An open set recognition and fine classification combined low-altitude target recognition system
By combining multimodal perception and hierarchical recognition neural networks, the problems of identifying unknown targets and classifying known targets in low-altitude target recognition systems are solved, enabling accurate identification and threat assessment of low-altitude targets and improving the system's real-time performance and fault tolerance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- PEKING UNIV SHENZHEN GRADUATE SCHOOL
- Filing Date
- 2026-02-02
- Publication Date
- 2026-05-29
AI Technical Summary
Existing low-altitude target identification systems are mostly based on closed-set conditions, which cannot identify unknown target categories. Furthermore, the classification granularity for known categories is too coarse, making it impossible to accurately identify and assess threats.
A multimodal perception module is used to acquire multimodal data of the target. Combined with a hierarchical open-set refined recognition neural network model, features are extracted through a shared backbone network. Recognition is performed using an open-set recognition head and a refined classification head. Threat assessment is then conducted through a data fusion and decision-making alarm module.
It enables proactive identification of unknown targets and accurate model determination of known targets, improving the system's identification accuracy and threat assessment capabilities. It also possesses fault tolerance and real-time processing capabilities, meeting the real-time requirements of low-altitude security.
Smart Images
Figure CN122116267A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and security monitoring technology, specifically a low-altitude target recognition system that combines open set recognition with refined classification. Background Technology
[0002] In recent years, with the rapid development of drone technology and the dramatic decrease in cost, various low-altitude aircraft, represented by low-altitude, slow-moving, and small targets, have experienced explosive growth globally. While these aircraft bring convenience to society, they also pose unprecedented challenges to the security of critical infrastructure, large-scale event venues, border areas, and military restricted zones. Their characteristics, such as small size, low radar cross-section, low flight altitude, and high maneuverability, make them difficult for traditional air defense early warning systems to effectively detect and identify.
[0003] Existing low-altitude target recognition systems typically rely on multi-sensor fusion technology to acquire the target's physical and motion characteristics. After acquiring the target data, the system usually employs a deep learning-based classification model to identify the target. However, this traditional recognition method still suffers from the following problems:
[0004] First, existing systems are mostly based on closed set conditions and can only identify predefined categories in the training set. When faced with a constant stream of new, modified, or homemade aircraft and other unknown categories of targets, the systems usually cannot accurately identify them, which poses a safety hazard.
[0005] Secondly, for known target categories, there are multiple models of existing aircraft, and different sub-models vary greatly in performance, payload, and threat level. The classification granularity of existing systems is too coarse to accurately identify and assess threats. Summary of the Invention
[0006] The purpose of this invention is to provide a low-altitude target recognition system that combines open set recognition with fine classification, so as to solve the problem of closed set recognition and coarse classification granularity with low accuracy in the traditional recognition method mentioned in the background art.
[0007] To this end, the present invention provides a low-altitude target identification system that combines open set recognition and refined classification, comprising: a multimodal sensing module for detecting low-altitude targets and acquiring their multimodal data, wherein the multimodal data includes at least the target's trajectory information and optical image;
[0008] The data fusion module is electrically connected to the multimodal sensing module and is used to receive and process multimodal data;
[0009] The data processing and recognition module is used to receive the processed multimodal data and run a hierarchical open-set refined recognition neural network model.
[0010] The decision alarm and display module is electrically connected to the data processing and recognition module, and is used to receive and display the recognition results, and to issue alarms according to preset rules.
[0011] Preferably, the multimodal sensing module includes a three-coordinate phased array radar and a set of optoelectronic servo turntables. The three-coordinate phased array radar is used to perform 360° full-space scanning. When an object that conforms to the motion characteristics of a low, slow and small target is detected, track data is generated.
[0012] The optoelectronic servo turntable includes a high-resolution visible light camera, a long-wave infrared thermal imager, and a high-precision servo turntable. The high-resolution visible light camera is used to provide visual details of color, texture, and shape in daylight or under good lighting conditions. The long-wave infrared thermal imager is used to sense the thermal radiation characteristics of the target. The high-precision servo turntable is used to carry the optical payload and complete fast, smooth, and accurate pointing and tracking.
[0013] Preferably, the specific data processing flow of the data fusion module is as follows:
[0014] Time unification: The three-coordinate phased array radar, optoelectronic servo turntable, and data processing and recognition module are connected to the same high-precision clock source, and all data are stamped with a unified timestamp;
[0015] Coordinate System 1: The polar coordinates measured by the three-coordinate phased array radar and the image pixel coordinates of the photoelectric sensor are uniformly transformed into a geodetic coordinate system centered on the deployment point;
[0016] Radar track-guided optoelectronic servo turntable: After the three-coordinate phased array radar detects the target and establishes a stable track, the data processing and recognition module calculates the azimuth and elevation angle commands required by the optoelectronic servo turntable based on the real-time three-dimensional coordinates of the target provided by the radar. The servo turntable executes the commands to align the visible light camera and long-wave infrared thermal imager with the predicted target position. In the image, the target block corresponding to the radar track is found through moving target detection.
[0017] Track image fusion: Establish a joint state vector that includes target position, velocity, acceleration, and possible appearance features;
[0018] Correlation output: Outputs a timestamp, a unique target ID, the fused 3D position and velocity, and a data packet of visible light and infrared image slices obtained from the photoelectric system that strictly corresponds to the ID and target position.
[0019] Preferably, the hierarchical open-set refined recognition neural network model includes a shared backbone network, an open-set recognition head, and a refined classification head;
[0020] The shared backbone network uses a deep network of Vision Transformer as a feature extractor to collect the preprocessed target image, extract high-dimensional feature maps containing rich semantic information, and output the final global feature vector z for open set judgment and intermediate layer feature map F for fine classification.
[0021] The open-set recognition head is used to receive the output global feature vector z and calculate the probability that a sample belongs to a known or unknown entity.
[0022] The fine-grained classification head receives the output intermediate layer feature map F, whose activation is controlled by the open-set recognition head, which only performs computations when the target is judged to be known.
[0023] Preferably, the open set identification head includes a prototype matching unit, an energy calculation unit, and a probability decision unit;
[0024] The prototype matching unit is used to calculate the weighted distance between the feature vector of the target and the pre-stored class prototypes of multiple known categories, wherein the class prototypes represent the center of the corresponding known category in the feature space;
[0025] An energy calculation unit is connected to the prototype matching unit and is used to calculate the energy score of the target based on the weighted distance, wherein the energy score represents the degree to which the target deviates from the overall distribution of all known categories;
[0026] A probability decision unit, connected to the energy calculation unit, internally stores a probability model established for each known category based on extreme value theory; the probability decision unit is configured to: input the energy score into the probability model of each known category, calculate the membership probability of the target belonging to each known category respectively; select the maximum membership probability and compare it with a preset confidence threshold;
[0027] If the maximum membership probability is less than the confidence threshold, the target is determined to be an unknown category. If the maximum membership probability is greater than or equal to the confidence threshold, the target is determined to be a known category and is classified into the known category corresponding to the maximum membership probability.
[0028] Preferably, the training steps for the open set recognition head are as follows:
[0029] Construct a training set containing feature vectors extracted from all samples belonging to the known major classes using a shared backbone network. and their corresponding category tags ;
[0030] For each known class c, calculate the mean of the feature vectors of all samples in that class as the class prototype vector. The calculation expression is: In the formula This represents the number of samples; then, the covariance matrix of the eigenvectors is calculated, expressed as follows: ,storage and As model parameters;
[0031] Using an independent validation set, we pass it through a pre-trained backbone network to obtain a feature vector set. For each validation sample's feature vector, we calculate its squared Mahalanobis distance and energy score with each type of prototype. The expression for calculating the squared Mahalanobis distance is as follows: , The formula for calculating energy fraction is as follows: , In the formula, K represents the number of known major categories. The non-normalized similarity of sample z belonging to class c is given by the fact that the closer the sample is, the larger the value. This indicates the degree of anomaly of the sample; the higher the value, the more unknown the sample.
[0032] After training, the energy scores of the known class samples in the validation set are calculated, and the distribution of these scores is fitted with a Weiber distribution to construct a cumulative distribution function for each known class.
[0033] Preferably, the refined classification head includes a condition control unit, a feature map input interface, a component-level attention module, a feature pooling and fusion unit, a subclass classifier, a loss calculation unit, and an output interface module;
[0034] The condition control unit receives decision signals from the open set recognition head. It activates the refined classification head only when the open set recognition head determines that the target is of a known category; otherwise, the module does not work.
[0035] The feature map input interface receives intermediate layer feature maps F from the shared backbone network;
[0036] The component-level attention module is used to focus on local regions in the target that are crucial to the molecular category of the region, enhance the feature expression of these regions, and suppress irrelevant background;
[0037] The feature pooling and fusion unit is used to aggregate the weighted feature maps of the component-level attention module into a global feature vector for subsequent classification;
[0038] The subclass classifier is used to map fine-grained feature vectors to probability distributions of subclasses;
[0039] The loss calculation unit calculates the refined classification loss, which is used to train the parameters of the refined classification head;
[0040] The output interface module is used to integrate information and generate fine-grained recognition results that can be used by the system.
[0041] Preferably, the refined classification loss includes:
[0042] Coarse-grained classification loss uses the standard cross-entropy loss function for basic classification tasks with known major classes. The calculation expression is as follows: In the formula Given the number of samples, For real category tags, This represents the model's predicted probability for the major category;
[0043] Open set space loss is used to pull the features of samples from known classes toward their corresponding class prototypes, while pushing the features of samples from unknown classes away from all known class prototypes, ensuring that there is a minimum margin m between them. The calculation expression is as follows: , In the formula Indicates Mahalanobis distance;
[0044] Refined classification loss is used to optimize the discriminative power between subclasses, and its calculation expression is as follows: In the formula For anchor point samples, For positive samples of its subclass, For negative samples of different subclasses, To ensure that samples of the same subclass are closer in the feature space than samples of different subclasses;
[0045] Total loss is , , and This is a hyperparameter used to balance the importance of the three tasks.
[0046] Preferably, the decision alarm and display module includes:
[0047] A data receiving and parsing unit is used to stably and efficiently receive and parse real-time data streams from the data processing and recognition core module.
[0048] A situation fusion and display unit is used to fuse multi-source information and provide an intuitive, multi-dimensional comprehensive display through a graphical interface.
[0049] The intelligent alarm decision unit is used to perform real-time threat assessment on each target and automatically trigger graded alarm responses.
[0050] Human-computer interaction interface unit, which provides all controls and interfaces for bidirectional interaction between the operator and the system;
[0051] The data recording and playback unit records all system events for review, evidence collection, and system optimization.
[0052] Preferably, the intelligent alarm decision unit includes a configurable rule base, multi-level alarm triggers, and alarm action executors;
[0053] The configurable rule base is used to define alarm conditions through a natural language-like or graphical rule editor;
[0054] The multi-level alarm trigger is used to receive target information in real time, match it with the rule base, and output alarm instructions and levels;
[0055] The alarm action executor executes preset actions according to the alarm level. For a level 1 alarm, the interface flashes in full screen, triggers a piercing alarm sound, automatically takes a screenshot and records video, sends a suppression command to the physical countermeasure system via API or message queue, and automatically reports intelligence to the superior command center. For a level 2 alarm, the target icon flashes, emits a prompt sound, and is pinned to the top of the situation summary bar. For a level 3 alarm, the target icon changes color and is recorded in the log.
[0056] The low-altitude target recognition system proposed in this invention combines open set recognition with refined classification, and its advantages are as follows:
[0057] This invention introduces a data processing and identification module with open set recognition capabilities, enabling the system to proactively discover and identify unknown targets. Through refined classification capabilities, it can not only identify the major categories of known targets but also accurately determine their specific models. The organic combination of these two capabilities allows the system to clearly distinguish low-altitude targets, laying a solid foundation for a comprehensive assessment of the threat situation.
[0058] The open-set identification head employs an adaptive decision-making mechanism based on Mahalanobis distance and extreme value theory. Its decision threshold originates from statistical modeling of the distribution of known class samples, rather than a fixed value. This allows the system to dynamically adjust the boundaries between known and unknown based on the data itself, exhibiting extremely high sensitivity to novel threats or disguised targets that are close to the known category but substantially different. Furthermore, the redundant design and data fusion of the multimodal sensing front-end ensure continuous detection capabilities in complex environments and under conditions of partial sensor failure. This gives the entire security system strong fault tolerance and future adaptability, enabling it to readily cope with new threats brought about by rapid technological iteration.
[0059] The refined classification head extracts subtle discriminative features of targets through a component-level attention mechanism, and can accurately output the subclass or model information of the target, so that operators or automated decision-making systems can judge the potential performance of the target, greatly improving the efficiency and compliance of the handling.
[0060] The system of this invention constructs open set recognition and fine classification heads in parallel on a shared backbone network, and performs end-to-end joint optimization through a composite loss function. This avoids the computational overhead and latency caused by multiple feature extractions and model switching in traditional serial schemes. Compared with multi-model stacking schemes, this architecture significantly reduces computational complexity and inference time while maintaining or even improving recognition accuracy. This makes it possible to achieve real-time, multi-target high-intelligence processing on edge computing devices, meeting the stringent real-time requirements of low-altitude security.
[0061] This invention not only overcomes the technical bottlenecks of existing technologies in the identification and fine-grained differentiation of unknown targets, but also, through systematic innovation, constructs a next-generation low-altitude intelligent security solution that is forward-looking, accurate, efficient, and adaptable. It truly realizes the leap from passive early warning to active intelligent defense, and has significant practical value and broad prospects for promotion and application. Attached Figure Description
[0062] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0063] Figure 1 This is an overall architecture diagram of the low-altitude target identification system of the present invention;
[0064] Figure 2 This is an architecture diagram of the decision alarm and display control module of the present invention. Detailed Implementation
[0065] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described in this specification are merely for explaining the invention and are not intended to limit the invention.
[0066] Example:
[0067] Please see Figure 1-2 This invention provides a low-altitude target recognition system that combines open set recognition with refined classification. The system hardware architecture includes: a multimodal perception module, a data processing and recognition core module, and a decision alarm and display control module.
[0068] The multimodal sensing module is deployed at a high point in the monitored area and consists of a Ku-band three-coordinate phased array radar and an optoelectronic servo turntable. The radar performs a 360° full-airspace scan. When it detects an object exhibiting low, slow, and small target motion characteristics, it immediately generates trajectory data and transmits the target's coordinates, velocity, and other information to the data processing and identification core module via an internal local area network using the TCP / IP protocol. In this embodiment, a mature commercial radar model is selected, with a detection range of 10km for targets with an RCS of 0.01㎡ and a data refresh rate of 20Hz. The optoelectronic servo turntable integrates a Sony FCB-EV9500L visible light camera module (4K resolution, 30x optical zoom) and a FLIR Boson 640 long-wave infrared thermal imager (640x512 resolution). The turntable receives guidance commands from the core module and precisely aligns the camera with the target indicated by the radar. The camera transmits H.265 encoded high-definition video streams to the core module in real time via the RTSP protocol.
[0069] The core data processing and recognition module is deployed in the monitoring center's server room, with a high-performance server at its heart. The server is equipped with an Intel Core i9-13900K CPU, 128GB of DDR5 memory, and two parallel NVIDIA RTX 4090 GPUs. Running on the server are the core software of this invention, including data receiving and preprocessing programs, a hierarchical open-set refined recognition neural network model, and communication services with the display and control module. The two GPUs are bridged via NVLink, supporting model parallelism and data parallelism to ensure real-time processing capabilities for multiple video streams.
[0070] The decision-making alarm and display module serves as the operator's workstation. Located in the security command center, a multi-screen display console is installed, including an 85-inch 4K main display screen for showing the overall GIS situation, and two 55-inch auxiliary displays for displaying real-time video of key targets, alarm lists, and system status, respectively. The operator workstation is equipped with a high-performance graphics workstation running custom-developed low-altitude intelligent defense command platform client software. Alarm output devices include an audible and visual alarm in the command room, integrated with the existing public address system and drone countermeasures equipment control interfaces.
[0071] The system's workflow is as follows: the radar detects the target and reports its trajectory; the core module receives the trajectory, calculates the target's bearing, and sends a follow-up command to the photoelectric turntable; the photoelectric turntable locks onto the target and sends the video stream to the core module; the core module decodes the video stream and performs target detection on each frame or key frame, cropping the image of the target area; the cropped target image is fed into a hierarchical open-set refined recognition neural network model deployed on the GPU for inference; the hierarchical open-set refined recognition neural network model outputs the recognition result (unknown or known major category plus known subclass) and confidence level; the core module binds the recognition result with the target's trajectory information and sends it to the display and control module; the display and control module updates the display on the interface and determines whether to trigger an alarm based on alarm rules.
[0072] The hierarchical open-set refined recognition neural network model is the core of this invention. Before system deployment, the model is specifically trained for common low-altitude targets in urban areas, constructing training sets for known and unknown target classes. The training set for known target classes is finely labeled (i.e., Category 1: Multi-rotor drones; subcategories: DJI Mavic 3 series, Air series, Matrice 300 / 350 series, Autel EVO II series, Anafi, etc., totaling 15 subcategories; Category 2: Birds; subcategories: pigeons, sparrows, seagulls, eagles (common), etc., 8 subcategories; Category 3: Fixed-wing model aircraft; subcategories: common training aircraft, aerobatic aircraft, etc., 5 subcategories, totaling approximately 250,000 collected and labeled images from local real-world footage and some open-source datasets). The unknown class for training includes approximately 50,000 images of kites, balloons, plastic bags, consumer-grade racing drones, etc., used to allow the model to learn open space.
[0073] The model training steps are as follows: Training is performed end-to-end. A batch of training images is input, mixing known and unknown class samples. The system performs forward propagation, extracting feature maps and the final feature vector z from the images through a shared backbone network. Then, the feature vector z is simultaneously fed into an open-set recognition head and a fine-grained classification head. The open-set recognition head calculates its Mahalanobis distance to each known major class prototype and outputs a coarse-grained classification (used for calculation). ) and the distance from the prototype (used for calculation) The attention module within the network acts on the feature maps of the backbone network to generate weighted features, which are then output as refined classifications by the classifier (used for computation). );
[0074] For all known class samples, calculate their coarse-grained classification loss. Calculate the open set space loss for all samples. For all known class samples, construct triples and calculate the refined classification loss. , According to the formula Calculate the total loss. , and Based on experience, the values can be set to 1.0, 0.5, and 0.8; calculate the gradient of the total loss with respect to the model parameters, and use the AdamW optimizer to update the model weights. After the model training is complete, deploy it to the core module for inference.
[0075] The system achieves all-weather, intelligent identification and response to low-altitude targets by organically combining multimodal perception, open set recognition, refined classification and intelligent alarm. It can proactively discover and identify unknown targets, and through refined classification capabilities, it can not only identify the general category of known targets, but also accurately determine their specific models. The organic combination of these two capabilities enables the system to clearly distinguish low-altitude targets, laying a solid foundation for a comprehensive assessment of the threat situation.
[0076] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A low-altitude target recognition system combining open set recognition and refined classification, characterized in that: include: A multimodal sensing module is used to detect low-altitude targets and acquire their multimodal data, wherein the multimodal data includes at least the target's trajectory information and optical image; The data fusion module, electrically connected to the multimodal sensing module, is used to receive and process multimodal data; The data processing and recognition module is used to receive the processed multimodal data and run a hierarchical open-set refined recognition neural network model. The decision alarm and display module is electrically connected to the data processing and recognition module, and is used to receive and display the recognition results, and to issue alarms according to preset rules.
2. The low-altitude target recognition system combining open set recognition and refined classification according to claim 1, characterized in that: The multimodal sensing module includes a three-coordinate phased array radar and a set of optoelectronic servo turntables. The three-coordinate phased array radar is used to perform 360° full-space scanning. When it detects an object that matches the motion characteristics of a low, slow and small target, it generates track data. The optoelectronic servo turntable includes a high-resolution visible light camera, a long-wave infrared thermal imager, and a high-precision servo turntable. The high-resolution visible light camera is used to provide visual details of color, texture, and shape in daylight or under good lighting conditions. The long-wave infrared thermal imager is used to sense the thermal radiation characteristics of the target. The high-precision servo turntable is used to carry the optical payload and complete fast, smooth, and accurate pointing and tracking.
3. The low-altitude target recognition system combining open set recognition and refined classification according to claim 2, characterized in that: The specific data processing flow of the data fusion module is as follows: Time unification: The three-coordinate phased array radar, optoelectronic servo turntable, and data processing and recognition module are connected to the same high-precision clock source, and all data are stamped with a unified timestamp; Coordinate System 1: The polar coordinates measured by the three-coordinate phased array radar and the image pixel coordinates of the photoelectric sensor are uniformly transformed into a geodetic coordinate system centered on the deployment point; Radar track-guided optoelectronic servo turntable: After the three-coordinate phased array radar detects the target and establishes a stable track, the data processing and recognition module calculates the azimuth and elevation angle commands required by the optoelectronic servo turntable based on the real-time three-dimensional coordinates of the target provided by the radar. The servo turntable executes the commands to align the visible light camera and long-wave infrared thermal imager with the predicted target position. In the image, the target block corresponding to the radar track is found through moving target detection. Track image fusion: Establish a joint state vector that includes target position, velocity, acceleration, and possible appearance features; Correlation output: Outputs a timestamp, a unique target ID, the fused 3D position and velocity, and a data packet of visible light and infrared image slices obtained from the photoelectric system that strictly corresponds to the ID and target position.
4. The low-altitude target recognition system combining open set recognition and refined classification according to claim 1, characterized in that: The hierarchical open-set refined recognition neural network model includes a shared backbone network, an open-set recognition head, and a refined classification head; The shared backbone network uses a deep network of Vision Transformer as a feature extractor to collect the preprocessed target image, extract high-dimensional feature maps containing rich semantic information, and output the final global feature vector z for open set judgment and intermediate layer feature map F for fine classification. The open-set recognition head is used to receive the output global feature vector z and calculate the probability that a sample belongs to a known or unknown entity. The fine-grained classification head receives the output intermediate layer feature map F, whose activation is controlled by the open-set recognition head, which only performs computations when the target is judged to be known.
5. A low-altitude target recognition system combining open set recognition and refined classification according to claim 4, characterized in that: The open set recognition head includes a prototype matching unit, an energy calculation unit, and a probability decision unit; The prototype matching unit is used to calculate the weighted distance between the feature vector of the target and the pre-stored class prototypes of multiple known categories, wherein the class prototypes represent the center of the corresponding known category in the feature space; An energy calculation unit is connected to the prototype matching unit and is used to calculate the energy score of the target based on the weighted distance, wherein the energy score represents the degree to which the target deviates from the overall distribution of all known categories; A probability decision unit, connected to the energy calculation unit, internally stores a probability model based on extreme value theory for each known category; The probability decision unit is configured to: input the energy score into the probability model of each known category, calculate the membership probability of the target belonging to each known category respectively; select the maximum membership probability and compare it with a preset confidence threshold. If the maximum membership probability is less than the confidence threshold, the target is determined to be an unknown category. If the maximum membership probability is greater than or equal to the confidence threshold, the target is determined to be a known category and is classified into the known category corresponding to the maximum membership probability.
6. The low-altitude target recognition system combining open set recognition and refined classification according to claim 5, characterized in that: The training steps for the open set recognition head are as follows: Construct a training set containing feature vectors extracted from all samples belonging to the known major classes using a shared backbone network. and their corresponding category tags ; For each known class c, calculate the mean of the feature vectors of all samples in that class as the class prototype vector. The calculation expression is: In the formula Indicates the number of samples; Then, the covariance matrix of the eigenvectors is calculated, and the calculation expression is: ,storage and As model parameters; Using an independent validation set, we pass it through a pre-trained backbone network to obtain a feature vector set. For each validation sample's feature vector, we calculate its squared Mahalanobis distance and energy score with each type of prototype. The expression for calculating the squared Mahalanobis distance is as follows: , The formula for calculating energy fraction is as follows: In the formula, K represents the number of known major categories. The non-normalized similarity of sample z belonging to class c is given by the fact that the closer the sample is, the larger the value. This indicates the degree of anomaly of the sample; the higher the value, the more unknown the sample. After training, the energy scores of the known class samples in the validation set are calculated, and the distribution of these scores is fitted with a Weiber distribution to construct a cumulative distribution function for each known class.
7. A low-altitude target recognition system combining open set recognition and refined classification according to claim 4, characterized in that: The refined classification head includes a condition control unit, a feature map input interface, a component-level attention module, a feature pooling and fusion unit, a subclass classifier, a loss calculation unit, and an output interface module; The condition control unit receives decision signals from the open set recognition head. It activates the refined classification head only when the open set recognition head determines that the target is of a known category; otherwise, the module does not work. The feature map input interface receives intermediate layer feature maps F from the shared backbone network; The component-level attention module is used to focus on local regions in the target that are crucial to the molecular category of the region, enhance the feature expression of these regions, and suppress irrelevant background; The feature pooling and fusion unit is used to aggregate the weighted feature maps of the component-level attention module into a global feature vector for subsequent classification; The subclass classifier is used to map fine-grained feature vectors to probability distributions of subclasses; The loss calculation unit calculates the refined classification loss, which is used to train the parameters of the refined classification head; The output interface module is used to integrate information and generate fine-grained recognition results that can be used by the system.
8. A low-altitude target recognition system combining open set recognition and refined classification according to claim 7, characterized in that: The refined classification loss includes: Coarse-grained classification loss uses the standard cross-entropy loss function for basic classification tasks with known major classes. The calculation expression is as follows: In the formula Given the number of samples, For real category tags, This represents the model's predicted probability for the major category; Open set space loss is used to pull the features of samples from known classes toward their corresponding class prototypes, while pushing the features of samples from unknown classes away from all known class prototypes, ensuring that there is a minimum margin m between them. The calculation expression is as follows: ， In the formula Indicates Mahalanobis distance; Refined classification loss is used to optimize the discriminative power between subclasses. The calculation expression is In the formula For anchor point samples, For positive samples of its subclass, For negative samples of different subclasses, To ensure that samples of the same subclass are closer in the feature space than samples of different subclasses; Total loss is , , and This is a hyperparameter used to balance the importance of the three tasks.
9. A low-altitude target recognition system combining open set recognition and refined classification according to claim 1, characterized in that: The decision alarm and display module includes: A data receiving and parsing unit is used to stably and efficiently receive and parse real-time data streams from the data processing and recognition core module; A situation fusion and display unit is used to fuse multi-source information and provide an intuitive, multi-dimensional comprehensive display through a graphical interface. The intelligent alarm decision unit is used to perform real-time threat assessment on each target and automatically trigger graded alarm responses. Human-computer interaction interface unit, which provides all controls and interfaces for bidirectional interaction between the operator and the system; The data recording and playback unit records all system events for review, evidence collection, and system optimization.
10. A low-altitude target recognition system combining open set recognition and refined classification according to claim 9, characterized in that: The intelligent alarm decision unit includes a configurable rule base, multi-level alarm triggers, and alarm action executors; The configurable rule base is used to define alarm conditions through a natural language-like or graphical rule editor; The multi-level alarm trigger is used to receive target information in real time, match it with the rule base, and output alarm instructions and levels; The alarm action executor executes preset actions according to the alarm level. For a level 1 alarm, the interface flashes in full screen, triggers a piercing alarm sound, automatically takes a screenshot and records video, sends a suppression command to the physical countermeasure system via API or message queue, and automatically reports intelligence to the superior command center. For a level 2 alarm, the target icon flashes, emits a prompt sound, and is pinned to the top of the situation summary bar. For a level 3 alarm, the target icon changes color and is recorded in the log.