Target detection method and device for complex remote sensing scene, equipment and medium

Abstract

The invention relates to a target detection method and device in a complex remote sensing scene, equipment and a medium, and the method comprises the steps: carrying out the processing of global features extracted from a remote sensing image to be subjected to target detection through employing a dynamic sparse attention unit, and obtaining key information features; a multi-scale convolution feature interaction network is adopted to extract specific features of different types of targets according to the key information features, multi-scale interaction features are obtained, the multi-scale convolution feature interaction network comprises a multi-scale convolution attention unit and a deformable convolution unit, and finally, a shared weight detection head is adopted to detect the specific features of different types of targets on the basis of the multi-scale interaction features. And carrying out joint regression prediction on the categories of the multiple different categories of targets and bounding box parameters to realize target detection. By adopting the method, the whole target detection framework can be lightened, and meanwhile, the detection accuracy of various different targets in a complex scene is improved.

Description

Technical Field

[0001] The present application relates to the technical field of target detection in aerial remote sensing images, and in particular to a target detection method, apparatus, device and medium for complex remote sensing scenes. Background Art

[0002] Aerial remote sensing imagery, thanks to its wide coverage, is widely used in disaster emergency response, urban traffic monitoring, agricultural resource surveys, and other fields. However, due to the high-altitude nature of image acquisition and resolution limitations, the target area in the image typically appears small, with weak contrast and low information density, presenting significant technical bottlenecks for target positioning and recognition. Furthermore, uncontrollable variables during remote sensing data collection, such as drastic changes in flight altitude, high-speed platform motion, and complex scene interference, further exacerbate the difficulty of distinguishing targets from backgrounds, placing even higher demands on high-precision detection. Existing object detection technologies, deep learning-based algorithms, are categorized into two-stage and single-stage approaches. Two-stage approaches, such as Faster R-CNN, achieve high-precision positioning but suffer from high computational complexity and poor real-time performance. Single-stage approaches, such as RetinaNet and the YOLO series, strike a balance between real-time performance and detection efficiency, but due to insufficient feature fusion layers, face performance bottlenecks in accurately locating small objects in complex scenarios. Recent Transformer-based detection methods, such as DETR, utilize self-attention mechanisms to achieve multi-scale feature interaction and long-range dependencies. However, these methods suffer from high computational complexity and output decoupling, limiting their practical deployment. While the improved RT-DETR improves detection accuracy and inference speed, the computational resource constraints of current aviation platforms still make the deployment of high-performance real-time detection systems challenging. Complex background interference leads to overlapping target and environmental features, existing feature extraction modules are unable to adequately represent multi-scale targets, and the difficulty in balancing detection accuracy and computational efficiency due to the limited resources of embedded hardware necessitates a new technical solution that establishes a dynamic balance between model lightweighting, feature representation completeness, and hardware compatibility. Summary of the Invention

[0003] Based on this, it is necessary to provide a target detection method, device, equipment and medium for complex remote sensing scenarios that can take into account both model lightweight and feature expression completeness to address the above technical problems.

[0004] A target detection method for complex remote sensing scenes, the method comprising: Acquiring a remote sensing image for target detection, wherein the remote sensing image includes multiple different types of targets; Extracting features from the remote sensing image using a global feature extraction unit to obtain global features; The global features are processed using a dynamic sparse attention unit to obtain key information features, wherein the dynamic sparse attention unit divides the global features into multiple regions, and dynamically selects key regions based on the degree of semantic association between the regions to obtain the key information features; A multi-scale convolutional feature interaction network is used to extract specific features of different types of targets based on the key information features to obtain multi-scale interaction features, wherein the multi-scale convolutional feature interaction network includes a multi-scale convolutional attention unit and a deformable convolution unit. The multi-scale convolutional attention unit uses a multi-branch convolution structure and an attention weighting strategy to perform multi-scale feature fusion on the key information features to enhance feature discrimination capability. The deformable convolution unit uses a deformable offset to dynamically adapt to target deformation to extract the multi-scale interaction features. A shared weight detection head is used to perform joint regression prediction of the categories and bounding box parameters of multiple different categories of targets based on the multi-scale interactive features to achieve target detection.

[0005] In one embodiment, the global feature extraction unit uses ResNet-18 as the backbone network.

[0006] In one embodiment, in the dynamic sparse attention unit: Dividing the global feature into a plurality of non-overlapping regions, each of the regions including a plurality of tokens; For each token in the region, a linear transformation is performed using three matrices to generate the corresponding query vector, key vector, and value vector. Based on the query vector and key vector corresponding to each token in each region, the affinity matrix between each region and other regions is obtained; Based on the affinity matrix of each region, a preset number of the most relevant regions are selected as key regions to achieve dynamic sparse screening; Attention calculation is performed on the filtered key areas to obtain the key information features.

[0007] In one embodiment, the query vector and key vector corresponding to each token in each region are processed to obtain an affinity matrix between each region and other regions, including: Aggregate the query vectors corresponding to the tokens in each region to obtain the query vector for that region. Aggregate the key vectors corresponding to the tokens in each region to obtain the key vector for that region. Perform matrix multiplication based on the query vector and key vector of each region to obtain the corresponding affinity matrix.

[0008] In one embodiment, in the multi-scale convolutional attention unit: Using a deep convolutional layer to perform local feature aggregation on the key information features to obtain a first intermediate feature; Performing multi-scale feature extraction on the first intermediate features using a multi-branch depth strip convolution block to obtain a plurality of second intermediate features of different scales; The second intermediate features of multiple different scales and the first intermediate features are de-summed by a sum operation to obtain a preliminary fusion feature; Processing the preliminary fusion features through a convolutional layer to obtain an attention weight map; The attention weight map is used to perform element-by-element matrix multiplication on the key information features to obtain weighted key information features.

[0009] In one embodiment, in the deformable convolution unit: Processing the weighted key information features using a first deformable convolutional layer to obtain initial features; generating a dynamic offset field according to the initial features, and using the dynamic offset field to drive a second deformable convolutional layer to perform irregular sampling to obtain a dynamic offset feature map; The dynamic offset feature map is fused with the weighted key information feature to obtain the multi-scale interaction feature.

[0010] In one embodiment, The convolution kernel size of the first deformable convolution layer is 3X3; The convolution kernel size of the second deformable convolution layer is 5X5.

[0011] The present application also provides a target detection device for complex remote sensing scenarios, the device comprising: A remote sensing image acquisition module for detection is used to acquire remote sensing images for target detection, wherein the remote sensing images include multiple different types of targets; A global feature extraction module is used to extract features from the remote sensing image using a global feature extraction unit to obtain global features; a key information feature extraction module, configured to process the global features using a dynamic sparse attention unit to obtain key information features, wherein the dynamic sparse attention unit divides the global features into multiple regions and dynamically selects key regions based on the degree of semantic association between the regions to obtain the key information features; A multi-scale interactive feature extraction module, configured to extract specific features of different types of targets based on the key information features using a multi-scale convolutional feature interaction network to obtain multi-scale interactive features, wherein the multi-scale convolutional feature interaction network includes a multi-scale convolutional attention unit and a deformable convolution unit. The multi-scale convolutional attention unit uses a multi-branch convolution structure and an attention weighting strategy to perform multi-scale feature fusion on the key information features to enhance feature discrimination capability. The deformable convolution unit uses a deformable offset to dynamically adapt to target deformation to extract the multi-scale interactive features. The target detection module is used to use a shared weight detection head to perform joint regression prediction of the categories and bounding box parameters of multiple different categories of targets based on the multi-scale interactive features to achieve target detection.

[0012] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of a target detection method in a complex remote sensing scene when executing the computer program.

[0013] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a target detection method in a complex remote sensing scene.

[0014] The above-mentioned target detection method, apparatus, device, and medium for complex remote sensing scenarios utilizes a dynamic sparse attention unit to process global features extracted from remote sensing images for target detection to obtain key information features. A multi-scale convolutional feature interaction network is then used to extract specific features of different types of targets based on the key information features, thereby obtaining multi-scale interactive features. The multi-scale convolutional feature interaction network includes a multi-scale convolutional attention unit and a deformable convolution unit. Finally, a shared weight detection head is used to perform joint regression prediction of the categories and bounding box parameters of multiple different target categories based on the multi-scale interactive features, thereby achieving target detection. This method can reduce the weight of the entire target detection architecture while improving the detection accuracy of multiple different targets in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 1 is a flow chart of a target detection method in a complex remote sensing scenario according to an embodiment; Figure 2 Schematic diagram of the structure of a dynamic sparse attention unit in one embodiment; Figure 3 Schematic diagram of the structure of a multi-scale convolutional attention unit in one embodiment; Figure 4 Schematic diagram of the structure of a deformable convolution unit in one embodiment; Figure 5Schematic diagram of the overall structure of a target detection network in one embodiment; Figure 6 A diagram of the confusion matrix of the VisDrone2019 dataset using this method in an experiment; Figure 7 A diagram of the confusion matrix of an experiment using this method on a SIMD dataset; Figure 8 A schematic diagram of visualizing ablation experimental results in an experiment; Figure 9 1 is a block diagram of a target detection device for a complex remote sensing scene according to an embodiment; Figure 10 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0017] In view of the low target detection accuracy in the existing technology in the multi-small target scene acquired from high altitude due to the multi-scale and low resolution of the small targets, and the requirement of lightweight under the limitation of computing resources must be taken into account, in this application, Figure 1 As shown, a target detection method for complex remote sensing scenes is provided, which specifically includes the following steps: Step S100: Acquire a remote sensing image for target detection, where the remote sensing image includes multiple different types of targets.

[0018] Step S110 , utilizing a global feature extraction unit to extract features from the remote sensing image to obtain global features.

[0019] In step S120, a dynamic sparse attention unit is used to process the global features to obtain key information features. The dynamic sparse attention unit divides the global features into multiple regions, and then dynamically selects key regions based on the degree of semantic correlation between the regions to obtain key information features.

[0020] In step S130, a multi-scale convolutional feature interaction network is used to extract specific features of different types of targets based on key information features to obtain multi-scale interaction features. The multi-scale convolutional feature interaction network includes a multi-scale convolutional attention unit and a deformable convolution unit. The multi-scale convolutional attention unit performs multi-scale feature fusion on the key information features by adopting a multi-branch convolution structure and an attention weighted strategy to enhance the feature discrimination ability. The deformable convolution unit uses a deformable offset to dynamically adapt to the target deformation to extract multi-scale interaction features.

[0021] In step S140 , a shared weight detection head is used to perform joint regression prediction of the categories and bounding box parameters of multiple different categories of targets based on multi-scale interactive features to achieve target detection.

[0022] In this application, a feature extraction unit is first used to extract global features from remote sensing images, and a dynamic sparse attention unit is used to implement adaptive sparse weight allocation. This strengthens the common feature representation of similar target elements (such as buildings and vehicles) in remote sensing images while suppressing background noise interference. Subsequently, a multi-scale convolutional feature interaction network is constructed using multi-scale convolutional attention units. Multi-scale convolutional kernel groups are used to extract spatial features of different granularities, and dynamic feature complementarity and spatial attention arousal are achieved through an element-wise multiplication mechanism. At the same time, a deformable convolution layer is introduced into the multi-scale convolutional feature interaction network to adaptively adjust the sampling grid to cope with the diversity of target scale, posture, and imaging mode. Finally, a shared weight detection head is used to perform joint regression prediction of target category and bounding box parameters, balancing computational efficiency and representation accuracy.

[0023] In step S100, the remote sensing image contains a variety of objects of different types and sizes. For example, in the remote sensing image, there are airplanes, cars, trains, pedestrians, and bicycles. Therefore, their sizes and scales are different, and their appearance in the image is also different.

[0024] It should be noted here that the remote sensing images targeted by this method can be obtained at low altitudes as well as high altitudes, and specifically can be targeted at remote sensing images acquired by drones.

[0025] In step S110, the global feature extraction unit uses ResNet-18 as the backbone network.

[0026] In step S120, in the dynamic sparse attention unit: the global features are divided into multiple non-overlapping regions, each of the regions includes multiple tokens, and the tokens in each region are linearly transformed through three matrices to generate corresponding query vectors, key vectors and value vectors, and are processed according to the query vectors and key vectors corresponding to each token in each region to obtain an affinity matrix between each region and other regions, and based on the affinity matrices of each region, the most relevant preset number of regions are selected as key regions to realize dynamic sparse screening, and attention calculation is performed on the screened key regions to obtain key information features.

[0027] In this embodiment, the dynamic sparse attention unit is obtained by optimizing the traditional Transformer self-attention mechanism. Its core idea is to dynamically screen the key areas in the input sequence while reducing the computational complexity while retaining the ability to perceive global information.

[0028] Specifically, the dynamic sparse attention unit first dynamically generates the importance score of each position based on the input features, namely the global features, selects the most relevant local regions and global key points through the Top-k strategy, and then performs intensive attention calculations only on the sparsely selected regions, thereby reducing the O(N) time of the traditional self-attention mechanism to 100. 2) The complexity is reduced to O(kN) (k ≪N\). This dynamic sparsity mechanism not only significantly reduces redundant computation but also adapts to the contextual features of different samples. It balances efficiency and accuracy in remote sensing imagery tasks (such as high-resolution image detection), providing a lightweight solution for complex perception tasks in resource-sensitive scenarios.

[0029] Furthermore, firstly transform the global feature X∈R H×W×C Divide into S×S non-overlapping regions, so that each region contains Each feature vector is converted into , expressed as:

[0030] In the above formula, It is represented as a global feature, which includes S×S non-overlapping regions, each of which includes tokens, where the mapping of each token is a C-dimensional vector.

[0031] In this embodiment, the query vector and key vector corresponding to each Token in each region are processed to obtain an affinity matrix between each region and other regions, including: aggregating the query vectors corresponding to the Tokens in each region to obtain the query vector of the region, aggregating the key vectors corresponding to the Tokens in each region to obtain the key vector of the region, and performing matrix multiplication on the query vectors and key vectors of each region to obtain the corresponding affinity matrix.

[0032] Specifically, three different tokens are selected and mapped to the query vector Q, key vector K and value vector V respectively.

[0033] Each feature vector in the input sequence is passed through three independent trainable parameter matrices ( 、 、 ) is linearly transformed, and the process is expressed as:

[0034]

[0035] Based on the projected vector space, the association matrix pairs Q and K are used to quantify the interaction weights between different regions, thereby adaptively determining the token range that each target region needs to be associated with in the attention calculation. The matrix multiplication between Qr and the transposed Kr derives the adjacency matrix of the region-to-region affinity graph, which is expressed as:

[0036] In the above formula, the entries in A measure the degree of semantic association between two regions, realizing the targeted allocation of computing resources.

[0037] In this embodiment, only the top-k indices of each region are retained to extract the correlation graph. The top-k connections are retained for each region to prune the correlation graph, and I contains the k indices of the most relevant regions.

[0038] Furthermore, the index matrix I can be applied to calculate the token-to-token attention between the tensors K and V. For each query token in region I, the key and value tensors are represented as:

[0039] Finally, the attention operation is used to calculate the aggregated And Q tensors to generate the attention between different tokens in the final image, such as Figure 2 As shown, the key information features are expressed as:

[0040] In step S130, within the multi-scale convolutional feature interaction network, a multi-scale convolutional attention unit (MCU) is employed to extract detailed features specific to the specific morphologies of various targets in drone imagery. This module aggregates local features through deep convolution and captures information at different scales using multi-branch deep stripe convolution. Furthermore, convolution is used to establish inter-channel correlations, and its output serves as attention weights to adaptively reweight the input features of the multi-scale convolution.

[0041] like Figure 3 As shown, in the multi-scale convolutional attention unit, a deep convolutional layer is used to perform local feature aggregation on the key information features to obtain the first intermediate features. A multi-branch deep strip convolution block is then used to perform multi-scale feature extraction on the first intermediate features to obtain multiple second intermediate features of different scales. The multiple second intermediate features of different scales and the first intermediate features are de-summed through a summation operation to obtain a preliminary fused feature. This preliminary fused feature is processed through a convolutional layer to obtain an attention weight map. Finally, the attention weight map is used to perform element-by-element matrix multiplication on the key information features to obtain the weighted key information features.

[0042] Specifically, the kernel sizes of each branch in the multi-branch depthwise strip convolution block are set to 7, 11, and 21, respectively, and two depthwise strip convolutions are superimposed to achieve the equivalent global receptive field of a large-scale convolution kernel. The choice of depthwise strip convolution is based on two main considerations: first, its factorized structure (e.g., replacing the 7×7 standard convolution with 7×1 and 1×7 convolution pairs) significantly reduces computational complexity; second, strip convolution complements traditional grid convolution for segmenting strip-shaped objects (such as human bodies and utility poles), effectively enhancing the ability to extract slender spatial features.

[0043] Furthermore, in the multi-scale convolution attention unit, the input feature F is subjected to DW-Conv depth convolution, and the feature effect after convolution is connected to the identity of Scalei in each branch to obtain the Att attention map, that is, the attention weight map. The Out output is obtained by performing an element-by-element matrix multiplication operation on the Att attention map and the feature F. The process is expressed as follows:

[0044] In this embodiment, a bionic visual perception mechanism is designed in the multi-scale convolutional feature interaction network. By using large-size convolution kernels to simulate the global perception capability of the self-attention mechanism, it avoids the high computational overhead of the traditional self-attention module while retaining the advantages of wide-area modeling. The designed variable large-kernel convolution module further introduces deformable convolution to dynamically optimize the sampling position, and enhances the model's feature adaptability to diverse data distributions through adaptive deformation capabilities. The variable large-kernel convolution module can improve the detection capability of small and irregularly shaped targets, especially in complex backgrounds or under different lighting conditions.

[0045] Specifically, the variable large kernel convolution module adopts a large-scale modeling mechanism, and constructs a large receptive field with a small number of parameters and high efficiency through the composite structure of depthwise separable convolution and dilated convolution. While simulating the global perception characteristics of the self-attention mechanism, this design significantly reduces the computational load through a layered convolution kernel decomposition strategy. Specifically, depthwise separable convolution is used to achieve spatial feature decoupling, combined with dilated convolution to expand the receptive field boundary, and finally the deformation sampling mechanism of deformable convolution is introduced. Deformable convolution adaptively adjusts the feature extraction area by dynamically offsetting the sampling coordinates, effectively enhancing the spatial modeling capability of irregular targets in drone images. Its elastic deformation mechanism breaks through the geometric constraints of traditional convolution kernels. It enables the network to adaptively adjust the feature extraction area, thereby enhancing the representation capability of irregular image structures.

[0046] like Figure 4 As shown in the figure, in the deformable convolution unit: the first deformable convolution layer (DefmConv) is used to process the weighted key information features to obtain initial features, a dynamic offset field is generated based on the initial features, and the dynamic offset field is used to drive the second deformable convolution layer (DefmConv) to perform irregular sampling to obtain a dynamic offset feature map, which is then fused with the weighted key information features to obtain multi-scale interaction features.

[0047] In this embodiment, the convolution kernel size of the first deformable convolution layer is 3×3, and the convolution kernel size of the second deformable convolution layer is 5×5.

[0048] Specifically, the deformable convolution unit uses large-scale convolution kernels to mimic the global perception capabilities of the self-attention mechanism. This approach retains the advantages of wide-area modeling while effectively avoiding the high computational overhead of traditional self-attention modules. The core operational flow of the deformable convolution architecture, illustrated in a schematic diagram, can be broken down into two key stages: First, a 3×3 deformable convolutional layer extracts initial features and generates a dynamic offset field (Offsets). This offset field, represented by a gridded visual metaphor with arrows, represents the adaptive spatial position offset of each sampling point on the feature map. These dynamic offsets then drive a 5×5 deformable convolution kernel to perform irregular sampling. The kernels are no longer constrained to a fixed grid position but instead autonomously adjust their sampling point distribution based on target geometry, achieving robust response to geometric transformations such as vehicle rotation and aircraft wing extension and retraction. The final output is refined into a spatially adaptable feature map by a standard convolution layer. By decoupling offset learning from feature extraction, the entire process significantly improves the model's representation of irregular targets with minimal increase in parameter count, providing underlying technical support for target detection in complex backgrounds for UAV remote sensing.

[0049] In step S140 , a shared weight detection head is used to perform target detection on the multi-scale interactive features.

[0050] In this embodiment, the above-mentioned global feature extraction unit, dynamic sparse attention unit, multi-scale convolutional feature interaction network and shared weight detection head can be constructed to obtain the target detection network (DFE-DETR model). The structure of the entire network is as follows: Figure 5 shown.

[0051] This paper also demonstrates the effectiveness of this method through experiments. Simulations were conducted using the widely used public datasets VisDrone2019 and SIMD (Satellite Imagery for Multiscale Vehicle Detection) data. To verify the robustness of this method, two additional datasets, RSOD and NWPU VHR-10, were also used for validation.

[0052] Experimental environment: All experiments were completed under a unified hardware configuration to eliminate the interference of environmental variables and ensure the reliability of the results and the reproducibility of the experiments. The study adopted a phased verification process: first, parameter optimization was completed based on the training set, and then verification was performed on an independent test set. In scenarios where the sample size is limited, this solution implements a retention verification strategy. Experiments were conducted based on VisDrone2019 and SIMD datasets to verify the effectiveness of the model. In order to demonstrate the robustness of the model, RSOD and NWPU VHR-10 datasets were additionally used to verify the model. Taking into account GPU limitations and training duration, the input size is 640x640 and the batch size is 4. The AdamW optimizer is used for parameter optimization in model training, the basic learning rate is configured to 1e-4, and the momentum coefficient is set to 0.9. In order to comprehensively evaluate the performance of the model, a multi-dimensional evaluation system covering classification accuracy is constructed. , recall rate and mean average precision (mAP@0.5 and mAP@0.5:0.95) and other core indicators are expressed as:

[0053]

[0054]

[0055] Performance experiment analysis: To verify the comprehensive effectiveness of this method, a multi-dimensional evaluation was conducted based on the VisDrone2019 and SIMD benchmark datasets, covering core indicators such as detection accuracy (Precision / Recall), real-time performance (FPS), model lightweightness (Params) and computational efficiency (GFLOPS). Through systematic comparative experiments, the practical application potential of the algorithm from the perspective of drones was revealed.

[0056] Experiments on the VisDrone2019 remote sensing dataset, using a 640×640 pixel input scale and 300 training epochs, demonstrate that this method demonstrates excellent performance in multi-scale object detection tasks (see Table 5 for details). Verification results on the VisDrone2019 dataset demonstrate the model's overall performance advantage: it achieves breakthrough detection accuracy with a mAP@0.5 of 47.34, a mAP@0.5:0.95 of 28.82%, and a precision of 63.12%. Furthermore, it achieves a recall of 0.4620 while maintaining a real-time processing speed of 65.8 FPS, demonstrating outstanding overall performance.

[0057] According to Table 1 and the confusion matrix, Figure 6As shown, in terms of detection accuracy, the car category performs best (P=0.812, mAP50=0.856), followed by buses (mAP50=0.605) and pedestrians (P=0.675). Due to the small number of samples in categories like bicycles and tricycles, detection still presents significant challenges. The recall and mAP50 for tricycles are only 0.165 and 20.5%, respectively. The overall model achieves a mAP50-95 index of 28.82%, and inference latency is kept to 65.8 FPS, achieving a balance between real-time performance and accuracy. The confusion matrix shows false positives for pedestrians (0.57) and background (0.3), suggesting that small objects and complex backgrounds can easily affect object recognition. Notably, while motorcycle detection achieves an accuracy of 0.69, it has a false positive rate of 0.2 for background. This demonstrates the model's ability to accurately perceive small-scale objects in complex scenes. Its multi-level feature fusion mechanism effectively extracts texture and morphological features of small objects. Although there is a delay in inference speed compared to lightweight models (affected by the deep network structure), its significant advantages in detection accuracy and missed detection rate control fully demonstrate the practical value of the algorithm in the field of remote sensing image interpretation.

[0058] Table 1 Experimental results of this method at VisDrone 2019

[0059] Table 5 shows that the SIMD dataset demonstrates comprehensive performance advantages in the UAV remote sensing target detection task: it achieves significant accuracy advantages with 82.79% mAP@0.5, 68.67% mAP@0.5:0.95, and 85.04% precision. The GIoU loss and Cls loss converge to 0.2075 and 0.5086, respectively, validating the effectiveness of the optimizer's dynamic adjustment strategy. Furthermore, the model's high recall rate of 0.8311 and real-time processing capability of 74.6 FPS demonstrate its superior target recognition performance and reduced background interference, providing strong support for its industrial deployment in structured scenarios such as airports and ports.

[0060] According to Table 2 and the confusion matrix, Figure 7The results show that DFE-DETR demonstrates superior performance in UAV remote sensing target detection: The model achieves a recall of 0.98 for structured objects such as passenger aircraft and a mAP@0.5 of 96.5% for training aircraft. Conventional vehicles such as sedans achieve 94.6% mAP@0.5 and ships achieve 96.9% mAP@0.5, demonstrating excellent detection robustness. Confusion matrix validation of the model using randomly sampled images demonstrates extremely high recognition purity for sedans, with a main diagonal value of 0.98. However, significant weaknesses remain for specialized industrial vehicles and similar objects. Recall is as low as 44.4% for pushback trucks and only 68.2% for stair trucks, highlighting deficiencies in geometric feature recognition. The "Other" category is the weakest, with a mAP@0.5 of 32.7% and a main diagonal value of 0.58, indicating that similar categories and background can affect the model's recognition. Particularly noteworthy is that while the recall rate for long vehicles reached 78.2%, the precision was only 51.7%. In the confusion matrix, 14% of long vehicles were misclassified as trucks. Meanwhile, helicopters had a mAP@0.5 ratio of 45.7% and a 0.95 ratio, with 14% misclassified as fighter jets. These findings reveal the model's limitations in locating objects with extreme aspect ratios and adaptability to rotation-sensitive targets. However, the average per-class recognition capability achieved superior accuracy compared to current state-of-the-art models. These results highlight the need for future optimization efforts to focus on three key areas: deformable modeling of specialized vehicles, rotation-enhanced training, and feature decoupling for "difficult" categories.

[0061] Table 2 Experimental results of this method in SIMD

[0062] To validate the generalization performance and domain adaptability of this method, a benchmarking framework was constructed on two remote sensing detection benchmarks, RSOD and NWPU VHR-10, with standardized experimental parameters (input resolution 640×640 and training epochs 150). As shown in the cross-dataset comparison analysis in Tables 3 and 4, the proposed model demonstrates superior accuracy in key metrics such as average precision and recall for the detection of typical remote sensing targets, such as aircraft and oil tankers. Experimental data demonstrates that, through the dynamic sparse attention module and stripe convolutional feature extraction mechanism, DFE-DETR effectively overcomes challenges such as long object strips, extreme scale variations, and complex ground interference in aerial imagery, achieving systematic optimization of technical indicators for remote sensing target detection. Particularly noteworthy is the mAP@0.5 score for oil tankers and playgrounds approaching 99%, demonstrating the model's strong recognition capability for regular objects. While the mAP@0.5 for overpass detection exceeds 90%, background interference reduces its mAP@0.5-0.95 to only 33.3%. Overall, the model demonstrates high detection accuracy and strong generalization for complex remote sensing targets. In terms of speed, the model achieves inference efficiency of 51.0-57.5 FPS, meeting real-time requirements. In particular, mAP@0.5-0.95 for small objects (such as ground track fields) and large-scale objects (such as storage tanks) reaches 83.6% and 67.6%, respectively, demonstrating the model's adaptability across multiple scenarios.

[0063] As shown in Table 5, DFE-DETR achieved balanced performance of 82.79% mAP@0.5 and 0.462 recall on VisDrone2019. On the SIMD dataset, the model achieved 47.34% mAP@0.5 and 0.8311 recall. DFE-DETR also performed well on the other two datasets, achieving 96.58% and 93.28% mAP@0.5 on the RSOD and NWPU VHR-10 datasets, respectively. The recall reached a maximum of 0.934 and 0.8775, with precision exceeding 90%.

[0064] Table 3 Experimental results of this method on RSOD

[0065] Table 4 Experimental results of this method on NWPU VHR-10

[0066] Table 5 Performance comparison of this method on four datasets

[0067] Effectiveness experiments of each unit: A unitized ablation experiment framework was constructed on the VisDrone2019 benchmark dataset, employing a phased and progressive validation strategy to explore the effectiveness of each innovative component. Using a controlled variable approach, three sets of comparative experiments were designed (as shown in Table 6): First, a deformable convolution kernel mechanism was independently embedded in the baseline network, followed by a multi-scale convolutional attention enhancement layer, and finally, a dynamic sparse attention enhancement architecture was integrated. Furthermore, a multi-module joint optimization group was set up to evaluate synergistic effects. This layer-by-layer stacking and combined validation approach accurately quantifies the performance gain of each module in the drone object detection task.

[0068] Table 6 Ablation experiment results

[0069] The experimental data in Table 6 demonstrates that the introduction of the Deformable Convolutional Module (DCN) significantly improves the model's performance compared to the baseline, with mAP@0.5 and mAP@0.5:0.95 increasing by 0.35% and 0.38%, respectively. This performance gain is primarily due to DCN's dynamic deformation sampling mechanism, which effectively captures the irregular geometric features of remote sensing targets through adaptive feature offset learning. While the combined use of DCN and the Multi-scale Convolutional Attention (MCA) module results in an increase in parameters, their joint optimization demonstrates a synergistic effect, yielding even more significant improvements in mAP@0.5 (+0.37%) and mAP@0.5:0.95 (+0.69%). In particular, MCA enhances small object detection capabilities through its cross-scale feature interaction mechanism. These experimental results demonstrate the complementary nature of DCN in spatial deformation modeling and MCA in multi-scale representation learning. Their integration provides a more robust feature representation paradigm for complex remote sensing target detection. After introducing the dynamic sparse attention mechanism, the model showed improved detection accuracy in both the baseline and various variant architectures. This mechanism uses an adaptive feature screening strategy to dynamically optimize features based on the target morphology during training, effectively accelerating model convergence and enhancing overall performance. Experimental data shows that the DFE-DETR model, which integrates three innovative technologies, achieved the best indicators (mAP@0.5:47.34%, mAP@0.5:0.95:28.82%) while maintaining a real-time inference speed of 65.8 FPS. Although the computational complexity and number of parameters have increased significantly, the coordinated optimization between modules has achieved a balance between performance and efficiency. Quantitative results verify the key contributions of core modules such as dynamic sparse attention and multi-scale convolution to the remote sensing target detection task from multiple dimensions.

[0070] Visual ablation experiment design Figure 8As shown, the necessity of each module is verified. Removing the dynamic sparse attention enhancement module weakens the morphological perception of small-scale objects. Traditional single convolutional architectures lack cross-directional feature interaction mechanisms, resulting in insufficient capture of details in multi-shaped objects, leading to missed detections of objects such as airplanes, stadiums, and vehicles. Misjudgments are seen in the second and third rows, such as misclassifying a vehicle as a long object and misclassifying a lane as a runway. The lack of a multi-scale convolutional attention enhancement module leads to misjudgment of vehicles and misidentification of small details, as shown in the third row. Although overall recognition performance improves, the confidence score is low, indicating that the module weakens the representation of slender objects such as pedestrians and vehicles. The unique directional sensitivity of axial depthwise strip convolution plays a key role in modeling linear spatial patterns. Disabling the deformable convolution module results in multiple object boxes, as shown in the fifth row, but the model recognizes objects more accurately and achieves improved confidence scores. Furthermore, the model can automatically generalize to recognize vehicles and parts of airplanes that were not originally labeled. It can be seen that the matching deviation between the fixed-size convolution kernel and the target geometric structure increases significantly, while the deformable convolution dynamically adjusts the receptive field distribution through an adaptive deformation mechanism, which can accurately fit the contour features of irregular objects and effectively improve the resolution accuracy of texture details.

[0071] Visual ablation experiment design Figure 8 As shown, the necessity of each unit is verified. Removing the dynamic sparse attention enhancement module weakens the morphological perception of small-scale objects. Traditional single convolutional architectures lack cross-directional feature interaction mechanisms, resulting in insufficient capture of details in multi-shaped objects, leading to missed detections of objects such as airplanes, stadiums, and vehicles. Misjudgments are seen in the second and third rows, such as misclassifying a vehicle as a long object and misclassifying a lane as a runway. The lack of a multi-scale convolutional attention enhancement module leads to misjudgment of vehicles and misidentification of small details, as shown in the third row. Although overall recognition performance improves, the confidence score is low, indicating that the module weakens the representation of slender objects such as pedestrians and vehicles. The unique directional sensitivity of axial depthwise strip convolution plays a key role in modeling linear spatial patterns. Disabling the deformable convolution module results in multiple object boxes, as shown in the fifth row, but the model recognizes objects more accurately and achieves improved confidence scores. Furthermore, the model can automatically generalize to recognize vehicles and parts of airplanes that were not originally labeled. It can be seen that the matching deviation between the fixed-size convolution kernel and the target geometric structure increases significantly, while the deformable convolution dynamically adjusts the receptive field distribution through an adaptive deformation mechanism, which can accurately fit the contour features of irregular objects and effectively improve the resolution accuracy of texture details.

[0072] Furthermore, a multi-dimensional comparative experimental framework was constructed on the VisDrone2019 remote sensing detection benchmark to systematically evaluate the performance advantages of DFE-DETR in remote sensing object detection. The experimental design encompassed three major technology approaches: single-stage detectors such as YOLO, detection architectures centered around the Transformer, and specialized models such as DTSSNet, specifically designed for remote sensing. Detailed detection accuracy and generalization data are provided in the cross-model comparative analysis module (Table 7). Compared to the traditional YOLO series, DFE-DETR achieved improvements in mAP@0.5 of 0.34%-14.14%, demonstrating significantly enhanced detection robustness in complex scenes. Despite utilizing a hybrid Transformer architecture, its inference speed surpasses conventional limitations, achieving a FPS advantage over the mainstream YOLO detector. This breakthrough in both accuracy and real-time performance sets a new benchmark for industrial applications. Compared with the mainstream Transformer architecture detection model, this model achieves better detection accuracy with a more streamlined network parameter scale. Experimental data shows that it leads most similar models with a significant improvement of 51.3% in the mAP@0.5 indicator. In terms of computational complexity, the model maintains a low level of GFLOPs through structural innovation. This feature enables it to maintain efficient inference speed while surpassing the DTSSNet model designed for remote sensing scenarios in comprehensive accuracy indicators. It is worth noting that although it did not reach the peak level in the accuracy comparison of cutting-edge detection frameworks such as Focus-DETR and DINO, DFE-DETR still maintained competitive detection accuracy, requiring only about 41.97 parameters to achieve higher detection accuracy, which verifies its dual breakthrough in model lightweighting and performance optimization.

[0073] In the drone image detection task, the DFE-DETR model was validated across multiple dimensions on the VisDrone 2019 benchmark dataset, and its performance was compared with current mainstream deformable detection architectures, including DqYOLOF, FSTD-Net, and DAB-Deformable-DETR. Deformable YOLOv10 L was used as a comparison model. This improved YOLOv10 L framework is achieved by replacing conventional convolutional layers with deformable convolutions. As shown in Table 7, this model demonstrates comprehensive advantages across the three key dimensions of parameter count (Params), computational complexity (FLOPs), and detection accuracy (mAP@0.5). Not only does it achieve efficient inference with a minimum of 23.6M parameters and 84.5 FLOPs, it also achieves a mAP@0.5 that matches the performance of current state-of-the-art models, achieving 47.34% mAP@0.5. This demonstrates that the model maintains its lightweight nature while delivering excellent detection performance. In terms of model performance optimization, although DFE-DETR's detection accuracy lags behind DAB-Deformable-DETR by approximately 4.36% mAP@0.5, it demonstrates significant advantages in computational efficiency: structural optimization reduces GFLOPS to 32.5% of DAB-Deformable-DETR, while also reducing the number of model parameters by nearly 50%. This makes DFE-DETR more valuable in latency-sensitive scenarios such as real-time drone detection and mobile deployment.

[0074] Table 7 Comparative experimental results of common object detectors

[0075] This method addresses the core challenges of small target detection in UAV remote sensing imagery (small target size, complex background, and limited computational resources) by proposing a dynamic feature enhancement detection framework, DFE-DETR, based on an improved RT-DETR. This model achieves breakthroughs in both accuracy and efficiency through a triple synergistic optimization mechanism: a dynamic sparse attention unit (SAEM) reduces computational complexity, a multi-scale convolutional attention unit (MSCAEM) enhances feature extraction of slender targets, and a deformable large kernel convolutional unit (DLKCM) optimizes the localization of irregular targets. Experiments on four datasets, including VisDrone2019 and SIMD, demonstrate that DFE-DETR achieves significant performance advantages with only 23.6M parameters: a mAP@0.5 of 47.34% on VisDrone2019, surpassing many superior models, and a mAP@0.5 of 82.79% on SIMD, while maintaining a real-time performance of 74.6 FPS. This demonstrates its value for lightweight deployment in embedded scenarios such as disaster response and traffic monitoring.

[0076] In the aforementioned target detection methods for complex remote sensing scenarios, a DFE-DETR detection model based on the Transformer architecture was designed to address the core challenges of target morphological diversity and scale complexity in remote sensing images. This model uses a dynamic sparse attention module to achieve adaptive feature screening and a dynamic spatial weight allocation strategy to enhance the contour feature perception of irregular targets (such as irregular aircraft and motor vehicles). A multi-scale convolutional attention structure is constructed, using an axial depth-separable convolution kernel group to capture edge texture information of linear objects and small targets, effectively addressing the multi-scale detection challenge under complex background interference. A deformable convolution driver module is designed to achieve adaptive adjustment of the convolution kernel morphology through a dynamic generation mechanism of geometric deformation parameters, significantly improving the model's feature matching accuracy for non-rigid targets such as rotation and distortion. While maintaining the global modeling advantages of the Transformer, this architecture deeply integrates a local detail enhancement strategy to systematically address the problem of accurate detection of geometrically sensitive targets in remote sensing scenarios.

[0077] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0078] In one embodiment, Figure 9 As shown, a target detection device for complex remote sensing scenes is provided, comprising: a remote sensing image acquisition module 200 to be detected, a global feature extraction module 210, a key information feature extraction module 220, a multi-scale interactive feature extraction module 230 and a target detection module 240, wherein: The remote sensing image acquisition module 200 is used to acquire a remote sensing image for target detection, wherein the remote sensing image includes multiple different types of targets; A global feature extraction module 210 is configured to extract features from the remote sensing image using a global feature extraction unit to obtain global features; A key information feature extraction module 220 is configured to process the global features using a dynamic sparse attention unit to obtain key information features, wherein the dynamic sparse attention unit divides the global features into multiple regions and dynamically selects key regions based on the degree of semantic association between the regions to obtain the key information features; A multi-scale interactive feature extraction module 230 is configured to extract specific features of different types of targets based on the key information features using a multi-scale convolutional feature interaction network to obtain multi-scale interactive features. The multi-scale convolutional feature interaction network includes a multi-scale convolutional attention unit and a deformable convolution unit. The multi-scale convolutional attention unit uses a multi-branch convolution structure and an attention weighting strategy to perform multi-scale feature fusion on the key information features to enhance feature discrimination capability. The deformable convolution unit uses a deformable offset to dynamically adapt to target deformation to extract the multi-scale interactive features. The target detection module 240 is configured to perform joint regression prediction of the categories and bounding box parameters of multiple different categories of targets based on the multi-scale interactive features using a shared weight detection head to achieve target detection.

[0079] Regarding the specific limitations of the target detection device for complex remote sensing scenes, please refer to the limitations of the target detection method for complex remote sensing scenes above, which will not be repeated here. The various modules in the above-mentioned target detection device for complex remote sensing scenes can be implemented in whole or in part by software, hardware, and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0080] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 10As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a target detection method for complex remote sensing scenarios is implemented. The display screen of the computer device can be a liquid crystal display or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0081] Those skilled in the art will understand that Figure 10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0082] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented: Acquiring a remote sensing image for target detection, wherein the remote sensing image includes multiple different types of targets; Extracting features from the remote sensing image using a global feature extraction unit to obtain global features; The global features are processed using a dynamic sparse attention unit to obtain key information features, wherein the dynamic sparse attention unit divides the global features into multiple regions, and dynamically selects key regions based on the degree of semantic association between the regions to obtain the key information features; A multi-scale convolutional feature interaction network is used to extract specific features of different types of targets based on the key information features to obtain multi-scale interaction features, wherein the multi-scale convolutional feature interaction network includes a multi-scale convolutional attention unit and a deformable convolution unit. The multi-scale convolutional attention unit uses a multi-branch convolution structure and an attention weighting strategy to perform multi-scale feature fusion on the key information features to enhance feature discrimination capability. The deformable convolution unit uses a deformable offset to dynamically adapt to target deformation to extract the multi-scale interaction features. A shared weight detection head is used to perform joint regression prediction of the categories and bounding box parameters of multiple different categories of targets based on the multi-scale interactive features to achieve target detection.

[0083] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented: Acquiring a remote sensing image for target detection, wherein the remote sensing image includes multiple different types of targets; Extracting features from the remote sensing image using a global feature extraction unit to obtain global features; The global features are processed using a dynamic sparse attention unit to obtain key information features, wherein the dynamic sparse attention unit divides the global features into multiple regions, and dynamically selects key regions based on the degree of semantic association between the regions to obtain the key information features; A multi-scale convolutional feature interaction network is used to extract specific features of different types of targets based on the key information features to obtain multi-scale interaction features, wherein the multi-scale convolutional feature interaction network includes a multi-scale convolutional attention unit and a deformable convolution unit. The multi-scale convolutional attention unit uses a multi-branch convolution structure and an attention weighting strategy to perform multi-scale feature fusion on the key information features to enhance feature discrimination capability. The deformable convolution unit uses a deformable offset to dynamically adapt to target deformation to extract the multi-scale interaction features. A shared weight detection head is used to perform joint regression prediction of the categories and bounding box parameters of multiple different categories of targets based on the multi-scale interactive features to achieve target detection.

[0084] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0085] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0086] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A target detection method for complex remote sensing scenes, characterized in that: The method comprises: Acquiring a remote sensing image for target detection, wherein the remote sensing image includes multiple different types of targets; Extracting features from the remote sensing image using a global feature extraction unit to obtain global features; The global features are processed using a dynamic sparse attention unit to obtain key information features, wherein the dynamic sparse attention unit divides the global features into multiple regions, and dynamically selects key regions based on the degree of semantic association between the regions to obtain the key information features; A multi-scale convolutional feature interaction network is used to extract specific features of different types of targets based on the key information features to obtain multi-scale interaction features, wherein the multi-scale convolutional feature interaction network includes a multi-scale convolutional attention unit and a deformable convolution unit. The multi-scale convolutional attention unit performs multi-scale feature fusion on the key information features by adopting a multi-branch convolution structure and an attention weighting strategy to enhance feature discrimination capability. The deformable convolution unit uses a deformable offset to dynamically adapt to target deformation to extract the multi-scale interaction features. A shared weight detection head is used to perform joint regression prediction of the categories and bounding box parameters of multiple different categories of targets based on the multi-scale interactive features to achieve target detection.

2. The target detection method for complex remote sensing scenes according to claim 1, characterized in that: The global feature extraction unit uses ResNet-18 as the backbone network.

3. The target detection method for complex remote sensing scenes according to claim 2, characterized in that: In the dynamic sparse attention unit: Dividing the global feature into a plurality of non-overlapping regions, each of the regions including a plurality of tokens; For each token in the region, a linear transformation is performed using three matrices to generate the corresponding query vector, key vector, and value vector. Based on the query vector and key vector corresponding to each token in each region, the affinity matrix between each region and other regions is obtained; Based on the affinity matrix of each region, a preset number of the most relevant regions are selected as key regions to achieve dynamic sparse screening; Attention calculation is performed on the filtered key areas to obtain the key information features.

4. The target detection method for complex remote sensing scenes according to claim 3, characterized in that: Based on the query vector and key vector corresponding to each token in each region, the affinity matrix between each region and other regions is obtained, including: Aggregate the query vectors corresponding to the tokens in each region to obtain the query vector for that region. Aggregate the key vectors corresponding to the tokens in each region to obtain the key vector for that region. Perform matrix multiplication based on the query vector and key vector of each region to obtain the corresponding affinity matrix.

5. The target detection method for complex remote sensing scenes according to claim 4, characterized in that: In the multi-scale convolutional attention unit: Using a deep convolutional layer to perform local feature aggregation on the key information features to obtain a first intermediate feature; Performing multi-scale feature extraction on the first intermediate features using a multi-branch depth strip convolution block to obtain a plurality of second intermediate features of different scales; The second intermediate features of multiple different scales and the first intermediate features are de-summed by a sum operation to obtain a preliminary fusion feature; Processing the preliminary fusion features through a convolutional layer to obtain an attention weight map; The attention weight map is used to perform element-by-element matrix multiplication on the key information features to obtain weighted key information features.

6. The target detection method for complex remote sensing scenes according to claim 5, characterized in that: In the deformable convolution unit: Processing the weighted key information features using a first deformable convolutional layer to obtain initial features; generating a dynamic offset field according to the initial features, and using the dynamic offset field to drive a second deformable convolutional layer to perform irregular sampling to obtain a dynamic offset feature map; The dynamic offset feature map is fused with the weighted key information feature to obtain the multi-scale interaction feature.

7. The target detection method for complex remote sensing scenes according to claim 6, characterized in that: The convolution kernel size of the first deformable convolution layer is 3X3; The convolution kernel size of the second deformable convolution layer is 5X5.

8. A target detection device for complex remote sensing scenes, characterized in that: The device comprises: A remote sensing image acquisition module for detection is used to acquire remote sensing images for target detection, wherein the remote sensing images include multiple different types of targets; A global feature extraction module is used to extract features from the remote sensing image using a global feature extraction unit to obtain global features; a key information feature extraction module, configured to process the global features using a dynamic sparse attention unit to obtain key information features, wherein the dynamic sparse attention unit divides the global features into multiple regions and dynamically selects key regions based on the degree of semantic association between the regions to obtain the key information features; A multi-scale interactive feature extraction module is configured to extract specific features of different types of targets based on the key information features using a multi-scale convolutional feature interaction network to obtain multi-scale interactive features. The multi-scale convolutional feature interaction network includes a multi-scale convolutional attention unit and a deformable convolution unit. The multi-scale convolutional attention unit uses a multi-branch convolution structure and an attention weighting strategy to perform multi-scale feature fusion on the key information features to enhance feature discrimination capability. The deformable convolution unit uses a deformable offset to dynamically adapt to target deformation to extract the multi-scale interactive features. The target detection module is used to use a shared weight detection head to perform joint regression prediction of the categories and bounding box parameters of multiple different categories of targets based on the multi-scale interactive features to achieve target detection.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images