it is found that the aerial image data has noise in the preliminary check, the Gaussian filtering algorithm is used to process the aerial image. Among them, the Gaussian filtering algorithm is the prior art content in this field and is not the creative solution of this application, so it will not be elaborated here; If it is found that the data has color deviation, the histogram equalization method is used to correct the color of the aerial image. Among them, the histogram equalization method is the prior art content in this field and is not the creative solution of this application, so it will not be elaborated here; (4) After noise removal and color correction, the preprocessed aerial image data is obtained and output.

[0027] A2: Through the formula Convert the preprocessed aerial image data from the original RGB color space to the CIELAB color space, and the formula satisfies the condition: and ; where X, Y, and Z respectively represent the red stimulus value, green stimulus value, and blue stimulus value, , , respectively represent the X, Y, and Z values of the reference white point, R, G, and B respectively represent the R channel value, G channel value, and B channel value of the preprocessed aerial image data, represents a constant that satisfies , and t represents the independent variable; It should be noted that by using , any visible color can be expressed as a linear combination of the three stimulus values X, Y, and Z. These three color matching functions respectively correspond to the stimulus values of the three primary colors red, green, and blue, and are obtained by statistically analyzing the color vision experimental data of a large number of observers.

[0028] It should be noted that the CIELAB color space aims to achieve a quantitative description of colors by simulating the human eye's visual perception, and the CIELAB value is a color space representation method widely used in the field of color, which is a device-independent color model and consists of three components, namely , and , where represents the brightness, and its value range is [0, 100], where 0 is absolute black and 100 is absolute white; is the red-green axis, representing the color component from green to red, with green being negative and red being positive in the value; is the yellow-blue axis, representing the color component from blue to yellow, with blue being negative and yellow being positive in the value.

[0029] A3: Split the CIELAB value in the CIELAB color space according to the preset color feature threshold, and separate the pixels in the aerial image that meet the color features to obtain the split hyperspectral data; Furthermore, the specific steps of A3 include: (1) For each pixel in the aerial image, extract the , and value of each pixel in the CIELAB color space; (2) Set the threshold ranges for the three components of , and ; (3)Traverse each pixel in the aerial image, and compare the , and values of each pixel with the threshold ranges of the three components of the preset , and ; (4)If the , and values of any pixel are all within their respective threshold ranges, then mark the data of this pixel as conforming to the color feature; otherwise, mark it as not conforming to the color feature; (5)Extract the pixel data of all pixels marked as conforming to the color feature to form a new data set, which is the split hyperspectral data.

[0030] A4: The central node collects the information of the GPU - equipped nodes in the cluster, calculates the comprehensive score of each node, and selects nodes according to the comprehensive scores of each node; Furthermore, the specific steps of A4 include: (1)The central node sends an information request message to all GPU - equipped nodes in the cluster through the network communication protocol TCP / IP; (2)After receiving the request message, the GPU - equipped nodes collect their own information, encapsulate the collected self - information into a message, and send it back to the central node through the network. Among them, the self - information includes computing power, current load situation, and network bandwidth; (3)After receiving the information returned by each node, the central node cleans and unifies the format of the data to ensure the accuracy and consistency of the data. For example, unify the video memory sizes represented in different units by different nodes into the same unit. At the same time, normalize the collected information to eliminate the influence of the dimension between different indicators; (4)The central node performs weighted summation on the information of each node according to the preset weight coefficients to obtain the comprehensive score of each node. Among them, the weight coefficients are determined based on specific business requirements and actual situations. For example, if more attention is paid to the computing power of the node, then increase the weight of the indicators related to computing power; (5)The central node sorts all nodes according to the calculated comprehensive scores; (6)Select the node with the highest comprehensive score from the sorted node list for task allocation.

[0031] A5: Package the split hyperspectral data and transmit it through the network to the node selected in A4.

[0032] The method of splitting the texture features of the aerial image data in the local edge cache unit and allocating the texture complex data obtained after splitting according to the texture features to the nodes equipped with FPGAs includes: B1: After receiving the texture feature splitting instruction, the local edge cache unit reads the aerial image data from the local edge cache and performs preprocessing; B2: Calculate the spatial relationship between different gray levels in the aerial image through the gray-level co-occurrence matrix to obtain texture feature parameters such as contrast, correlation, energy, and homogeneity. Among them, the gray-level co-occurrence matrix is obtained by traversing each pixel in the aerial image and statistically calculating all pixels according to the determined distance and direction parameters. The formula is the prior art content in this field and not the creative solution of this application, so it will not be elaborated here; B3: According to the obtained texture features, evaluate the complexity of the texture through contrast calculation. Among them, the contrast calculation method is the prior art content in this field and not the creative solution of this application, so it will not be elaborated here; B4: Set the texture complexity threshold to divide the pixels in the aerial image into two categories: texture complex and texture simple; B5: According to the texture complexity evaluation result, split the aerial image data into texture complex data and texture simple data according to the preset texture complexity threshold, and extract the texture complex data separately to form a data subset to be allocated; B6: The central node collects information about the nodes equipped with FPGAs in the cluster, including computing power, current load, and communication bandwidth; B7: According to the information of the nodes equipped with FPGAs, use the weighted round-robin algorithm to select FPGA nodes. Among them, the weighted round-robin algorithm is the prior art content in this field and not the creative solution of this application, so it will not be elaborated here; B8: Package the data subset to be allocated and transmit it to the FPGA node selected in B7 through the network.

[0033] Embodiment 3 Please refer to Figure 4 , in this embodiment, each node in the UAV cluster obtains the aerial image data allocated to itself from the data cache module, and uses an improved intelligent collaborative recognition algorithm in the data recognition module to jointly perform target recognition on the aerial image, and stores the marked target recognition result in the global cache unit, including: C1: Each node in the UAV cluster locates in the distributed index system of the data cache module according to the parsed double-layer task allocation information; the specific process of the location includes: The node sends a query request to the index server, and the index server returns the physical address of the aerial image data in the cache module according to the task identifier and data features in the request; C2: The node obtains the aerial image data allocated to itself from the data cache module through the distributed file system according to the physical address returned by the index server, and performs preprocessing; C3: Load the improved intelligent collaborative recognition algorithm model, and each node inputs the preprocessed aerial image into the improved intelligent collaborative recognition algorithm model for target recognition; the improved intelligent collaborative recognition algorithm model is configured based on the deep convolutional neural network, attention mechanism and multi-scale feature fusion; C4: After the target recognition is completed, each node marks the target recognition result; the marking is to draw a bounding box, label the target category and confidence on the original aerial image, where the confidence calculation formula is the prior art content in the field and is not the creative solution of this application, so it will not be elaborated here; Further, the specific steps of C4 include: (1) Each node extracts target information from the output of the target recognition algorithm, and the target information includes the category of the target, the coordinates of the bounding box and the confidence score, where the coordinates of the bounding box refer to the pixel coordinates of the upper left corner and the lower right corner; (2) Load the corresponding original aerial image from local storage or cache to ensure that the resolution and format of the image are the same as those used during recognition; (3) Define the style of the bounding box, such as the color and thickness of the line, and use the image processing library OpenCV to draw a rectangle on the original aerial image according to the extracted bounding box coordinates, where OpenCV is the prior art content in the field and is not the creative solution of this application, so it will not be elaborated here; (4) Perform annotation from above the bounding box to avoid blocking the target and the bounding box; (5) Use the image processing library to add text information of the target category and confidence at the determined position, and at the same time, define the font, size and color of the text to ensure that the annotation is clear and readable; (6) Determine the save path and file name of the marked image, and use the image processing library to save the marked image.

[0034] C5: Each node encapsulates the marked target recognition result, and stores the encapsulated target recognition result into the global cache through the network according to the storage location determined by the distributed index system of the double-layer task assignment information and the global cache.

[0035] The specific steps of C3 include: C3.1: Each node loads the pre-trained improved intelligent collaborative recognition algorithm model in local storage; C3.2: Each node inputs the preprocessed aerial images into the improved intelligent collaborative recognition algorithm model, performs convolution operations on the aerial images through the convolutional layer to extract multi-scale feature maps; in the convolution operation, the first 3 layers of the convolutional layer use separable convolution, and the last 2 layers of the convolutional layer use dilated convolution; C3.3: Perform weighted processing on the multi-scale feature maps through the attention mechanism to obtain weighted feature maps; C3.4: Input the weighted feature maps into the region proposal network, generate candidate regions containing the target through a sliding window, and perform classification and bounding box regression on the candidate regions to obtain the target recognition results; the target recognition results include the target category, position coordinates, and confidence. Among them, the region proposal network is the prior art content in the field and is not the creative solution of this application, so it will not be elaborated here; Further, the specific steps of C3.4 include: (1) Input the weighted feature maps into the region proposal network; (2) Determine the size and stride of the sliding window. Among them, the window size is selected according to the possible size of the target, and the stride determines the sliding interval of the window on the feature map; (3) Use the defined sliding window to traverse the weighted feature maps and process each window position; (4) Preset a set of anchor boxes with different scales and aspect ratios for each window position; (5) At each window position, generate the corresponding anchor boxes according to the preset anchor box parameters; (6) For each generated anchor box, extract its corresponding feature vector from the weighted feature maps; (7) Use the fully connected layer to perform classification prediction on the extracted anchor box features to determine whether each anchor box contains the target and the category of the target. In the present invention, the Softmax function is used to output the probability of each category. At the same time, another fully connected layer is used to perform regression prediction on the position and size of the anchor box to obtain more accurate target bounding box coordinates. Among them, the bounding box regression usually predicts four offsets, that is, the offsets of the center coordinates and width and height of the anchor box; (8) According to the probability of the classification prediction and the result of the bounding box regression, set a threshold. For example, filter out the anchor boxes with a classification probability lower than the preset threshold; (9) Perform non-maximum suppression processing on the remaining candidate regions, remove the redundant candidate regions with a high degree of overlap, and retain the candidate region with the highest confidence as the final target recognition result.

[0036] C3.5: Each node encapsulates the recognition results into structured data, encrypts them and sends them to 3 adjacent nodes within the communication radius. The adjacent nodes perform intersection over union calculation on the received target recognition results and the local target recognition results of the adjacent nodes; If the intersection over union (IoU) is greater than 0.5 and the confidence of the received target recognition result is greater than the confidence of the local target recognition result of the adjacent node, then replace the local target recognition result of the adjacent node with the received target recognition result; Specifically, each node locally maintains a data structure for storing recognition results, such as a list, where each element corresponds to a detected target and contains target category, bounding box coordinates, and confidence information. When the recognition result of an external node meets the conditions that the IoU is greater than 0.5 and the confidence is higher, this node will find the corresponding target element in the local data structure whose IoU with this external result is greater than 0.5, and update its category, bounding box coordinates, and confidence fields to the corresponding content of the external result. For example, in the local result, a certain target is a car, with bounding box coordinates (10, 10, 50, 50) and confidence 0.6, while in the external result, the corresponding target is a car, with bounding box coordinates (12, 12, 52, 52) and confidence 0.8, and the IoU between the two is greater than 0.5. At this time, all the information of this target in the local data structure will be replaced with the information of the external result. Here, the external result refers to the received target recognition result.

[0037] If the IoU is less than 0.3, then retain the received target recognition result as a new detection result; Specifically, when the IoU between the external result and any local recognition result is less than 0.3, it indicates that the target detected by this external result is not fully covered or repeatedly detected in the local result. Therefore, directly add this external result to the recognition result data structure maintained by this node. For example, if 2 targets are detected in the local result, and an external node transmits a new target detection result whose IoU with the 2 local targets is less than 0.3, then this new external result will be added as a new element to the local recognition result data structure and become one of the new target recognition results of this node.

[0038] C3.6: Output the final set of target recognition results.

[0039] The construction process of the pre-trained improved intelligent collaborative recognition algorithm model in C3.1 includes: C3.11: Define the application scenario of the model and collect the data of the model's application scenario. For example, if it is used for aerial image target recognition, aerial images in different regions, different weather conditions, and different times need to be collected. Here, the data source of the model's application scenario can be a public dataset, accumulated from actual projects, or self-collected; C3.12: Use the LabelImg annotation tool to accurately annotate the application scenario data of the collected model. The annotation content shall depend on the specific task. The LabelImg annotation tool is the prior art in this field and is not the creative solution of this application, so it will not be elaborated here. C3.13: Load the model architecture based on the deep convolutional neural network and conduct model testing. C3.14: Introduce an attention mechanism into the well-tested model architecture based on the deep convolutional neural network, and fuse the feature maps of different layers by combining the multi-scale feature fusion method, enabling the model to capture both the global and local features of the target simultaneously. At the same time, use the separable convolution technique to reduce the number of model parameters and the computational amount, improving the running efficiency of the model. The separable convolution technique is the prior art in this field and is not the creative solution of this application, so it will not be elaborated here. C3.15: Design the information interaction method between nodes, such as data transmission through a wireless communication network, and adopt broadcast and point-to-point communication methods to ensure that nodes can share recognition results and feature information in a timely and accurate manner. At the same time, formulate a cooperation strategy to form an improved intelligent cooperation recognition algorithm model. The cooperation strategy adopts the weighted voting method. For example, when multiple nodes identify the same target, the final recognition result is determined through the weighted voting method. The weights are determined according to factors such as the confidence level and computing power of the nodes, and the weighted voting method is the prior art in this field and is not the creative solution of this application, so it will not be elaborated here. C3.16: Divide the preprocessed aerial image data into a training set, a validation set, and a test set in the ratio of 7:1:2. At the same time, select the cross-entropy loss function and the stochastic gradient descent optimizer according to the task type. The cross-entropy loss function and the stochastic gradient descent optimizer are the prior art in this field and are not the creative solution of this application, so it will not be elaborated here. C3.17: Use the training set to train the improved intelligent cooperation recognition algorithm model, and continuously adjust the parameters of the improved intelligent cooperation recognition algorithm model during the training process to gradually reduce the loss function and obtain the trained improved intelligent cooperation recognition algorithm model.

[0040] The specific steps of the above C3.3 include: C3.31: Organize the multi-scale feature maps obtained in C3.2 according to the channel dimension, and perform global average pooling and global max pooling operations on each feature map to obtain the global information feature vector G. The global average pooling operation is implemented by calculating the average value of all pixels in each channel. The global max pooling operation is implemented by extracting the maximum value of all pixels in each channel. The channel dimension is [batch_size, channels, height, width], where batch_size represents the number of samples processed at one time, channels represents the number of channels of the aerial image data, height represents the height of the image or feature map, and width represents the width of the aerial image; C3.32: Input the global information feature vector into the shared multi-layer perceptron to obtain the channel attention weight matrix; Furthermore, the specific steps of C3.32 include: (1) Receive the global information feature vector G with the channel dimension of batch_size×2×channels; (2) Based on the global information feature vector G, perform a linear transformation through the weight matrix and the bias , and then obtain the output result after ReLU activation; (3) Perform a second linear transformation through the weight matrix and the bias , and use the Sigmoid activation function to obtain the channel attention weight matrix .

[0041] C3.33: Perform channel average pooling and channel max pooling operations on the multi-scale feature maps in the channel dimension to generate the spatial attention weight matrix; The channel average pooling is implemented by calculating the average value of all channels at each position; The channel max pooling operation is implemented by calculating the maximum value of all channels at each position; By concatenating the average value of all channels at each position and the maximum value of all channels at each position in the channel dimension, perform feature fusion through a 7×7 convolutional layer, and then generate the spatial attention weight matrix through the Sigmoid activation function; C3.34: Multiply the channel attention weight matrix and the spatial attention weight matrix element by element to obtain the attention weight matrix; C3.35: Multiply the attention weight matrix and the multi-scale feature map in C3.2 element by element to obtain the weighted feature map.

[0042] The task allocation unit uses the weighted minimum connection number algorithm, combines the load weights of each node in the UAV cluster monitored in real time by the load monitoring unit, and allocates the newly arrived data processing tasks, including: D1: Create an information record for each node in the UAV cluster; the information record includes the identification of the node, the current connection count, and the load weight information, where the load weight is comprehensively determined based on factors such as the computing power, memory size, and battery power of the node. D2: Start the load monitoring unit to enable it to start real-time monitoring of the load conditions of each node and update the load weights of each node. D3: After the task allocation unit receives a new data processing task, it obtains the current connection count and load weight of each node from the load monitoring unit. D4: For each node, the task allocation unit calculates its priority using the weighted least connection number algorithm; the weighted least connection number algorithm is implemented by calculating the ratio of the current connection count of node i to the load weight of node i. D5: According to the calculated node priorities, select the node with the highest priority to process the new task, and update the connection count of the selected node by adding 1 to it. D6: The selected node starts to execute the task assigned to it. At the same time, the load monitoring unit continues to real-time monitor the load conditions of each node and dynamically updates the load weights of the nodes. D7: When the node completes the task, subtract 1 from its connection count and wait for the arrival of a new task.

[0043] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make changes, modifications, substitutions, and variations to the above embodiments without departing from the purpose and scope of the present invention. All of these are within the protection scope of the present invention.

Claims

1. A distributed intelligent recognition and processing system for collaborative aerial photography images of UAV swarms, characterized in that, Including: A data acquisition module, a data caching module, a data recognition module, and a load balancing module; the data caching module includes a local edge caching unit, a regional caching unit, and a global caching unit; the load balancing module includes a load monitoring unit and a task allocation unit; The aerial image data collected by the data acquisition module in the local edge caching unit is split into double-layer tasks according to the split feature types, and the split subtasks and metadata are uploaded to the regional caching unit; Each node in the UAV cluster obtains the aerial image data allocated to itself from the data caching module, and uses an improved intelligent collaborative recognition algorithm in the data recognition module to jointly perform target recognition on the aerial images, and stores the marked target recognition results in the global caching unit; the improved intelligent collaborative recognition algorithm is configured based on a deep convolutional neural network, an attention mechanism, and multi-scale feature fusion; The task allocation unit uses the weighted least connection number algorithm, and combines the load weights of each node in the UAV cluster monitored by the load monitoring unit in real time to allocate the arriving new data processing tasks; After the new data processing task is completed, the central node aggregates all the recognition results from the global cache according to the grid ID to form an aerial image recognition report and outputs the final result.

2. The distributed intelligent recognition and processing system for collaborative aerial photography images of an unmanned aerial vehicle cluster according to claim 1, characterized in that, The double-layer tasks include: Performing color feature splitting on the aerial image data in the local edge caching unit, and allocating the hyperspectral data obtained after the color feature splitting to the nodes equipped with GPUs in the cluster; Performing texture feature splitting on the aerial image data in the local edge caching unit, and allocating the texture complex data obtained after the texture feature splitting to the nodes equipped with FPGAs; The metadata includes a grid ID, a feature type, and a priority.

3. The distributed intelligent recognition and processing system for collaborative aerial photography images of an unmanned aerial vehicle cluster according to claim 2, characterized in that, The performing color feature splitting on the aerial image data in the local edge caching unit, and allocating the hyperspectral data obtained after the color feature splitting to the nodes equipped with GPUs in the cluster includes: A1: After the local edge caching unit receives a read instruction from the system, it extracts the aerial image data from the local edge cache and performs a preliminary check on the extracted aerial image data; If it is found that the extracted aerial image data has noise or color deviation, the Gaussian filtering algorithm and the histogram equalization method are used to remove the noise and correct the color to obtain the preprocessed aerial image data; A2: Converting the preprocessed aerial image data from the original RGB color space to the CIELAB color space; A3: Splitting the CIELAB values in the CIELAB color space according to a preset color feature threshold, and separating the pixel data in the aerial image that conforms to the color feature to obtain the split hyperspectral data; A4: The central node collects the information of the nodes equipped with GPUs in the cluster, calculates the comprehensive score of each node, and selects nodes according to the comprehensive score of each node; A5: Encapsulating the split hyperspectral data and transmitting it through the network to the nodes selected in A4.

4. The distributed intelligent recognition and processing system for collaborative aerial photography images of an unmanned aerial vehicle cluster according to claim 3, wherein The splitting of the texture features of the aerial image data in the local edge cache unit and the allocation of the texture complex data obtained after splitting according to the texture features to the nodes equipped with FPGAs include: B1: After receiving the texture feature splitting instruction, the local edge cache unit reads the aerial image data from the local edge cache and performs preprocessing; B2: Calculate the spatial relationship between different gray levels in the aerial image through the gray level co-occurrence matrix to obtain texture feature parameters; the texture feature parameters include contrast, correlation, energy, and homogeneity; B3: Evaluate the complexity of the texture by calculating the contrast according to the obtained texture features; B4: Set a texture complexity threshold to divide the pixels in the aerial image into two categories: texture complex and texture simple; B5: According to the texture complexity evaluation result, split the aerial image data into texture complex data and texture simple data according to the preset texture complexity threshold, and separately extract the texture complex data to form a data subset to be allocated; B6: The central node collects information of the nodes equipped with FPGAs in the cluster; B7: According to the information of the nodes equipped with FPGAs, use the weighted round-robin algorithm to select FPGA nodes; B8: Package the data subset to be allocated and transmit it to the FPGA node selected in B7 through the network.

5. The distributed intelligent recognition and processing system for collaborative aerial photography images of an unmanned aerial vehicle cluster according to claim 4, characterized in that, Each node in the UAV cluster obtains the aerial image data allocated to itself from the data cache module, and uses an improved intelligent collaborative recognition algorithm in the data recognition module to jointly perform target recognition on the aerial image, and stores the marked target recognition result in the global cache unit, including: C1: Each node in the UAV cluster locates in the distributed index system of the data cache module according to the parsed double-layer task allocation information; the specific process of the location includes: The node sends a query request to the index server, and the index server returns the physical address of the aerial image data in the cache module according to the task identifier and data features in the request; C2: The node obtains the aerial image data allocated to itself from the data cache module through the distributed file system according to the physical address returned by the index server and performs preprocessing; C3: Load the improved intelligent collaborative recognition algorithm model, and each node inputs the preprocessed aerial image into the improved intelligent collaborative recognition algorithm model for target recognition; C4: After the target recognition is completed, each node marks the target recognition result; the marking is to draw a bounding box, label the target category and confidence on the original aerial image; C5: Each node packages the marked target recognition result and stores the packaged target recognition result in the global cache according to the storage location determined by the double-layer task allocation information and the distributed index system of the global cache through the network.

6. The distributed intelligent recognition and processing system for collaborative aerial photography images of an unmanned aerial vehicle cluster according to claim 5, characterized in that The specific steps of C3 include: C3.1: Each node loads the improved intelligent collaborative recognition algorithm model in the local storage; C3.2: Each node inputs the pre - processed aerial images into the improved intelligent collaborative recognition algorithm model. The convolutional layer performs convolutional operations on the aerial images to extract multi - scale feature maps. In the convolutional operations, the first 3 layers of the convolutional layer use depthwise separable convolutions, and the last 2 layers of the convolutional layer use dilated convolutions; C3.3: The multi - scale feature maps are weighted through an attention mechanism to obtain weighted feature maps; C3.4: The weighted feature maps are input into the region proposal network. Candidate regions containing the target are generated through a sliding window, and the candidate regions are classified and the bounding boxes are regressed to obtain the target recognition results. The target recognition results include the target category, location coordinates, and confidence; C3.5: Each node encapsulates the recognition results into structured data, encrypts them, and sends them to 3 adjacent nodes within the communication radius. The adjacent nodes calculate the intersection - over - union (IoU) between the received target recognition results and the local target recognition results of the adjacent nodes; If the IoU is greater than 0.5 and the confidence of the received target recognition results is greater than the confidence of the local target recognition results of the adjacent nodes, then the local target recognition results of the adjacent nodes are replaced with the received target recognition results; If the IoU is less than 0.3, then the received target recognition results are retained as new detection results; C3.6: Output the final target recognition result set.

7. The distributed intelligent recognition and processing system for collaborative aerial photography images of an unmanned aerial vehicle cluster according to claim 6, characterized in that, The construction process of the pre - trained improved intelligent collaborative recognition algorithm model in C3.1 includes: C3.11: Define the application scenario of the model and collect the data of the model's application scenario; C3.12: Use the LabelImg annotation tool to accurately annotate the collected data of the model's application scenario; C3.13: Load the model architecture based on the deep convolutional neural network and conduct tests; C3.14: Introduce an attention mechanism into the well - tested model architecture based on the deep convolutional neural network, and fuse the feature maps of different layers by combining the multi - scale feature fusion method and the depthwise separable convolution technology; C3.15: Design the information interaction method and communication method between nodes, and formulate a collaborative strategy to form an improved intelligent collaborative recognition algorithm model; C3.16: Divide the pre - processed aerial image data into a training set, a validation set, and a test set, and select the cross - entropy loss function and the stochastic gradient descent optimizer according to the task type; C3.17: Use the training set to train the improved intelligent collaborative recognition algorithm model to obtain the trained improved intelligent collaborative recognition algorithm model.

8. The distributed intelligent recognition and processing system for collaborative aerial photography images of an unmanned aerial vehicle cluster according to claim 7, characterized in that The specific steps of C3.3 include: C3.31: Organize the multi - scale feature maps obtained in C3.2 according to the channel dimension, and perform global average pooling and global max pooling operations on each feature map to obtain global information feature vectors; The global average pooling operation is implemented by calculating the average value of all pixels in each channel; The global max pooling operation is implemented by extracting the maximum value of all pixels in each channel; The channel dimension is [batch_size, channels, height, width], where batch_size represents the number of samples processed at one time, channels represents the number of channels of the aerial image data, height represents the height of the image or feature map, and width represents the width of the aerial image; C3.32: Input the global information feature vector into the shared multi-layer perceptron to obtain the channel attention weight matrix; C3.33: Perform channel average pooling and channel max pooling operations on the multi-scale feature maps in the channel dimension to generate the spatial attention weight matrix; The channel average pooling is implemented by calculating the average value of all channels at each position; The channel max pooling operation is implemented by calculating the maximum value of all channels at each position; C3.34: Multiply the channel attention weight matrix and the spatial attention weight matrix element by element to obtain the attention weight matrix; C3.35: Multiply the attention weight matrix and the multi-scale feature maps in C3.2 element by element to obtain the weighted feature maps.

9. The distributed intelligent recognition and processing system for collaborative aerial photography images of an unmanned aerial vehicle cluster according to claim 8, wherein The task allocation unit uses the weighted least connection number algorithm and combines the load weights of each node in the UAV cluster monitored in real time by the load monitoring unit to allocate the newly arrived data processing tasks, including: D1: Create an information record for each node in the UAV cluster; the information record includes the identification of the node, the current connection number, and the load weight information; D2: Start the load monitoring unit to make it start monitoring the load conditions of each node in real time and update the load weights of each node; D3: After the task allocation unit receives a new data processing task, obtain the current connection number and load weight of each node from the load monitoring unit; D4: For each node, the task allocation unit calculates its priority using the weighted least connection number algorithm; the weighted least connection number algorithm is implemented by calculating the ratio of the current connection number of node i to the load weight of node i; D5: According to the calculated node priorities, select the node with the highest priority to process the new task, and update the connection number of the selected node by adding 1 to it; D6: The selected node starts to execute the task assigned to it. At the same time, the load monitoring unit continues to monitor the load conditions of each node in real time and dynamically updates the load weights of the nodes; D7: When the node completes the task, subtract 1 from its connection number and wait for the arrival of a new task.

Images