Distributed intelligent recognition and processing system for drone cluster collaborative aerial images

Abstract

The present invention discloses a distributed intelligent recognition and processing system for collaborative aerial images of a cluster of unmanned aerial vehicles (UAVs), which belongs to the technical field of UAV image processing. The system specifically comprises: UAVs collect aerial images according to a geographical division scheme and store them in a local edge cache, and then split the aerial image data into two-layer tasks according to color and texture features and upload them to a regional cache; each node obtains data from a three-level cache system, completes target recognition using an intelligent collaborative recognition algorithm after preliminary processing, and stores the results in a global cache; the central node monitors the load weight of each node in real time, and allocates new tasks using a weighted minimum connection number algorithm; the central node aggregates the recognition results in the global cache according to the grid ID, eliminates splicing dislocations through a consistent hashing algorithm, and generates an aerial image recognition report output; the present invention effectively solves the problems of low processing efficiency and large collaborative errors in a centralized architecture, and improves the efficiency and accuracy of aerial image processing of UAV clusters.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle (UAV) image processing, and specifically relates to a distributed intelligent recognition and processing system for collaborative aerial images of a UAV cluster. Background Art

[0002] Traditional drone aerial image processing typically uses a centralized processing model, where all drones transmit the collected image data back to a ground station or cloud server for unified analysis. However, this approach has the following problems: a single server takes 30 minutes to process 20GB of imagery. As the workload of a drone cluster increases, a large amount of image data accumulates on a single node, causing a sharp drop in system processing efficiency and even system crashes, seriously hindering the timely completion of tasks. Multiple drones identify the same target, resulting in repeated calculations and prone to misjudgments during the recognition process, affecting image recognition accuracy. In electromagnetic interference environments, the data transmission error rate in a centralized architecture increases significantly, leading to problems such as image stitching misalignment, making it impossible to meet the requirements of high-quality image processing. Therefore, a new processing method is urgently needed to address these issues. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention proposes a distributed intelligent recognition and processing system for collaborative aerial images of drone clusters. The drones collect aerial images according to a geographical division scheme and store them in a local edge cache. The aerial image data is then split into two-layer tasks according to color and texture features and uploaded to the regional cache; each node obtains data from the three-level cache system, completes target recognition using an intelligent collaborative recognition algorithm after preliminary processing, and stores the results in a global cache; the central node monitors the load weight of each node in real time, and assigns new tasks using a weighted minimum connection number algorithm; the central node aggregates the recognition results in the global cache according to the grid ID, eliminates splicing dislocations through a consistent hashing algorithm, and generates an aerial image recognition report output; the present invention effectively solves the problems of low processing efficiency and large collaborative errors in a centralized architecture, and improves the efficiency and accuracy of drone cluster aerial image processing.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A distributed intelligent recognition and processing system for collaborative aerial imagery of a drone cluster includes: 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;

[0006] Split the aerial image data collected by the data acquisition module in the local edge cache unit into two-layer tasks according to the split feature type, and upload the split subtasks and metadata to the regional cache unit;

[0007] Each node in the drone cluster obtains its assigned aerial image data from the data cache module, and uses an improved intelligent collaborative recognition algorithm in the data recognition module to jointly identify targets in the aerial images. The target recognition results are marked and stored 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.

[0008] The task allocation unit uses a weighted minimum connection number algorithm, combined with the load weight of each node in the drone cluster monitored in real time by the load monitoring unit, to allocate the newly arrived data processing tasks;

[0009] 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.

[0010] Specifically, the two-layer tasks include:

[0011] Perform color feature splitting on the aerial image data in the local edge cache unit, and distribute the hyperspectral data obtained after color feature splitting to the GPU-equipped nodes in the cluster;

[0012] Perform texture feature segmentation on the aerial image data in the local edge cache unit, and distribute the complex texture data obtained after the texture feature segmentation to the node equipped with FPGA;

[0013] The metadata includes grid ID, feature type, and priority.

[0014] Specifically, performing color feature splitting on the aerial image data in the local edge cache unit and distributing the hyperspectral data obtained after the color feature splitting to the GPU-equipped nodes in the cluster includes:

[0015] A1: After receiving the read instruction from the system, the local edge cache unit extracts the aerial image data from the local edge cache and performs a preliminary check on the extracted aerial image data;

[0016] If the extracted aerial image data is found to have 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;

[0017] A2: Convert the pre-processed aerial image data from the original RGB color space to the CIELAB color space;

[0018] A3: Split the CIELAB values ​​in the CIELAB color space according to the preset color feature threshold, separate the pixel data in the aerial image that meets the color feature, and obtain the split hyperspectral data;

[0019] A4: The central node collects information about GPU-equipped nodes in the cluster, calculates the comprehensive score of each node, and selects nodes based on the comprehensive score.

[0020] A5: Encapsulate the split hyperspectral data and transmit it to the node selected in A4 through the network.

[0021] Specifically, the aerial image data in the local edge cache unit is split according to texture features, and the complex texture data obtained after the split according to texture features is distributed to the node equipped with FPGA, including:

[0022] 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;

[0023] B2: Calculating the spatial relationship between different gray levels in the aerial image using the gray level co-occurrence matrix to obtain texture feature parameters; the texture feature parameters include contrast, correlation, energy, and homogeneity;

[0024] B3: Based on the obtained texture features, the complexity of the texture is evaluated by contrast calculation;

[0025] B4: Set the texture complexity threshold to classify the pixels in the aerial image into two categories: complex texture and simple texture;

[0026] B5: Based on the texture complexity evaluation results, the aerial image data is split into texture-complex data and texture-simple data according to a preset texture complexity threshold, and the texture-complex data is extracted separately to form a data subset to be allocated;

[0027] B6: The central node collects information about FPGA-equipped nodes in the cluster;

[0028] B7: Based on the information of the equipped FPGA nodes, a weighted round-robin algorithm is used to select the FPGA node;

[0029] B8: Encapsulate the data subset to be distributed and transmit it to the FPGA node selected in B7 through the network.

[0030] Specifically, each node in the drone cluster obtains the aerial image data assigned to it from the data cache module, uses the improved intelligent collaborative recognition algorithm in the data recognition module to jointly identify the target in the aerial image, and marks the target recognition result and stores it in the global cache unit, including:

[0031] C1: Each node in the drone cluster is located in the distributed index system of the data cache module based on the parsed two-layer task allocation information. The specific process of positioning includes:

[0032] 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 based on the task identifier and data characteristics in the request;

[0033] C2: Based on the physical address returned by the index server, the node obtains the aerial image data assigned to it from the data cache module through the distributed file system and performs preprocessing.

[0034] C3: Loading the improved intelligent collaborative recognition algorithm model. Each node inputs the pre-processed 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.

[0035] C4: After target recognition is completed, each node marks the target recognition result; the marking is to draw a bounding box on the original aerial image and mark the target category and confidence level;

[0036] C5: Each node encapsulates the marked target recognition results and stores the encapsulated target recognition results into the global cache through the network according to the storage location determined by the two-layer task allocation information and the distributed index system of the global cache.

[0037] Specifically, the steps of C3 include:

[0038] C3.1: Each node loads the improved intelligent collaborative recognition algorithm model into local storage;

[0039] C3.2: Each node inputs the preprocessed aerial imagery into the improved intelligent collaborative recognition algorithm model, performs convolution operations on the aerial imagery through convolutional layers, and extracts multi-scale feature maps. In this convolution operation, the first three layers of the convolutional layer use separable convolutions, and the last two layers of the convolutional layer use dilated convolutions.

[0040] C3.3: Weighted processing of multi-scale feature maps is performed through the attention mechanism to obtain weighted feature maps;

[0041] C3.4: Input the weighted feature map 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 target recognition results; the target recognition results include target category, location coordinates, and confidence level;

[0042] C3.5: Each node encapsulates the recognition result into structured data, encrypts it, and sends it to three neighboring nodes within the communication radius. The neighboring nodes then perform an intersection-and-comparison calculation on the received target recognition result with their local target recognition result.

[0043] If the intersection-over-union ratio 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, the local target recognition result of the adjacent node is replaced with the received target recognition result;

[0044] If the intersection-over-union ratio is less than 0.3, the received target recognition result is retained as the new detection result;

[0045] C3.6: Output the final target recognition result set.

[0046] Specifically, the construction process of the improved intelligent collaborative recognition algorithm model pre-trained in C3.1 includes:

[0047] C3.11: Define the application scenarios of the model and collect data on the application scenarios of the model;

[0048] C3.12: Use the LabelImg annotation tool to accurately label the collected model application scenario data;

[0049] C3.13: Load and test the deep convolutional neural network model architecture;

[0050] C3.14: Introduce an attention mechanism into a tested deep convolutional neural network model architecture, combining multi-scale feature fusion and separable convolution techniques to fuse feature maps from different layers.

[0051] C3.15: Design information interaction and communication methods between nodes, formulate collaborative strategies, and form an improved intelligent collaborative recognition algorithm model;

[0052] C3.16: Divide the preprocessed aerial image data into training, validation, and test sets, and select the cross-entropy loss function and stochastic gradient descent optimizer based on the task type.

[0053] C3.17: Use the training set to train the improved intelligent collaborative recognition algorithm model to obtain a trained improved intelligent collaborative recognition algorithm model.

[0054] Specifically, the steps of C3.3 include:

[0055] C3.31: Arrange the multi-scale feature maps obtained in C3.2 according to the channel dimension and perform global average pooling and global maximum pooling operations on each feature map to obtain the global information feature vector;

[0056] The global average pooling operation is implemented by calculating the average value of all pixels in each channel;

[0057] The global maximum pooling operation is implemented by extracting the maximum value of all pixels in each channel;

[0058] The channel dimension is [batch_size, channels, height, width], where batch_size represents the number of samples processed at a 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;

[0059] C3.32: Input the global information feature vector into the shared multi-layer perceptron to obtain the channel attention weight matrix;

[0060] C3.33: Perform channel average pooling and channel maximum pooling operations on the multi-scale feature map in the channel dimension to generate a spatial attention weight matrix;

[0061] The channel average pooling is achieved by calculating the average value of all channels at each position;

[0062] The channel maximum pooling operation is implemented by calculating the maximum value of all channels at each position;

[0063] C3.34: Multiply the channel attention weight matrix and the spatial attention weight matrix element by element to obtain the attention weight matrix;

[0064] C3.35: Multiply the attention weight matrix by the multi-scale feature map in C3.2 element-wise to obtain the weighted feature map.

[0065] Specifically, the task allocation unit uses a weighted minimum connection number algorithm, combined with the load weight of each node in the drone cluster monitored in real time by the load monitoring unit, to allocate the newly arrived data processing tasks, including:

[0066] D1: Create an information record for each node in the drone cluster; the information record includes the node's identification, current number of connections, and load weight information;

[0067] D2: Start the load monitoring unit to monitor the load of each node in real time and update the load weight of each node;

[0068] D3: After receiving the new data processing task, the task allocation unit obtains the current number of connections and load weight of each node from the load monitoring unit;

[0069] D4: The task allocation unit calculates the priority of each node using a weighted minimum connection number algorithm. The weighted minimum connection number algorithm is implemented by calculating the ratio of the current number of connections of node i to the load weight of node i.

[0070] D5: Based on the calculated node priority, select the node with the highest priority to process the new task, and update the connection count of the selected node by adding 1;

[0071] 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 of each node in real time and dynamically updates the load weight of the node;

[0072] D7: When a node completes a task, it reduces its connection count by 1 and waits for new tasks to arrive.

[0073] Compared with the prior art, the present invention has the following beneficial effects:

[0074] The present invention proposes a distributed intelligent recognition and processing system for collaborative aerial imagery from drone swarms. During the data collection and processing phase, a geographic partitioning and feature splitting strategy distributes aerial imagery tasks to each node, preventing excessive processing pressure on a single node. The coordination of local edge caching and regional caching enables efficient data storage and distribution. Nodes equipped with GPUs and FPGAs divide processing tasks based on image features, leveraging hardware advantages and significantly reducing processing time. The application of intelligent collaborative recognition algorithms integrates multi-node data, effectively reducing the misjudgment rate of repeated recognition by multiple machines, improving target recognition accuracy, and making image recognition results more reliable. The present invention proposes a distributed intelligent recognition and processing system for collaborative aerial imagery from drone swarms. Load balancing and result integration mechanisms ensure system stability and processing quality. The central node monitors the load weight of each node in real time and dynamically allocates tasks using a weighted minimum connection algorithm to prevent single-node overload and ensure efficient and stable system operation. During result processing, recognition results are aggregated by grid ID, and a consistent hashing algorithm is used to eliminate splicing misalignments, generating a complete and accurate aerial image recognition report. This enhances the system's adaptability in complex environments, effectively addresses the shortcomings of traditional centralized architectures, and provides a superior solution for drone swarm aerial imagery processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 This is a flow chart showing the principle of the distributed intelligent recognition and processing system for collaborative aerial photography by drone clusters of the present invention;

[0076] Figure 2 This is an architecture diagram of the distributed intelligent recognition and processing system for collaborative aerial photography of drone clusters of the present invention;

[0077] Figure 3This is a flow chart of hyperspectral data acquisition for the distributed intelligent recognition and processing system for collaborative aerial imagery of UAV clusters of the present invention;

[0078] Figure 4 This is a target recognition flow chart of the distributed intelligent recognition and processing system for collaborative aerial imagery of drone clusters of the present invention. DETAILED DESCRIPTION

[0079] Example 1

[0080] See also Figure 1 and Figure 2 The present invention provides an embodiment of a distributed intelligent recognition and processing system for drone cluster collaborative aerial images, comprising:

[0081] Data acquisition module, data cache module, data identification module, load balancing module, report generation module;

[0082] The data acquisition module uses drones to collect aerial images of each grid area, and performs temporary storage and data splitting and uploading;

[0083] The data cache module is used to build a three-level cache system to provide data storage and access services for each node in the drone cluster, ensuring the efficient flow of data at different processing stages;

[0084] Data recognition module, used for preliminary processing of acquired aerial image data and target recognition;

[0085] Load balancing module, used to monitor the load weight of each node in real time and reasonably allocate new data processing tasks;

[0086] The report generation module is used to integrate the recognition results of each node, eliminate splicing dislocations, and form and output the final report.

[0087] The data acquisition module includes: an image acquisition unit and a data splitting unit;

[0088] The image acquisition unit is used to control the drones in the drone cluster to collect aerial images of the grid areas they are responsible for according to the pre-set geographical division plan;

[0089] The data splitting unit is used to split the aerial 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.

[0090] The data cache module includes: local edge cache unit, regional cache unit, and global cache unit;

[0091] The local edge cache unit is used to temporarily store the raw aerial image data collected by the drone for reading and processing by the feature segmentation and task allocation unit;

[0092] The regional cache unit is used to store the split subtasks and metadata uploaded by the feature splitting and task assignment unit, so that each node can obtain the corresponding data for processing according to the task assignment;

[0093] The global cache unit is used to store the result data marked by each node after target recognition, providing data support for subsequent result aggregation and report generation.

[0094] The data recognition module includes: data acquisition unit, preliminary processing unit, and intelligent collaborative recognition unit;

[0095] A data acquisition unit is used to acquire the aerial image data allocated to it from the regional cache in the three-level cache system;

[0096] A preliminary processing unit is used to perform preliminary processing on the acquired aerial image data, such as data format conversion, noise reduction, etc.

[0097] The intelligent collaborative recognition unit is used to utilize the intelligent collaborative recognition algorithm to interact with non-self nodes to jointly identify targets in aerial images, and mark the target recognition results and store them in the global cache.

[0098] The load balancing module includes: load monitoring unit and task allocation unit;

[0099] The load monitoring unit monitors the load weight of each node in the drone cluster in real time through the central node, and calculates the load weight by comprehensively considering factors such as the node's processing capacity and the current task load;

[0100] The task allocation unit is used to use the weighted minimum 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 give priority to allocating tasks to idle nodes with high weights to ensure system load balancing.

[0101] The report generation module includes: result aggregation unit, splicing processing unit, and report generation unit;

[0102] Result aggregation unit, after each node completes the task of aerial image recognition, the central node aggregates all recognition results from the global cache according to the grid ID;

[0103] The splicing processing unit uses a consistent hashing algorithm to process the aggregated recognition results, eliminate splicing dislocations, and ensure the integrity and accuracy of the image recognition results;

[0104] The report generation unit is used to organize the recognition results after splicing processing, form an aerial image recognition report, and output the final results.

[0105] In summary, the overall implementation process of the distributed intelligent recognition and processing system for drone cluster collaborative aerial images includes:

[0106] S1: The drones in the drone cluster collect aerial images of their respective grid areas according to the pre-set geographical division scheme and store them in the local edge cache in real time;

[0107] Furthermore, the specific steps of S1 include:

[0108] (1) Based on the geographical characteristics, size, and mission requirements of the aerial photography area, a geographic division plan is formulated in advance to divide the entire aerial photography area into N grid areas, and the boundaries, coordinate range, and other information of each grid area are clearly defined;

[0109] (2) Based on the geographical division scheme, each drone in the drone cluster is assigned its own grid area. Mission information is transmitted to each drone via a wireless communication network. The mission information includes the location information of the grid area and aerial photography parameters. The aerial photography parameters include flight altitude, speed, shooting angle, and shooting time interval.

[0110] (3) The drones fly to the airspace above the grid area they are responsible for according to the received mission information and collect images according to the preset aerial photography parameters. During the collection process, the camera or other imaging equipment on the drone continuously takes photos or records videos to obtain aerial image data of the grid area;

[0111] (4) After collecting aerial image data, the drone immediately stores the data in a 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 drone to ensure that the data can be saved in a timely and secure manner for subsequent processing and transmission.

[0112] S2: Split the aerial image data in the local edge cache into two-layer tasks according to the split feature type, and upload the split subtasks and metadata to the regional cache;

[0113] S3: Each node in the drone cluster obtains the aerial image data assigned to it from the three-level cache system and performs preliminary processing. It then uses an intelligent collaborative recognition algorithm to jointly identify targets in the aerial imagery and stores the target recognition results in a global cache after marking them. The three-level cache system includes a local edge cache, a regional cache, and a global cache.

[0114] S4: The central node monitors the load weight of each node in the drone cluster in real time and uses the weighted minimum connection number algorithm to allocate new data processing tasks that arrive; the data processing tasks include the data to be collaboratively processed transmitted by the nodes and the processing tasks of newly collected aerial images;

[0115] S5: After each node completes the aerial image recognition task, the central node aggregates all recognition results from the global cache according to the grid ID, and uses the consistent hashing algorithm to eliminate splicing dislocations, forms an aerial image recognition report, and outputs the final result.

[0116] Furthermore, the specific steps of S5 include:

[0117] (1) After completing the aerial image recognition task, each node sends a task completion message to the central node, which contains the node identifier and the ID information of the grid that has been identified;

[0118] (2) The central node receives and records the task completion information of each node to ensure that the identification tasks of all nodes have been completed;

[0119] (3) The central node aggregates all recognition results from the global cache according to the grid ID based on the recorded task completion information of each node. The recognition results stored in the global cache usually contain the target category, location, and confidence information, and are associated with the corresponding grid ID;

[0120] (4) The central node collects the identification results corresponding to each grid ID to form preliminary aggregated data;

[0121] (5) A 32-bit hash ring is defined, and each storage node in the global cache is mapped to the hash ring through a hash function. In the present invention, the hash function adopts the MD5 algorithm, which is a prior art in this field and is not an inventive solution of the present application, and is not described in detail here;

[0122] The MD5 algorithm takes data of arbitrary length as input and generates a 128-bit hash value through calculation. This hash value is usually represented by a 32-bit hexadecimal number. Its calculation process mainly includes grouping, initialization, and iterative processing steps. It utilizes logical operations and shift operations, and finally obtains a fixed-length hash value by performing multiple transformations on the input data.

[0123] (6) For each grid ID, use the same hash function to map it to the hash ring. Then, starting from the mapping point, search clockwise for the first storage node encountered on the hash ring. This node is the home node of the grid ID identification result.

[0124] (7) The identification results of the same grid ID collected at different nodes may be misaligned, and the consistent hashing algorithm is used to ensure that the identification results of the same grid ID are always mapped to the same node, reducing the problem of inconsistent results caused by node changes. At the same time, the collected results are compared and merged to eliminate duplicate and contradictory information. The consistent hashing algorithm is the existing technology content in this field and is not the inventive solution of this application, so it will not be described in detail here;

[0125] (8) The central node organizes and counts the recognition results processed by the consistent hashing algorithm according to the target category;

[0126] (9) Generate an aerial image recognition report based on the statistical results, which includes the total number of targets identified, the number and distribution of each type of target, and the recognition accuracy;

[0127] (10) The central node outputs the generated aerial image recognition report in JSON format for users to view and analyze.

[0128] Example 2

[0129] See also Figure 3 , the dual-layer tasks in this embodiment include:

[0130] Perform color feature splitting on the aerial image data in the local edge cache unit, and distribute the hyperspectral data obtained after color feature splitting to the GPU-equipped nodes in the cluster;

[0131] Perform texture feature segmentation on the aerial image data in the local edge cache unit, and distribute the complex texture data obtained after the texture feature segmentation to the node equipped with FPGA;

[0132] The metadata includes grid ID, feature type, and priority.

[0133] The method of performing color feature splitting on the aerial image data in the local edge cache unit and distributing the hyperspectral data obtained after the color feature splitting to the GPU-equipped nodes in the cluster includes:

[0134] A1: After receiving the read instruction from the system, the local edge cache unit extracts the aerial image data from the local edge cache and performs a preliminary check on the extracted aerial image data;

[0135] If the extracted aerial image data is found to have 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;

[0136] Furthermore, the specific steps of A1 include:

[0137] (1) The local edge cache unit always listens to the instructions from the system. When receiving the read instruction, the local edge cache unit starts to prepare to extract the aerial image data from the local edge cache;

[0138] (2) According to the requirements of the read instruction, the local edge cache unit locates the location where the aerial image data is stored and extracts the aerial image data from the local edge cache;

[0139] (3) Conduct a preliminary inspection of the extracted aerial image data, mainly by calculating the mean of the aerial image data to check whether there is noise and color deviation. Noise appears as random pixels in the aerial image, while color deviation causes the image color to be inconsistent with the actual scene;

[0140] If noise is found in the aerial image data during the preliminary inspection, the aerial image is processed using a Gaussian filtering algorithm. The Gaussian filtering algorithm is a prior art in this field and is not an inventive solution of the present application, so it will not be described in detail here.

[0141] If color deviation is found in the data, the aerial image is color corrected using a histogram equalization method. The histogram equalization method is a prior art in this field and is not an inventive solution of the present application, so it will not be described in detail here.

[0142] (4) After noise removal and color correction, the preprocessed aerial image data is obtained and output.

[0143] A2: By formula The pre-processed aerial image data is converted from the original RGB color space to the CIELAB color space, and the formula meets the following conditions:

[0144] and ;

[0145] Among them, X, Y, and Z represent the red stimulus value, green stimulus value, and blue stimulus value respectively. 、 、 Respectively represent the X, Y, and Z values ​​of the reference white point, and R, G, and B represent the R channel value, G channel value, and B channel value of the pre-processed aerial image data. represents a constant, satisfying , t represents the independent variable;

[0146] It should be noted that by using Any visible color can be expressed as a linear combination of three stimulus values, X, Y, and Z. These three color matching functions correspond to the stimulus values ​​of the three primary colors, red, green, and blue, respectively. They are obtained by statistical analysis of color vision experimental data from a large number of observers.

[0147] It should be noted that the CIELAB color space is designed to achieve quantitative description of color by simulating human visual perception, and the CIELAB value is a color space representation method widely used in the color field. It is a device-independent color model composed of three components, namely 、 and ,in, Indicates brightness, the value range is [0, 100], 0 is absolute black, 100 is absolute white; The red-green axis represents the color components from green to red, where green is a negative value and red is a positive value; The yellow-blue axis represents the color components from blue to yellow, where blue is a negative value and yellow is a positive value.

[0148] A3: Split the CIELAB values ​​in the CIELAB color space according to the preset color feature threshold, separate the pixels in the aerial image that meet the color features, and obtain the split hyperspectral data;

[0149] Furthermore, the specific steps of A3 include:

[0150] (1) For each pixel in the aerial image, extract the 、 Sum ;

[0151] (2) Settings 、 and Threshold ranges for the three components;

[0152] (3) Traverse each pixel in the aerial image and convert each pixel 、 and Value and preset 、 and The threshold ranges of the three components were compared;

[0153] (4) If any pixel 、 and If the values ​​are within their respective threshold ranges, the pixel data is marked as meeting the color feature; otherwise, it is marked as not meeting the color feature.

[0154] (5) All pixel data marked as meeting the color features are extracted to form a new data set, which is the split hyperspectral data.

[0155] A4: The central node collects information about GPU-equipped nodes in the cluster, calculates the comprehensive score of each node, and selects nodes based on the comprehensive score.

[0156] Furthermore, the specific steps of A4 include:

[0157] (1) The central node sends information request messages to all GPU-equipped nodes in the cluster through the network communication protocol TCP / IP;

[0158] (2) After receiving the request message, the GPU-equipped node collects its own information, encapsulates the collected information into a message, and sends it back to the central node through the network. The information includes computing power, current load, and network bandwidth.

[0159] (3) After receiving the information returned by each node, the central node cleans the data and unifies the format to ensure the accuracy and consistency of the data. For example, the memory size expressed in different units of different nodes is converted to the same unit. At the same time, the collected information is normalized to eliminate the impact of different dimensions of different indicators.

[0160] (4) The central node performs a weighted summation of the information of each node according to the preset weight coefficient to obtain the comprehensive score of each node. The weight coefficient is determined based on the specific business needs and actual situation. For example, if more emphasis is placed on the computing power of the node, the weight of the computing power-related indicators will be increased;

[0161] (5) The central node ranks all nodes according to the calculated comprehensive scores;

[0162] (6) Select the node with the highest comprehensive score from the sorted node list for task assignment.

[0163] A5: Encapsulate the split hyperspectral data and transmit it to the node selected in A4 through the network.

[0164] The method of performing texture feature splitting on the aerial image data in the local edge cache unit and distributing the texture complex data obtained after the texture feature splitting to the node equipped with the FPGA includes:

[0165] 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;

[0166] B2: Calculate the spatial relationship between different gray levels in the aerial image using a gray-level co-occurrence matrix to obtain texture feature parameters such as contrast, correlation, energy, and homogeneity. The gray-level co-occurrence matrix is ​​calculated by traversing each pixel in the aerial image and counting all pixels based on the determined distance and direction parameters. This formula is the prior art in this field and does not constitute the inventive solution of this application, so it will not be described in detail here.

[0167] B3: Evaluate the complexity of the texture by contrast calculation based on the obtained texture features. The contrast calculation method is a prior art in this field and does not constitute an inventive solution of the present application, and is not described in detail here.

[0168] B4: Set the texture complexity threshold to classify the pixels in the aerial image into two categories: complex texture and simple texture;

[0169] B5: Based on the texture complexity evaluation results, the aerial image data is split into texture-complex data and texture-simple data according to a preset texture complexity threshold, and the texture-complex data is extracted separately to form a data subset to be allocated;

[0170] B6: The central node collects information about the FPGA nodes in the cluster, including computing power, current load, and communication bandwidth.

[0171] B7: Select an FPGA node using a weighted round-robin algorithm based on information about the equipped FPGA nodes. The weighted round-robin algorithm is a prior art in this field and does not constitute an inventive solution of the present application, and is not described in detail here.

[0172] B8: Encapsulate the data subset to be distributed and transmit it to the FPGA node selected in B7 through the network.

[0173] Example 3

[0174] See also Figure 4 In this embodiment, each node in the drone cluster obtains the aerial image data assigned to it from the data cache module, uses the improved intelligent collaborative recognition algorithm in the data recognition module to jointly identify the target in the aerial image, and marks the target recognition result and stores it in the global cache unit, including:

[0175] C1: Each node in the drone cluster is located in the distributed index system of the data cache module based on the parsed two-layer task allocation information. The specific process of positioning includes:

[0176] 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 based on the task identifier and data characteristics in the request;

[0177] C2: Based on the physical address returned by the index server, the node obtains the aerial image data assigned to it from the data cache module through the distributed file system and performs preprocessing.

[0178] C3: Loading the improved intelligent collaborative recognition algorithm model. Each node inputs the pre-processed 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.

[0179] C4: After target recognition is completed, each node marks the target recognition result; the marking is to draw a bounding box on the original aerial image, annotate the target category and confidence level. The confidence level calculation formula is the existing technical content in this field and does not constitute the inventive solution of this application, so it is not detailed here;

[0180] Furthermore, the specific steps of C4 include:

[0181] (1) Each node extracts target information from the output of the target recognition algorithm. The target information includes the target category, the coordinates of the bounding box, and the confidence score. The coordinates of the bounding box refer to the pixel coordinates of the upper left corner and the lower right corner.

[0182] (2) Load the corresponding original aerial image from local storage or cache, ensuring that the resolution and format of the image are consistent with those used during recognition;

[0183] (3) defining the style of the bounding box, such as the color and thickness of the lines, and using the image processing library OpenCV to draw a rectangular box on the original aerial image according to the extracted bounding box coordinates. OpenCV is a prior art in this field and is not an inventive solution of this application, and is not described in detail here;

[0184] (4) Annotate from above the bounding box to avoid occluding the target and the bounding box;

[0185] (5) Use the image processing library to add text information of the target category and confidence level at the determined location. At the same time, define the font, size, and color of the text to ensure that the annotation is clear and readable;

[0186] (6) Determine the save path and file name of the marked image, and use the image processing library to save the marked image.

[0187] C5: Each node encapsulates the marked target recognition results and stores the encapsulated target recognition results into the global cache through the network according to the storage location determined by the two-layer task allocation information and the distributed index system of the global cache.

[0188] The specific steps of C3 include:

[0189] C3.1: Each node loads the pre-trained improved intelligent collaborative recognition algorithm model into local storage;

[0190] C3.2: Each node inputs the preprocessed aerial imagery into the improved intelligent collaborative recognition algorithm model, performs convolution operations on the aerial imagery through convolutional layers, and extracts multi-scale feature maps. In this convolution operation, the first three layers of the convolutional layer use separable convolutions, and the last two layers of the convolutional layer use dilated convolutions.

[0191] C3.3: Weighted processing of multi-scale feature maps is performed through the attention mechanism to obtain weighted feature maps;

[0192] C3.4: Input the weighted feature map into a region proposal network, generate candidate regions containing the target using a sliding window, and perform classification and bounding box regression on the candidate regions to obtain target recognition results. The target recognition results include target category, location coordinates, and confidence level. The region proposal network is prior art in this field and does not constitute an inventive solution of this application, so it is not described in detail here.

[0193] Furthermore, the specific steps of C3.4 include:

[0194] (1) Input the weighted feature map into the region proposal network;

[0195] (2) Determine the size and step size of the sliding window, where the window size is selected based on the possible size of the target, and the step size determines the sliding interval of the window on the feature map;

[0196] (3) Use the defined sliding window to traverse the weighted feature map and process each window position;

[0197] (4) Preset a set of anchor boxes of different scales and aspect ratios for each window position;

[0198] (5) At each window position, generate the corresponding anchor frame according to the preset anchor frame parameters;

[0199] (6) For each generated anchor box, extract its corresponding feature vector from the weighted feature map;

[0200] (7) Use the fully connected layer to classify and predict the extracted anchor box features to determine whether each anchor box contains the target and the category of the target. The present invention uses the Softmax function to output the probability of each category. At the same time, another fully connected layer is used to regress the position and size of the anchor box to obtain more accurate target bounding box coordinates. The bounding box regression usually predicts four offsets, namely the center coordinates and width and height offsets of the anchor box;

[0201] (8) According to the probability of classification prediction and the result of bounding box regression, a threshold is set. For example, anchor boxes with a classification probability lower than a preset threshold are filtered out.

[0202] (9) Perform non-maximum suppression on the remaining candidate regions to remove redundant candidate regions with high overlap, and retain the candidate region with the highest confidence as the final target recognition result.

[0203] C3.5: Each node encapsulates the recognition result into structured data, encrypts it, and sends it to three neighboring nodes within the communication radius. The neighboring nodes then perform an intersection-and-comparison calculation on the received target recognition result with their local target recognition result.

[0204] If the intersection-over-union ratio 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, the local target recognition result of the adjacent node is replaced with the received target recognition result;

[0205] In specific operations, each node locally maintains a data structure for storing recognition results, such as a list, in which each element corresponds to a detected target, including target category, bounding box coordinates, and confidence information. When the recognition result of the external node meets the conditions that the intersection ratio is greater than 0.5 and the confidence is higher, this node will find the corresponding target element in the local data structure whose intersection ratio with the 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, a target in the local result is a car, the bounding box coordinates are (10, 10, 50, 50), and the confidence is 0.6, while the corresponding target in the external result is a car, the bounding box coordinates are (12, 12, 52, 52), the confidence is 0.8, and the intersection ratio of the two is greater than 0.5. At this time, all information about the target in the local data structure will be replaced by the information of the external result, where the external result refers to the received target recognition result.

[0206] If the intersection-over-union ratio is less than 0.3, the received target recognition result is retained as the new detection result;

[0207] In specific operations, when the intersection-and-union ratio of the external result and any local recognition result is less than 0.3, it means that the target detected by the external result is not fully covered or repeatedly detected in the local result. Therefore, the external result is directly added to the recognition result data structure maintained by this node. For example, two targets are detected in the local result, and a new target detection result is transmitted from the external node, and its intersection-and-union ratio with the two local targets is less than 0.3, then this new external result will be added as a new element to the data structure of the local recognition result and become one of the new target recognition results of this node.

[0208] C3.6: Output the final target recognition result set.

[0209] The construction process of the improved intelligent collaborative recognition algorithm model pre-trained in C3.1 includes:

[0210] C3.11: Define the model's application scenarios and collect data for these scenarios. For example, if the model is used for aerial imagery object recognition, it is necessary to collect aerial images from different regions, weather conditions, and times. The data for the model's application scenarios can come from public datasets, actual project data, or self-collection.

[0211] C3.12: Use the LabelImg annotation tool to accurately label the collected model application scenario data. The annotation content should be determined based on the specific task. The LabelImg annotation tool is prior art in this field and does not constitute an inventive solution of this application. A detailed description of this tool is omitted here.

[0212] C3.13: Load a deep convolutional neural network model architecture and perform model testing;

[0213] C3.14: Introducing an attention mechanism into the tested deep convolutional neural network model architecture and fusing feature maps from different layers with a multi-scale feature fusion approach enables the model to simultaneously capture both global and local features of the target. Furthermore, using separable convolution techniques reduces the number of model parameters and computational complexity, thereby improving the model's operational efficiency. Separable convolution techniques are prior art in this field and do not constitute an inventive solution of this application, so they are not described in detail here.

[0214] C3.15: Design information exchange methods between nodes, such as data transmission via wireless communication networks, 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 collaborative strategies to form improved intelligent collaborative recognition algorithm models. The collaborative strategy adopts a weighted voting method. For example, when multiple nodes identify the same target, the final recognition result is determined by weighted voting, where the weights are determined based on the node's confidence and computing power. The weighted voting method is the prior art content in this field and does not constitute the inventive solution of this application. It is not described in detail here.

[0215] C3.16: Divide the preprocessed aerial image data into a training set, a validation set, and a test set in a ratio of 7:1:2. At the same time, select a cross-entropy loss function and a stochastic gradient descent optimizer based on the task type. The cross-entropy loss function and the stochastic gradient descent optimizer are prior art in this field and do not constitute the inventive solution of this application, so they are not described here in detail.

[0216] C3.17: Use the training set to train the improved intelligent collaborative recognition algorithm model. During the training process, continuously adjust the parameters of the improved intelligent collaborative recognition algorithm model to gradually reduce the loss function, thereby obtaining a trained improved intelligent collaborative recognition algorithm model.

[0217] The specific steps of C3.3 include:

[0218] C3.31: Arrange the multi-scale feature maps obtained in C3.2 according to the channel dimension and perform global average pooling and global maximum pooling operations on each feature map to obtain the global information feature vector G;

[0219] The global average pooling operation is implemented by calculating the average value of all pixels in each channel;

[0220] The global maximum pooling operation is implemented by extracting the maximum value of all pixels in each channel;

[0221] The channel dimension is [batch_size, channels, height, width], where batch_size represents the number of samples processed at a 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;

[0222] C3.32: Input the global information feature vector into the shared multi-layer perceptron to obtain the channel attention weight matrix;

[0223] Furthermore, the specific steps of C3.32 include:

[0224] (1) Receive a global information feature vector G with a channel dimension of batch_size×2×channels;

[0225] (2) Based on the global information feature vector G, through the weight matrix and bias Perform linear transformation and then activate with ReLU to get the output result ;

[0226] (3) Through the weight matrix and bias Perform a quadratic linear transformation and use the Sigmoid activation function to obtain the channel attention weight matrix .

[0227] C3.33: Perform channel average pooling and channel maximum pooling operations on the multi-scale feature map in the channel dimension to generate a spatial attention weight matrix;

[0228] The channel average pooling is achieved by calculating the average value of all channels at each position;

[0229] The channel maximum pooling operation is implemented by calculating the maximum value of all channels at each position;

[0230] The spatial attention weight matrix is ​​generated 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, performing feature fusion through a 7×7 convolution layer, and then performing a Sigmoid activation function.

[0231] C3.34: Multiply the channel attention weight matrix and the spatial attention weight matrix element by element to obtain the attention weight matrix;

[0232] C3.35: Multiply the attention weight matrix by the multi-scale feature map in C3.2 element-wise to obtain the weighted feature map.

[0233] The task allocation unit uses a weighted minimum connection number algorithm, combined with the load weight of each node in the drone cluster monitored in real time by the load monitoring unit, to allocate the newly arrived data processing tasks, including:

[0234] D1: Create an information record for each node in the drone cluster; the information record contains the node's identity, current number of connections, and load weight information, where the load weight is determined based on factors such as the node's computing power, memory size, and battery power;

[0235] D2: Start the load monitoring unit to monitor the load of each node in real time and update the load weight of each node;

[0236] D3: After receiving the new data processing task, the task allocation unit obtains the current number of connections and load weight of each node from the load monitoring unit;

[0237] D4: The task allocation unit calculates the priority of each node using a weighted minimum connection number algorithm. The weighted minimum connection number algorithm is implemented by calculating the ratio of the current number of connections of node i to the load weight of node i.

[0238] D5: Based on the calculated node priority, select the node with the highest priority to process the new task, and update the connection count of the selected node by adding 1;

[0239] 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 of each node in real time and dynamically updates the load weight of the node;

[0240] D7: When a node completes a task, it reduces its connection count by 1 and waits for new tasks to arrive.

[0241] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose and scope of protection of the present invention. These are all protected by the present invention.

Claims

1. A distributed intelligent recognition and processing system for drone cluster collaborative aerial images, characterized by: include: Data acquisition module, data cache module, data identification module, 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 image data collected by the data acquisition module in the local edge cache unit into two-layer tasks according to the split feature type, and upload the split subtasks and metadata to the regional cache unit; Each node in the drone cluster obtains the aerial image data assigned to it from the data cache module, and uses the improved intelligent collaborative recognition algorithm in the data recognition module to jointly identify the target in the aerial image. The target recognition results are marked and stored in the global cache unit. The improved intelligent collaborative recognition algorithm is configured based on deep convolutional neural networks, attention mechanisms, and multi-scale feature fusion, specifically including: C1: Each node in the drone cluster is located in the distributed index system of the data cache module based on the parsed two-layer task allocation information. The specific process of positioning 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 based on the task identifier and data characteristics in the request; C2: Based on the physical address returned by the index server, the node obtains the aerial image data assigned to it from the data cache module through the distributed file system and performs preprocessing. C3: Load the improved intelligent collaborative recognition algorithm model. Each node inputs the pre-processed aerial image into the improved intelligent collaborative recognition algorithm model for target recognition. C4: After target recognition is completed, each node marks the target recognition result; the marking is to draw a bounding box on the original aerial image and mark the target category and confidence level; C5: Each node encapsulates the marked target recognition results and stores them in the global cache through the network according to the storage location determined by the two-layer task allocation information and the distributed index system of the global cache; The task allocation unit uses a weighted minimum connection number algorithm, combined with the load weight of each node in the drone cluster monitored in real time by the load monitoring unit, to allocate 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; The two-tiered tasks include: Perform color feature splitting on the aerial image data in the local edge cache unit, and distribute the hyperspectral data obtained after color feature splitting to the GPU-equipped nodes in the cluster; Perform texture feature segmentation on the aerial image data in the local edge cache unit, and distribute the complex texture data obtained after the texture feature segmentation to the node equipped with FPGA; The metadata includes grid ID, feature type, and priority.

2. The distributed intelligent recognition and processing system for drone cluster collaborative aerial images according to claim 1, characterized in that: The method of performing color feature splitting on the aerial image data in the local edge cache unit and distributing the hyperspectral data obtained after the color feature splitting to the GPU-equipped nodes in the cluster includes: A1: After receiving the read instruction from the system, the local edge cache unit extracts the aerial image data from the local edge cache and performs a preliminary check on the extracted aerial image data; If the extracted aerial image data is found to have 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 pre-processed 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, separate the pixel data in the aerial image that meets the color feature, and obtain the split hyperspectral data; A4: The central node collects information about GPU-equipped nodes in the cluster, calculates the comprehensive score of each node, and selects nodes based on the comprehensive score. A5: Encapsulate the split hyperspectral data and transmit it to the node selected in A4 through the network.

3. The distributed intelligent recognition and processing system for drone cluster collaborative aerial images according to claim 2, characterized in that: The method of performing texture feature splitting on the aerial image data in the local edge cache unit and distributing the texture complex data obtained after the texture feature splitting to the node equipped with the FPGA 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: Calculating the spatial relationship between different gray levels in the aerial image using the gray level co-occurrence matrix to obtain texture feature parameters; the texture feature parameters include contrast, correlation, energy, and homogeneity; B3: Based on the obtained texture features, the complexity of the texture is evaluated by contrast calculation; B4: Set the texture complexity threshold to classify the pixels in the aerial image into two categories: complex texture and simple texture; B5: Based on the texture complexity evaluation results, the aerial image data is split into texture-complex data and texture-simple data according to a preset texture complexity threshold, and the texture-complex data is extracted separately to form a data subset to be allocated; B6: The central node collects information about FPGA-equipped nodes in the cluster; B7: Based on the information of the equipped FPGA nodes, a weighted round-robin algorithm is used to select the FPGA node; B8: Encapsulate the data subset to be distributed and transmit it to the FPGA node selected in B7 through the network.

4. The distributed intelligent recognition and processing system for drone cluster collaborative aerial images according to claim 3, characterized in that: The specific steps of C3 include: C3.1: Each node loads the improved intelligent collaborative recognition algorithm model into local storage; C3.2: Each node inputs the preprocessed aerial imagery into the improved intelligent collaborative recognition algorithm model, performs convolution operations on the aerial imagery through convolutional layers, and extracts multi-scale feature maps. In this convolution operation, the first three layers of the convolutional layer use separable convolutions, and the last two layers of the convolutional layer use dilated convolutions. C3.3: Weighted processing of multi-scale feature maps is performed through the attention mechanism to obtain weighted feature maps; C3.4: Input the weighted feature map 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 target recognition results; the target recognition results include target category, location coordinates, and confidence level; C3.5: Each node encapsulates the recognition result into structured data, encrypts it, and sends it to three neighboring nodes within the communication radius. The neighboring nodes then perform an intersection-and-comparison calculation on the received target recognition result with their local target recognition result. If the intersection-over-union ratio 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, the local target recognition result of the adjacent node is replaced with the received target recognition result; If the intersection-over-union ratio is less than 0.3, the received target recognition result is retained as the new detection result; C3.6: Output the final target recognition result set.

5. The distributed intelligent recognition and processing system for drone cluster collaborative aerial images according to claim 4, characterized in that: The construction process of the improved intelligent collaborative recognition algorithm model pre-trained in C3.1 includes: C3.11: Define the application scenarios of the model and collect data on the application scenarios of the model; C3.12: Use the LabelImg annotation tool to accurately label the collected model application scenario data; C3.13: Load and test the deep convolutional neural network model architecture; C3.14: Introduce an attention mechanism into a tested deep convolutional neural network model architecture, combining multi-scale feature fusion and separable convolution techniques to fuse feature maps from different layers. C3.15: Design information interaction and communication methods between nodes, formulate collaborative strategies, and form an improved intelligent collaborative recognition algorithm model; C3.16: Divide the preprocessed aerial image data into training, validation, and test sets, and select the cross-entropy loss function and stochastic gradient descent optimizer based on the task type. C3.17: Use the training set to train the improved intelligent collaborative recognition algorithm model to obtain a trained improved intelligent collaborative recognition algorithm model.

6. The distributed intelligent recognition and processing system for drone cluster collaborative aerial images according to claim 5, characterized in that: The specific steps of C3.3 include: C3.31: Arrange the multi-scale feature maps obtained in C3.2 according to the channel dimension and perform global average pooling and global maximum pooling operations on each feature map to obtain the global information feature vector; The global average pooling operation is implemented by calculating the average value of all pixels in each channel; The global maximum 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 a 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 maximum pooling operations on the multi-scale feature map in the channel dimension to generate a spatial attention weight matrix; The channel average pooling is achieved by calculating the average value of all channels at each position; The channel maximum 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 by the multi-scale feature map in C3.2 element-wise to obtain the weighted feature map.

7. The distributed intelligent recognition and processing system for drone cluster collaborative aerial images according to claim 6, characterized in that: The task allocation unit uses a weighted minimum connection number algorithm, combined with the load weight of each node in the drone 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 drone cluster; the information record includes the node's identification, current number of connections, and load weight information; D2: Start the load monitoring unit to monitor the load of each node in real time and update the load weight of each node; D3: After receiving the new data processing task, the task allocation unit obtains the current number of connections and load weight of each node from the load monitoring unit; D4: The task allocation unit calculates the priority of each node using the weighted minimum connection number algorithm; the weighted minimum connection number algorithm is to calculate the priority of each node. i The current number of connections and nodes i The ratio of the load weight is achieved; D5: Based on the calculated node priority, select the node with the highest priority to process the new task, and update the connection count of the selected node by adding 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 of each node in real time and dynamically updates the load weight of the node; D7: When a node completes a task, it reduces its connection count by 1 and waits for new tasks to arrive.

Images