Real-time Remote Sensing Open-World Target Detection Method Guided by Large Model Priors

Abstract

This invention relates to the field of remote sensing image processing, specifically to a real-time open-world target detection method based on large model prior guidance. The method includes: acquiring training remote sensing images; generating a set of pseudo-labels for unknown targets using SAM (Simultaneous Aspect-Oriented Model); inputting the training images into a target detection neural network to extract multi-scale features and obtain a dense candidate point set; performing spatial diversity queries on the dense candidate point set to filter out an initial target query set with spatial differences; performing one-to-many matching between the initial query set and known category labels and unknown target pseudo-labels, mapping queries matched to pseudo-labels to a virtual space for deduplication; updating network parameters through uncertainty-aware object-oriented learning; and inputting the image to be tested into the trained network to output the category labels and bounding box coordinates of each target. This invention effectively improves the recall rate of dense small targets, alleviates the problem of sparse supervision by pseudo-labels, and adaptively resists interference from complex background noise.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing, specifically to a real-time remote sensing open-world target detection method based on large model prior guidance. Background Technology

[0002] In recent years, high-resolution remote sensing image data has experienced explosive growth. Deep learning-based target detection techniques, especially those represented by convolutional neural networks and Transformers, have achieved significant success in identifying predefined categories of targets such as vehicles, ships, and buildings in remote sensing images. However, most existing detection paradigms strictly adhere to the closed-world assumption, which requires that the training and test sets contain exactly the same set of target categories.

[0003] In real-world, open, dynamic remote sensing scenarios, models inevitably encounter a large number of unlabeled or never-before-seen unknown objects (such as buildings damaged by sudden disasters, rare man-made objects, etc.). Traditional closed-set detectors lack the ability to handle unknown concepts, typically forcing these unknown objects into one of the known categories or simply ignoring them as background. This severely limits the robustness and reliability of models in practical deployments, spurring research into Open-World Object Detection (OWOD) technology.

[0004] The OWOD task requires the model to simultaneously achieve two conflicting objectives: first, to accurately identify targets of known categories; and second, to discover and label targets of unknown categories in order to continuously expand the knowledge base during subsequent incremental learning phases. Current mainstream OWOD methods, such as ORE (Open-World Object Detector) and OW-DETR (Open-World DETR), have shown potential in processing conventional images, but their direct application to remote sensing images suffers from the following significant limitations:

[0005] (I) Query Redundancy and Missed Detection of Dense Small Targets under Extreme Scale Span. Remote sensing images have extreme intra-class scale spans. An image may simultaneously contain giant targets such as airports and ports occupying large areas, as well as tiny targets such as vehicles and ships with only a few dozen pixels. Existing methods typically employ a Top-K query selection strategy based on classification confidence, which naturally favors large targets that occupy a large number of pixels and have strong feature responses. With a fixed query budget, a large amount of computational resources are occupied by redundant queries on large targets, causing the weak features of densely arranged small targets in space to be completely masked, resulting in serious missed detection problems.

[0006] (II) The Dilemma of Sparse Supervision and Interference from Complex Background Noise. To improve the recall rate of unknown targets, recent studies have attempted to introduce large-scale visual models such as SAM (Segment Anything Model) to generate dense pseudo-labels as geometric priors to guide detector training. However, existing solutions employ a traditional one-to-one hard matching mechanism, requiring each SAM pseudo-label to uniquely supervise a single detection query. Due to the dense and mutually occluded targets in remote sensing scenes, queries corresponding to a large number of visually real but tiny targets cannot obtain effective pseudo-label supervision, resulting in a sparse supervision problem. Simultaneously, remote sensing scenes contain complex background interference such as sea surface wakes, building shadows, and cloud occlusion. Existing methods rely on fixed geometric thresholds to filter pseudo-labels, easily leading to the erroneous rejection of irregularly shaped but real tiny unknown targets as noise, further exacerbating information loss.

[0007] In summary, how to effectively overcome the challenges of low recall rate of dense small targets, sparse supervision of false labels, and interference from complex background noise in remote sensing images are technical problems that urgently need to be solved in this field. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a remote sensing open-world target detection method based on large model prior guidance. It aims to solve the technical problems of low recall rate of dense small targets, sparse pseudo-label supervision, and poor filtering ability of existing open-world detection networks in remote sensing scenarios.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] This invention proposes a real-time remote sensing open-world target detection method based on large model prior guidance, characterized by the following steps:

[0011] S1. Obtain training remote sensing images and their corresponding known category label sets; input the training remote sensing images into a pre-trained SAM for dense grid sampling, generate multiple candidate masks, perform quality filtering on the candidate masks, and obtain an unknown target pseudo-label set;

[0012] S2. Input the training remote sensing image into the target detection neural network, extract multi-scale features through the backbone network, and input the extracted multi-scale features into the hybrid encoder for feature interaction to obtain a dense candidate point set;

[0013] S3. Perform spatial diversity query processing on the dense candidate point set, construct an adaptive repulsion field based on the scale prior of the candidate points, and use a greedy iterative sampling strategy to filter out an initial target query set with spatial differences from the dense candidate point set.

[0014] S4. Match the initial target query set with the known category label set to obtain the initial target query set matched with the known category label; match the initial target query set with the unknown target pseudo label set, and for each unknown target pseudo label, assign it to the initial target query that overlaps with its spatial position to obtain the initial target query set matched with the unknown target pseudo label; map the initial target queries matched with the unknown target pseudo label to non-overlapping virtual coordinate spaces, and perform non-maximum suppression deduplication to obtain the deduplicated query set;

[0015] S5. For initial target queries that are not assigned to known category labels and are not assigned to unknown target pseudo-labels, calculate negative sample soft weights based on their classification uncertainty; calculate negative sample targetability loss based on the negative sample soft weights; for initial target queries that match known category labels, calculate positive sample targetability loss, classification loss, and regression loss; for queries in the deduplicated query set, calculate spatial constraint loss; calculate comprehensive object-oriented learning loss using the positive sample targetability loss and negative sample targetability loss; sum the comprehensive object-oriented learning loss, classification loss, regression loss, and spatial constraint loss using weighted averages to obtain the total loss; update the weight parameters of the target detection neural network using gradient descent to obtain the trained target detection neural network.

[0016] S6. Input the remote sensing image to be tested into the trained target detection neural network, and output the category label and bounding box coordinates of each target.

[0017] Furthermore, in S3, the process of performing spatial diversity query processing on the dense candidate point set specifically includes:

[0018] S301. Obtain the dense candidate point set, which includes the confidence score, center coordinates, prediction width and prediction height of each candidate point.

[0019] S302. Calculate the adaptive rejection radius for each candidate point:

[0020]

[0021] In the formula, For the dense candidate point set, the first Candidate points The adaptive repulsion radius, To control the hyperparameters of the repulsive field strength, Candidate points The prediction width, Candidate points The predicted height;

[0022] S303, Initialize the selection set The dense candidate point set is iterated until a selection set is reached. Once the number of candidate points added reaches the preset value, an initial target query set with spatial differences is obtained;

[0023] In each iteration, the candidate point with the highest confidence score among the remaining candidate points is selected. Add to selection set ; Calculate each remaining candidate point With candidate points The square of the Euclidean distance between them ,in, For the remaining candidate points The center coordinates, Candidate points The center coordinates, Define the L2 norm; define the suppression mask. It will fall into the candidate point The confidence score of the remaining candidate points within the exclusion field is set to 0, where... Candidate points The adaptive repulsion radius, This is an indicator function.

[0024] Furthermore, in S4, the initial target queries that match the unknown target pseudo-labels are mapped to non-overlapping virtual coordinate spaces, and non-maximum suppression deduplication is performed, specifically including:

[0025] Initial target queries to match different unknown target pseudo-labels Apply virtual coordinate offsets respectively Then, it is mapped to a virtual coordinate space: ,in Query for the initial target The predicted bounding box, Initial target query in virtual coordinate space The bounding box, Query for the initial target Matching pseudo-tag index, A constant larger than the image size;

[0026] Perform nonmaximum suppression on all bounding boxes in the virtual coordinate space to obtain a deduplicated query set.

[0027] Furthermore, in S5, the process of calculating the negative sample targeting loss includes:

[0028] S501. For initial target queries that are not assigned to known category labels and not assigned to unknown target pseudo-labels. Its object response energy value Transformed into object probability through exponential mapping :

[0029]

[0030] In the formula, Temperature coefficient;

[0031] S502, Define the initial target query Classification uncertainty :

[0032]

[0033] In the formula, Given a known set of category labels, For activation function, Query for the initial target For known categories The predicted value;

[0034] S503, Introducing Initial Target Query Corresponding unknown category inhibitory factor :

[0035]

[0036] In the formula, Query for the initial target Predicted values ​​for unknown categories;

[0037] S504, Calculate the initial target query Negative sample soft weights :

[0038]

[0039] In the formula, The total number of initial target queries that were not assigned to known category labels and not assigned to unknown target pseudo-labels;

[0040] S505. Calculate the negative sample targeting loss based on the soft weights. :

[0041]

[0042] Furthermore, in S5, the formula for calculating the total loss is:

[0043]

[0044]

[0045] In the formula, For the total loss, For classifying losses, To regress the loss, For spatial constraint loss, To comprehensively address object-oriented learning loss, These are the weighting coefficients for the regression loss. These are the weighting coefficients for the spatial constraint loss. The weighting coefficients for the overall object-oriented learning loss, For positive sample targeting loss, For negative sample targeting loss.

[0046] Furthermore, the target detection neural network is a real-time target detector based on the RT-DETR (Real-Time Detection Transformer) architecture.

[0047] This invention also proposes a real-time remote sensing open-world target detection system based on large model prior guidance for implementing the above method, comprising:

[0048] The image acquisition and prior generation module is used to acquire training remote sensing images and their corresponding known category label sets, and to generate unknown target pseudo-label sets through pre-trained SAM;

[0049] The detection processing module includes a target detection neural network, which executes steps S2 to S5 to obtain a trained target detection neural network. The detection processing module is also used to execute step S6 to output the category label and bounding box coordinates of each target.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] (1) This invention uses spatial diversity query processing to construct an adaptive rejection field based on the scale prior of candidate points and effectively eliminates spatial redundancy by using greedy iterative sampling. This method makes the rejection range of large targets large and the rejection range of small targets small, thus reserving sufficient query quota for small targets, effectively solving the problem of query perception blind spots under extremely large scale spans in remote sensing scenes, and significantly improving the recall rate of dense small targets.

[0052] (2) This invention abandons the traditional exclusive single-point matching by using collaborative query allocation. It adopts a one-to-many allocation strategy, allowing the same pseudo-label to supervise multiple initial target queries that overlap with its spatial location. Through virtual coordinate space offset and non-maximum suppression operations, it maximizes the utilization of geometric priors and effectively alleviates the sparse supervision problem. At the same time, this invention uses an uncertainty-aware object learning strategy to mine the classification uncertainty within the model and dynamically allocates negative sample soft weights based on this, achieving adaptive suppression of complex background noise and effectively protecting real, small, and unknown targets with different shapes.

[0053] (3) This invention combines spatial diversity query processing, collaborative query allocation and uncertainty-aware object learning, achieving significant accuracy improvement in remote sensing open-world target detection tasks. It also has a stronger ability to retain old categories when learning new categories, effectively overcoming the forgetting problem. Attached Figure Description

[0054] Figure 1 The flowchart illustrates the real-time remote sensing open-world target detection method based on large model prior guidance provided by this invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1

[0057] refer to Figure 1 This embodiment provides a real-time remote sensing open-world target detection method based on large model prior guidance, which is performed according to the following steps:

[0058] S1. Obtain the training remote sensing images and their corresponding known category label sets. The known category label sets contain pre-labeled category labels, such as ships, vehicles, buildings, etc.

[0059] The trained remote sensing image is input into a pre-trained SAM for dense grid sampling. In this embodiment, the SAM adopts a full-image dense sampling mode, configuring a 32×32 dense grid sampling point array. The SAM uses each grid point as a cue point to segment the corresponding object region and generate multiple candidate masks.

[0060] To ensure the quality of the pseudo-labels, the candidate masks are subjected to quality filtering. Specifically, the predicted intersection-over-union (IoU) threshold and stability score threshold are both set to 0.95, and the minimum mask area is limited to 200 pixels to filter out background debris. The bounding boxes of the remaining candidate masks after filtering are used as pseudo-labels for unknown targets, resulting in a set of pseudo-labels for unknown targets. These pseudo-labels do not contain category information; they only indicate that an object may exist at that location and are used for the generation of subsequent supervision signals.

[0061] S2. The trained remote sensing image is input into the target detection neural network. The backbone network performs multi-level downsampling on the input image to extract multi-scale features. The extracted multi-scale features are then input into a hybrid encoder for feature interaction and fusion to obtain a dense candidate point set. The dense candidate point set contains each candidate point... ( ;in The confidence score, center coordinates, prediction width, and prediction height are calculated for the total number of candidate points in the dense candidate point set. In this embodiment, the target detection neural network adopts the RT-DETR architecture, and the backbone network of the target detection neural network adopts ResNet-50 (Residual Network 50).

[0062] S3. Perform spatial diversity query processing on the dense candidate point set, specifically including:

[0063] S301. Obtain the dense candidate point set.

[0064] S302. Calculate the adaptive rejection radius for each candidate point:

[0065]

[0066] In the formula, For dense candidate points, the candidate points are concentrated. The adaptive repulsion radius, To control the hyperparameters of the repulsive field strength, Candidate points The prediction width, Candidate points The predicted altitude.

[0067] In this embodiment, =0.25. This formula results in a large repulsion radius for large targets and a small repulsion radius for small targets, thereby achieving adaptive spatial suppression.

[0068] S303, Initialize the selection set The dense candidate point set is iteratively sampled using a greedy iterative sampling strategy until a selection set is reached. When the number of candidate points added reaches a preset value (100 in this embodiment), an initial target query set (query set A) with spatial differences is obtained.

[0069] In each iteration, select the candidate point with the highest confidence score from the remaining candidate points. Add to selection set ; Calculate each remaining candidate point ( ;in Candidate points Add to selection set (the total number of remaining candidate points) and candidate points The square of the Euclidean distance between them ,in, For the remaining candidate points The center coordinates, Candidate points The center coordinates, Define the L2 norm; define the remaining candidate points. Suppression mask It will fall into the candidate point The confidence score of the remaining candidate points within the exclusion field is set to 0, where... Candidate points The adaptive repulsion radius, This is an indicator function.

[0070] The query set A obtained in this step is evenly distributed in the image space, which avoids the problem of small targets being missed due to large targets occupying all query resources.

[0071] S4. Match query set A with the known category label set to obtain the initial target query set (query set B) that matches the known category labels. Simultaneously, match query set A with the unknown target pseudo-label set to obtain the initial target query set (query set C) that matches the unknown target pseudo-labels. For matching unknown target pseudo-labels, to address the issue of sparse pseudo-label supervision, a one-to-many allocation strategy is adopted: calculate the intersection-union ratio (IU) of each predicted bounding box in query set A with all pseudo-labels in the unknown target pseudo-label set, obtain the maximum IU value and the corresponding best-matching pseudo-label index, and select query sets with an IU value greater than a preset matching threshold. (In this embodiment) The initial target query (=0.5) is used to form a query set C. For each unknown target pseudo-label, it is simultaneously assigned to the initial target query that overlaps with its spatial location. Unlike the one-to-one hard matching mechanism in traditional methods, the one-to-many assignment strategy allows the same pseudo-label to supervise multiple queries with spatial complementarity, thereby alleviating the sparse supervision problem.

[0072] For query set C, further deduplication is needed. Specifically, this involves deduplicating the initial target queries in query set C that match different unknown target pseudo-labels. ( ;in For query set Apply virtual coordinate offsets to the total number of initial target queries. Then, it is mapped to a virtual coordinate space:

[0073]

[0074] In the formula, Query for the initial target The predicted bounding box, Initial target query in virtual coordinate space The bounding box, Query for the initial target Matching pseudo-tag index, A constant larger than the image size, for example =2000 pixels, which is greater than the image width and height (the image size in this embodiment is 1024×1024 pixels).

[0075] This step decouples queries that match different pseudo-labels into non-overlapping virtual coordinate spaces through mapping.

[0076] Perform nonmaximum suppression on all mapped bounding boxes in the virtual coordinate space to remove redundant queries and obtain a deduplicated query set (query set D).

[0077] S5. Calculate the total loss of the object detection neural network and update the weight parameters of the object detection neural network, specifically including the following sub-steps:

[0078] S501. For initial target queries that are not assigned to known category labels and not assigned to unknown target pseudo-labels. Perform the following operations:

[0079] Its object response energy value Transformed into object probability through exponential mapping :

[0080]

[0081] In the formula, This is a temperature coefficient used to adjust the scaling of energy values. The objectivity probability represents the likelihood that an object exists in the corresponding query region.

[0082] Define the initial target query Classification uncertainty :

[0083]

[0084] In the formula, Given a known set of category labels, For activation function, Query for the initial target For known categories Predicted values; classification uncertainty It reflects the confidence level of the model in predicting the known class: the closer the value is to 0, the more certain the model is in predicting the known class; the closer the value is to 1, the less certain the model is in predicting the known class.

[0085] Introducing the initial target query Corresponding unknown category inhibitory factor :

[0086]

[0087] In the formula, Query for the initial target Predicted values ​​for unknown categories; the unknown category suppression factor is used to reduce the negative sample penalty for targets predicted as unknown by the model: if the model believes that the query has a high probability of belonging to the unknown category, then... If the value is very small, the soft weight of the negative samples in this query is reduced accordingly.

[0088] Calculate the initial target query Negative sample soft weights :

[0089]

[0090] In the formula, The total number of initial target queries that are not assigned to known category labels and not assigned to unknown target pseudo-labels; negative sample soft weights give greater negative sample penalties to queries with high uncertainty and those not predicted as unknown.

[0091] Based on negative sample soft weights Calculate the negative sample targeting loss :

[0092]

[0093] S502. Calculate the positive sample targeting loss of query set B. Classification loss and regression loss Specifically, the loss value is calculated for each initial target query in query set B, and then the loss values ​​of all initial target queries are summed and averaged.

[0094] Positive sample target loss The calculation formula is:

[0095]

[0096] In the formula, This represents the total number of initial target queries in query set B. For the initial target query in query set B The objectivity probability is calculated using the same method as... .

[0097] In this embodiment, classification loss The regression loss is calculated using the Varifocal (zoom) loss function. The standard bounding box regression loss function is used for calculation.

[0098] S503. For the initial target query in query set D, calculate the spatial constraint loss. This loss is used to constrain the spatial consistency between queries that match unknown target pseudo-labels and the pseudo-labels themselves.

[0099] S504, using the positive sample targeting loss and negative sample target loss Calculate the comprehensive object-oriented learning loss :

[0100]

[0101] S505, the comprehensive object learning loss Classification loss Regression loss and spatial constraint loss We calculate the total loss by weighted summation. ;

[0102]

[0103] In the formula, These are the weighting coefficients for the regression loss. These are the weighting coefficients for the spatial constraint loss. These are the weighting coefficients for the overall object-oriented learning loss.

[0104] S506. The weight parameters of the target detection neural network are updated using the gradient descent algorithm through the AdamW optimizer to obtain the trained target detection neural network.

[0105] S6. Input the remote sensing image to be tested into the trained target detection neural network, and output the category label and bounding box coordinates of each target. For targets with known categories, output the specific category name (e.g., ships); for targets with unknown categories, output an unknown label.

[0106] To verify the effectiveness of this invention, the proposed open-world object detection method and existing models ORE, OW-DETR, and PROB (Probabilistic Objectness for Open World Object Detection) were incrementally learned and validated on the DOTA v1.5 remote sensing dataset. The DOTA v1.5 dataset uses an incremental learning evaluation protocol, dividing the categories into four tasks (Task 1 to Task 4), with the model learning new categories sequentially for each task. Evaluation metrics include: Unknown Class Recall (U-Recall), Known Class Mean Aptitude (mAP), Frames Per Second (FPS), and FLOPs (Functional Flow Rate). Table 1 shows the performance comparison of each model on the DOTA v1.5 dataset at a 1024×1024 resolution input.

[0107] Table 1. Comparison of known class mAP for each model on the DOTA v1.5 test set

[0108]

[0109] As shown in Table 1, the present invention significantly outperforms existing methods in mAP across all four tasks, particularly demonstrating a marked improvement in detection accuracy across all categories. Furthermore, the present invention exhibits stronger retention of older categories while learning new ones, effectively overcoming the forgetting problem.

[0110] Table 2 Comparison of Unknown Class Recall and Inference Efficiency of Each Model on the DOTA v1.5 Test Set

[0111]

[0112] As shown in Table 2, this invention significantly outperforms existing methods in terms of recall rate for unknown categories, demonstrating a strong ability to discover unknown targets. In terms of inference speed, this invention offers a significant improvement over existing methods. Regarding computational complexity, this invention has the lowest computational cost among all compared methods.

[0113] In summary, this invention effectively solves the technical problems of low recall rate of dense small targets, sparse supervision of false labels, and large interference from complex background noise in the prior art, and achieves significant improvement in real-time performance and accuracy in remote sensing open-world target detection tasks.

[0114] The specific embodiments of the present invention are provided to enable those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention.

[0115] It should be understood that the present invention is not limited to the content already described above, and various modifications and changes can be made without departing from its scope. The scope of the present invention is limited only by the appended claims.

Claims

1. A real-time remote sensing open world target detection method based on large model prior guidance, characterized in that, Includes the following steps: S1. Obtain training remote sensing images and their corresponding known category label sets; input the training remote sensing images into a pre-trained SAM for dense grid sampling, generate multiple candidate masks, perform quality filtering on the candidate masks, and obtain an unknown target pseudo-label set; S2. Input the training remote sensing image into the target detection neural network, extract multi-scale features through the backbone network, and input the extracted multi-scale features into the hybrid encoder for feature interaction to obtain a dense candidate point set; S3. Perform spatial diversity query processing on the dense candidate point set, construct an adaptive repulsion field based on the scale prior of the candidate points, and use a greedy iterative sampling strategy to filter out an initial target query set with spatial differences from the dense candidate point set. The process of performing spatial diversity query processing on the dense candidate point set specifically includes: S301. Obtain the dense candidate point set, which includes the confidence score, center coordinates, prediction width and prediction height of each candidate point. S302. Calculate the adaptive rejection radius for each candidate point: In the formula, For the dense candidate point set, the first Candidate points The adaptive repulsion radius, To control the hyperparameters of the repulsive field strength, Candidate points The prediction width, Candidate points The predicted height; S303, Initialize the selection set The dense candidate point set is iterated until a selection set is reached. Once the number of candidate points added reaches the preset value, an initial target query set with spatial differences is obtained; In each iteration, the candidate point with the highest confidence score among the remaining candidate points is selected. Add to selection set ; Calculate each remaining candidate point With candidate points The square of the Euclidean distance between them ,in, For the remaining candidate points The center coordinates, Candidate points The center coordinates, Define the L2 norm; define the suppression mask. It will fall into the candidate point The confidence score of the remaining candidate points within the exclusion field is set to 0, where... Candidate points The adaptive repulsion radius, For indicator functions; S4. Match the initial target query set with the known category label set to obtain the initial target query set matched with the known category label; match the initial target query set with the unknown target pseudo label set, and for each unknown target pseudo label, assign it to the initial target query that overlaps with its spatial position to obtain the initial target query set matched with the unknown target pseudo label; map the initial target queries matched with the unknown target pseudo label to non-overlapping virtual coordinate spaces, and perform non-maximum suppression deduplication to obtain the deduplicated query set; The process of obtaining the deduplicated query set specifically includes: Initial target queries to match different unknown target pseudo-labels Apply virtual coordinate offsets respectively Then, it is mapped to a virtual coordinate space: ,in Query for the initial target The predicted bounding box, Initial target query in virtual coordinate space The bounding box, Query for the initial target Matching pseudo-tag index, A constant larger than the image size; Perform nonmaximum suppression on all bounding boxes in the virtual coordinate space to obtain a deduplicated query set; S5. For initial target queries that are not assigned to known category labels and are not assigned to unknown target pseudo-labels, calculate negative sample soft weights based on their classification uncertainty; calculate negative sample targetability loss based on the negative sample soft weights; for initial target queries that match known category labels, calculate positive sample targetability loss, classification loss, and regression loss; for queries in the deduplicated query set, calculate spatial constraint loss; calculate comprehensive object-oriented learning loss using the positive sample targetability loss and negative sample targetability loss; sum the comprehensive object-oriented learning loss, classification loss, regression loss, and spatial constraint loss using weighted averages to obtain the total loss; update the weight parameters of the target detection neural network using gradient descent to obtain the trained target detection neural network. The process of calculating the negative sample targeting loss includes: S501. For initial target queries that are not assigned to known category labels and not assigned to unknown target pseudo-labels. Its object response energy value Transformed into object probability through exponential mapping : In the formula, Temperature coefficient; S502, Define the initial target query Classification uncertainty : In the formula, Given a known set of category labels, For activation function, Query for the initial target For known categories The predicted value; S503, Introducing Initial Target Query Corresponding unknown category inhibitory factor : In the formula, Query for the initial target Predicted values ​​for unknown categories; S504, Calculate the initial target query Negative sample soft weights : In the formula, The total number of initial target queries that were not assigned to known category labels and not assigned to unknown target pseudo-labels; S505, Based on Negative Sample Soft Weights Calculate the negative sample targeting loss : S6. Input the remote sensing image to be tested into the trained target detection neural network, and output the category label and bounding box coordinates of each target.

2. The real-time remote sensing open-world target detection method based on large model prior guidance as described in claim 1, characterized in that, In S5, the formula for calculating the total loss is: In the formula, For the total loss, For classifying losses, To regress the loss, For spatial constraint loss, To comprehensively address object-oriented learning loss, These are the weighting coefficients for the regression loss. These are the weighting coefficients for the spatial constraint loss. The weighting coefficients for the overall object-oriented learning loss, For positive sample targeting loss, For negative sample targeting loss.

3. The real-time remote sensing open-world target detection method based on large model prior guidance according to claim 1, characterized in that, The target detection neural network is a real-time target detector based on the RT-DETR architecture.

4. A real-time remote sensing open-world target detection system based on large model prior guidance for implementing the method of any one of claims 1-3, characterized in that, include: The image acquisition and prior generation module is used to acquire training remote sensing images and their corresponding known category label sets, and to generate unknown target pseudo-label sets through pre-trained SAM; The detection processing module includes a target detection neural network, which executes steps S2 to S5 to obtain a trained target detection neural network. The detection processing module is also used to execute step S6 to output the category label and bounding box coordinates of each target.

Images