A remote sensing target detection method, system, device and medium

Abstract

The application provides a remote sensing target detection method, system, device and medium, and belongs to the technical field of target detection. The method comprises the following steps: pre-processing a remote sensing image and inputting a backbone network to extract a multi-scale feature map; inputting the feature map into an efficient hybrid encoder for feature interaction, wherein the encoder comprises a scale-in feature interaction module based on proxy attention, a nonlinear representation module based on dynamic tanh, and a cross-scale feature fusion module based on a C3K2 module and dynamic upsampling; screening a target query through an uncertainty minimization query selection mechanism from the output feature of the encoder; and inputting the target query into a decoder with a prediction head to obtain a prediction result of a target category and a bounding box. The application solves the problems of high computational complexity of high-resolution images, serious small target missing detection and large model parameter quantity difficult to be deployed in a lightweight manner in remote sensing image target detection, reduces the computing and storage overhead, and improves the detection accuracy and edge deployment capability.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a remote sensing target detection method, system, device and medium. Background Technology

[0002] With the rapid development of high-resolution remote sensing technology, target detection in remote sensing images is increasingly being applied in fields such as smart cities, national defense and security, and environmental monitoring. However, it also faces a series of severe technical challenges. The current technical contradictions are mainly reflected in the following aspects: Target detection in remote sensing images is challenging due to the prevalence of numerous small-scale targets (such as vehicles and ships). Traditional convolutional neural network detectors are limited by their local receptive fields and fixed anchor boxes, leading to false positives and false negatives. While the DETR series of Transformer-based models possess global modeling capabilities, their attention mechanisms are insufficient for responding to extremely subtle features, resulting in limited accuracy in detecting small targets. Furthermore, when dealing with targets that typically have extremely high spatial resolution (such as...),... For remote sensing images with pixel counts (N), advanced models such as RT-DETR have a computational complexity of O(N^2) due to their core global self-attention mechanism. 2 When processing high-resolution input, the computational load and memory usage increase explosively, severely restricting deployment in real-time and resource-constrained scenarios. Furthermore, the strict limitations imposed by edge computing platforms (such as satellites and drones) on model size and computational resources directly contradict the large number of parameters in existing high-performance detection models (e.g., RT-DETR-L has 32M parameters), resulting in high training requirements, heavy storage and memory pressure at the edge, and slow inference speed, making it difficult to meet the stringent requirements of onboard real-time processing. Although the Real-Time Detection Transformer (RT-DETR) has made significant progress in speed and accuracy through hybrid encoders and dynamic query mechanisms, it has not fundamentally solved the above three major contradictions: its attention complexity remains O(N). 2 The model is inefficient at processing ultra-high resolution images; the multi-scale fusion mechanism has limited effect on enhancing the features of small targets; and the model has a high number of parameters, which is not conducive to edge deployment.

[0003] Therefore, existing technologies lack a remote sensing target detection method that can simultaneously and efficiently process high-resolution inputs, accurately detect small-scale targets, and has a sufficiently lightweight model. Summary of the Invention

[0004] The purpose of this invention is to provide a remote sensing target detection method, system, device, and medium, aiming to resolve the contradictions existing methods face in terms of small target detection accuracy, high-resolution image processing efficiency, and lightweight model deployment.

[0005] To achieve the above objectives, the present invention provides a remote sensing target detection method, comprising the following steps: Step S1: Preprocess the acquired remote sensing images; Step S2: Input the preprocessed image into the backbone network for feature extraction to obtain multi-scale feature maps; Step S3: Input the multi-scale feature map into the efficient hybrid encoder for feature interaction. The efficient hybrid encoder includes: An intra-scale feature interaction module based on proxy attention is used to reduce the computational complexity of self-attention from O(N) 2 The time complexity is reduced to O(NM), where N is the length of the input sequence and M is the length of the proxy tokens. A nonlinear characterization module based on dynamic tanh is used to replace the traditional normalization layer; A cross-scale feature fusion module based on the C3K2 module and the dynamic upsampling module is used to reduce the number of parameters and computational cost while fusing multi-scale features. Step S4: The features output by the high-efficiency hybrid encoder are filtered to select the target query through an uncertainty minimization query selection mechanism; Step S5: Input the selected target query into the decoder with the prediction head to obtain the prediction results of the target category and bounding box.

[0006] Preferably, in step S3, the implementation of the intra-scale feature interaction module based on agent attention includes: The input feature map is transformed linearly to obtain the query vector Q, the key vector K, and the value vector V. The query vector Q is used to generate proxy A through a pooling operation; Two-stage attention computation is performed. In the first stage, agent A, acting as the query vector Q in the attention mechanism, is combined with the key vector K and agent bias using a softmax function and multiplied by the value vector V. The result is... In the second-level computation, agent A, acting as the key K in the attention mechanism, is combined with query Q and bias to perform a softmax function calculation and multiply. ; The result of the two-level attention calculation is added to the value vector V processed by deep convolution (DWC) to output a feature map with global dependencies.

[0007] Preferably, in step S3, the function expression of the nonlinear representation module based on dynamic tanh is... ;in For learnable scaling parameters, The hyperbolic tangent activation function is used. , For learnable channel-level recovery parameters, The input is the feature tensor of this module.

[0008] Preferably, in step S3, the implementation of the C3K2 module includes: The input feature map is passed through Convolutional layers, batch normalization, and the SiLU activation function are used to expand the number of channels. The expanded features are then equally divided into two parts along the channel dimension. The first part is directly retained. The second part is passed through multiple bottleneck network modules. The two processed feature parts are then concatenated along the channel dimension. Finally, a... Convolutional layers, batch normalization, and SiLU activation function are used to adjust the number of output channels and complete feature fusion. Each bottleneck network contains two Each convolutional layer is followed by a batch normalization and SiLU activation function, and residual connections are provided.

[0009] Preferably, in step S3, the implementation of the dynamic upsampling module includes: A sample set is generated by a sample point generator, which uses either a static scaling factor or a dynamic scaling factor. The input feature map is resampled using a grid sampling function to obtain an upsampled feature map.

[0010] Preferably, in step S4, the uncertainty minimization query selection mechanism quantifies the uncertainty score by calculating the cross-union ratio variation between the predicted bounding box and the ground truth annotation during the training phase, and performs gradient-weighted optimization on the query based on the score; during the inference phase, the query is pruned based on the uncertainty score.

[0011] Preferably, in step S5, the decoder with a prediction head includes multiple sequentially stacked decoding layers, each decoding layer including a self-attention module, a cross-attention module, and a prediction head module for outputting classification scores and bounding box coordinates.

[0012] The present invention also provides a remote sensing target detection system, comprising: The preprocessing module is used to preprocess the acquired remote sensing images; The backbone network module is used to extract features from the preprocessed image to obtain multi-scale feature maps; The high-efficiency hybrid encoder module is used for feature interaction on multi-scale feature maps. The high-efficiency hybrid encoder module includes an intra-scale feature interaction module based on surrogate attention, a nonlinear representation module based on dynamic tanh, and a cross-scale feature fusion module based on C3K2 module and dynamic upsampling module. The query selection module, connected to the high-efficiency hybrid encoder module, is used to filter target queries through an uncertainty-minimizing query selection mechanism. The decoder module, connected to the query selection module, processes the filtered target queries to obtain prediction results for the target category and bounding box.

[0013] The present invention also provides a computer device, including: a memory and a processor; the memory stores a computer program, and the processor executes the computer program to implement the steps of the remote sensing target detection method described above.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the remote sensing target detection method described above.

[0015] Therefore, the present invention employs the above-described remote sensing target detection method, system, equipment, and medium, and the beneficial technical effects are as follows: (1) By introducing a proxy attention mechanism, this invention reduces the computational complexity of self-attention in the feature interaction stage from quadratic to near linear, effectively alleviating the computational explosion and memory consumption problems when processing high-resolution images. Combined with dynamic upsampling and the lightweight C3K2 module, the number of parameters and computational overhead of the model in the feature fusion stage are further reduced, making the model more suitable for deployment environments with limited computing resources.

[0016] (2) To address the problem of high proportion and weak features of small targets in remote sensing images, this invention uses the C3K2 module and the dynamic upsampling module to perform multi-scale feature fusion, which enhances the preservation and reconstruction of detailed information. Combined with the dynamic tanh activation function to enhance nonlinear representation, it further improves the feature response and discrimination ability for extremely small targets, thereby significantly reducing the false negative and false positive rates of small targets.

[0017] (3) This invention improves detection accuracy while effectively controlling the number of model parameters and computational burden by organically combining proxy attention, lightweight C3K2 structure, dynamic upsampling module and dynamic tanh module. This design enables the model to maintain high detection performance while having smaller storage footprint and lower inference overhead, and can better adapt to the real-time detection needs of edge computing platforms such as satellites and UAVs. Attached Figure Description

[0018] Figure 1 A schematic diagram of the overall structure of the improved RT-DETR model; Figure 2 This is a schematic diagram of the intra-scale feature interaction module based on agent attention; Figure 3 This is a schematic diagram of the nonlinear characterization module based on dynamic tanh. Figure 4 This is a schematic diagram of the C3K2 module. Figure 5 This is a schematic diagram of the bottleneck network structure; Figure 6This is a schematic diagram of the dynamic upsampling module. Figure 7 This is a schematic diagram of a static scaling sampling point generator; Figure 8 A schematic diagram of a dynamically scaled sampling point generator; Figure 9 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0021] Example 1 This embodiment provides a remote sensing target detection method, the overall process of which is as follows: Figure 9 As shown, the specific steps include: 1. Data preprocessing.

[0022] To acquire raw remote sensing images (e.g., from drones or satellites), the first step is to standardize the images, resizing them to a fixed size (e.g., ...). (pixels). Then, normalization is performed to scale the pixel values ​​to the [0, 1] range. To enhance the model's generalization ability, multi-scale random cropping and horizontal flipping are used for data augmentation. Simultaneously, the annotation information (bounding box coordinates and class labels) is converted to the format required for model training.

[0023] 2. Feature extraction.

[0024] A schematic diagram of the overall structure of the improved RT-DETR model is shown below. Figure 1 As shown, the preprocessed image is input into the backbone network for multi-scale feature extraction. This embodiment uses ResNet18 as the backbone network. The input image sequentially passes through an initial convolutional layer and four stages. Each stage consists of two basic modules and residual connections, and each basic module consists of a... The process consists of convolution, batch normalization, and the ReLU activation function. Each stage generates feature maps of different scales, ultimately outputting three feature maps of different scales (denoted as S3, S4, and S5), which form a feature pyramid for subsequent multi-scale fusion and interaction.

[0025] 3. High-efficiency hybrid encoder processing.

[0026] The multi-scale feature maps output from the backbone network are input into an efficient hybrid encoder, which is the core improvement of this invention: (1) Intra-scale feature interaction module based on agent attention.

[0027] Intra-scale feature interaction is performed on feature map S5, and the final output is the feature map after intra-scale feature interaction processing. .like Figure 2 As shown, the feature map is first flattened and then linearly transformed to obtain the query vector Q, the key vector K, and the value vector V. This means projecting the input features into a query vector Q. This means projecting the input features into a key vector K. This means projecting the input features into a value vector V. The query vector Q is pooled to generate a set of agents A. Then, two levels of attention computation are performed: In the first level, agent A is used in the attention mechanism to perform a softmax function calculation on the query vector Q, key vector K, and agent bias, and then multiplied by the value vector V. The result is... In the second stage, agent A, acting as the key vector K in the attention mechanism, performs a softmax function calculation with the query vector Q and the bias, and then multiplies it by the result obtained in the previous step. Finally, the attention result is added to the value vector V processed by depthwise convolution (DWC) to obtain the enhanced feature with global dependencies. This process reduces the computational complexity of attention from O(N) to O(N). 2 The value is significantly reduced to O(NM), where M is a value much smaller than N.

[0028] (2) Nonlinear characterization module based on dynamic tanh.

[0029] After attention calculation, a dynamic tanh module is used instead of the traditional normalization layer. For example... Figure 3 As shown, this module handles the input Apply learnable scaling parameters After activation by the tanh function, it is then processed by learnable channel-level parameters. and To scale and translate, i.e. This operation enhances the model's nonlinear fitting ability while avoiding the computational overhead of the normalization layer.

[0030] (3) Cross-scale feature fusion module based on C3K2 module and dynamic upsampling module.

[0031] To integrate features from different scales from the backbone network, this invention designs a lightweight C3K2 module that works in conjunction with dynamic upsampling.

[0032] C3K2 module: structure as follows Figure 4 As shown. The input features first pass through a... A convolutional layer is followed by batch normalization (BatchNorm) and SiLU activation function to expand the number of channels; the expanded features are equally divided into two parts along the channel dimension; the first part is directly retained as a shortcut connection; the second part passes through multiple bottleneck network modules in sequence (structure as follows). Figure 5 As shown, each bottleneck network contains two Each convolutional layer is followed by batch normalization and SiLU activation, and residual connections are provided. The two processed feature sets are concatenated along the channel dimension; finally, a... The convolutional layer (followed by batch normalization and SiLU activation function) adjusts the number of output channels to complete feature fusion.

[0033] Dynamic upsampling module: structure as follows Figure 6 As shown, where , , These represent the height, width, and number of channels of the input feature map, respectively. and These are the upsampling scale factor and the number of groups, respectively. To align and fuse deep and shallow features, dynamic upsampling is used instead of traditional interpolation methods. The input features generate a sample set through a sampling point generator and are then resampled with the original input features using a grid sampling function, reducing computational cost while reconstructing details.

[0034] Sample point generator in dynamic upsampling, such as Figure 7 , Figure 8 As shown (symbols in the figure) and (These represent element-wise addition and dot product, respectively). Figure 7 This is a static range factor version of the sampling point generator, an upsampling method that directly predicts the translation of all upsampling points using a linear layer. Figure 8 This is a dynamic range factor version sampling point generator. The dynamic range factor version is an improvement on the static range factor version, designed to increase the flexibility of translation. It introduces a dynamic range modulation mechanism, allowing the translation amplitude at different positions to be adaptively adjusted, effectively constraining the movement range of sampling points and avoiding disordered sampling. In the static range factor version, the input features generate 2 channels through a linear transformation. 2 The translation tensor is then reshaped using pixel shuffling and a scaling factor of 0.25. The translation amount is then added to the original sampling grid to obtain the sampling set. The dynamic range factor version of the sampling point generator contains two branches: one branch generates the basic translation amount through a linear transformation, and the other branch, after performing a linear transformation, passes it through a 0.5... (i.e., the 0.5 sigmoid function) scales to the interval [0, 0.5]. The final translation is obtained by multiplying the dynamic range factor and the base translation element-wise, and then adding it to the original sampling grid to obtain the sampling set. The high-efficiency hybrid encoder achieves efficient and lightweight interaction and fusion of multi-scale features through the collaborative work of the above modules.

[0035] 4. Minimize uncertainty in query selection.

[0036] The high-level features output by the efficient hybrid encoder are input into a query selection module that minimizes uncertainty. This module intelligently filters and weights query vectors in the decoder by establishing a probabilistic query evaluation system. Specifically, each query generates two parallel outputs: a bounding box prediction and an uncertainty score, which is quantified by the variational Intersection over Union (IoU) between the predicted box and the ground truth annotation. During training, the model prioritizes queries with high uncertainty for optimization, applying larger gradient weights to these queries through a variant of Focal Loss, thus concentrating network resources on difficult-to-classify samples. Simultaneously, during inference, an entropy-based query pruning strategy is employed to automatically filter out low-quality queries (those with uncertainty scores below a threshold). =0.5), retaining the most discriminative queries for the final prediction.

[0037] 5. Decoding and Prediction.

[0038] The decoder with a prediction head in RT-DETR is the core component of its object detection architecture, achieving high-precision detection through a multi-level progressive decoding mechanism. This decoder consists of six identical stacked decoding layers, each containing three key modules: First, a self-attention module models target relationships through interactions between queries, employing a sparse attention mechanism to reduce computational complexity; second, a cross-attention module dynamically aligns the query with the multi-scale features output by the encoder, focusing on potential target regions through deformable attention; finally, the prediction head module outputs classification scores and bounding box coordinates in parallel. The classification head uses dynamic focal loss to address class imbalance, while the regression head iteratively refines the box positions (predicting relative translations in each decoding layer). RT-DETR innovatively embeds an uncertain prediction sub-head in the decoder, generating a confidence score for each query, used for subsequent query filtering and result weighting. During decoding, the prediction results of the previous layer serve as the initial values ​​for the next layer's query, forming a cascaded optimization that ultimately outputs the detection results.

[0039] 6. Output results.

[0040] The prediction results output by the decoder are post-processed (such as threshold filtering) to obtain the final target detection results of the remote sensing image, including the category label of each detected target and its location bounding box in the image.

[0041] The invention will be further illustrated below with specific examples.

[0042] To verify the effectiveness of the method of this invention, a comparative experiment was conducted on the public dataset NWPUVHR10. During training, the AdamW optimizer was used, with an initial learning rate set to 1e-4 (1e-5 for the backbone network), and weight decay was introduced to optimize the training process. The batch size was set to 8.

[0043] Under the same experimental conditions, the original RT-DETR model and the improved model of this invention were compared and evaluated. As shown in Table 1, the improved model shows significant advantages in several key indicators: Table 1 Comparison Results

[0044] Experimental results show that this invention, through the introduction of surrogate attention, dynamic tanh, and the collaborative design of C3K2 and dynamic upsampling, effectively improves detection accuracy while significantly reducing the number of model parameters and computational complexity. Specifically, the model achieves an improvement in the core evaluation metric, mean accuracy (mAP@50); both the number of parameters and computational cost are reduced, demonstrating its superior lightweight characteristics and computational efficiency. This provides a reliable technical solution for achieving high-precision real-time remote sensing target detection on resource-constrained edge devices.

[0045] Example 2 A remote sensing target detection system, comprising: The preprocessing module is used to preprocess the acquired remote sensing images; The backbone network module is used to extract features from the preprocessed image to obtain multi-scale feature maps; The high-efficiency hybrid encoder module is used for feature interaction on multi-scale feature maps. The high-efficiency hybrid encoder module includes an intra-scale feature interaction module based on surrogate attention, a nonlinear representation module based on dynamic tanh, and a cross-scale feature fusion module based on C3K2 module and dynamic upsampling module. The query selection module, connected to the high-efficiency hybrid encoder module, is used to filter target queries through an uncertainty-minimizing query selection mechanism. The decoder module, connected to the query selection module, processes the filtered target queries to obtain prediction results for the target category and bounding box.

[0046] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0047] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0048] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0049] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0050] Therefore, the present invention adopts the above-mentioned remote sensing target detection method, system, device and medium. By introducing a proxy attention mechanism, a dynamic tanh activation function and a collaborative design of C3K2 module and dynamic upsampling, it effectively solves the three major technical contradictions in high-resolution remote sensing image processing: high computational complexity, insufficient accuracy of small target detection and large number of model parameters that are difficult to deploy in a lightweight manner. It significantly reduces computation and storage overhead and improves the overall detection accuracy and edge deployment feasibility of the model.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A remote sensing target detection method, characterized in that, Includes the following steps: Step S1: Preprocess the acquired remote sensing images; Step S2: Input the preprocessed image into the backbone network for feature extraction to obtain multi-scale feature maps; Step S3: Input the multi-scale feature map into the efficient hybrid encoder for feature interaction. The efficient hybrid encoder includes: An intra-scale feature interaction module based on proxy attention is used to reduce the computational complexity of self-attention. A nonlinear characterization module based on dynamic tanh is used to replace the traditional normalization layer; A cross-scale feature fusion module based on the C3K2 module and the dynamic upsampling module is used to reduce the number of parameters and computational cost while fusing multi-scale features. Step S4: The features output by the high-efficiency hybrid encoder are filtered to select the target query through an uncertainty minimization query selection mechanism; Step S5: Input the selected target query into the decoder with the prediction head to obtain the prediction results of the target category and bounding box.

2. The remote sensing target detection method according to claim 1, characterized in that, In step S3, the implementation of the intra-scale feature interaction module based on agent attention includes: The input feature map is transformed linearly to obtain the query vector Q, the key vector K, and the value vector V. The query vector Q is used to generate proxy A through a pooling operation; Two-stage attention computation is performed. In the first stage, agent A, acting as the query vector Q in the attention mechanism, is combined with the key vector K and agent bias using a softmax function and multiplied by the value vector V. The result is... In the second-level computation, agent A, acting as the key K in the attention mechanism, is combined with query Q and bias to perform a softmax function calculation and multiply. ; The result of the two-level attention calculation is added to the value vector V processed by deep convolution (DWC) to output a feature map with global dependencies.

3. The remote sensing target detection method according to claim 1, characterized in that, In step S3, the function expression of the nonlinear representation module based on dynamic tanh is... ;in For learnable scaling parameters, The hyperbolic tangent activation function is used. , For learnable channel-level recovery parameters, The input is the feature tensor of this module.

4. The remote sensing target detection method according to claim 1, characterized in that, In step S3, the implementation of the C3K2 module includes: The input feature map is passed through Convolutional layers, batch normalization, and the SiLU activation function are used to expand the number of channels. The expanded features are then equally divided into two parts along the channel dimension. The first part is directly retained. The second part is passed through multiple bottleneck network modules. The two processed feature parts are then concatenated along the channel dimension. Finally, a... Convolutional layers, batch normalization, and SiLU activation function are used to adjust the number of output channels and complete feature fusion. Each bottleneck network contains two Each convolutional layer is followed by a batch normalization and SiLU activation function, and residual connections are provided.

5. The remote sensing target detection method according to claim 1, characterized in that, In step S3, the implementation of the dynamic upsampling module includes: A sample set is generated by a sample point generator, which uses either a static scaling factor or a dynamic scaling factor. The input feature map is resampled using a grid sampling function to obtain an upsampled feature map.

6. The remote sensing target detection method according to claim 1, characterized in that, In step S4, the uncertainty minimization query selection mechanism quantifies the uncertainty score by calculating the cross-union ratio variation between the predicted bounding box and the ground truth annotation during the training phase, and performs gradient-weighted optimization of the query based on the score; during the inference phase, the query is pruned based on the uncertainty score.

7. The remote sensing target detection method according to claim 1, characterized in that, In step S5, the decoder with a prediction head contains multiple decoding layers stacked sequentially. Each decoding layer includes a self-attention module, a cross-attention module, and a prediction head module for outputting classification scores and bounding box coordinates.

8. A remote sensing target detection system, characterized in that, include: The preprocessing module is used to preprocess the acquired remote sensing images; The backbone network module is used to extract features from the preprocessed image to obtain multi-scale feature maps; The high-efficiency hybrid encoder module is used for feature interaction on multi-scale feature maps. The high-efficiency hybrid encoder module includes an intra-scale feature interaction module based on surrogate attention, a nonlinear representation module based on dynamic tanh, and a cross-scale feature fusion module based on C3K2 module and dynamic upsampling module. The query selection module, connected to the high-efficiency hybrid encoder module, is used to filter target queries through an uncertainty-minimizing query selection mechanism. The decoder module, connected to the query selection module, processes the filtered target queries to obtain prediction results for the target category and bounding box.

9. A computer device, comprising: Memory and processor; The memory stores a computer program, characterized in that when the processor executes the computer program, it implements the steps of the remote sensing target detection method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When a computer program is executed by a processor, it implements the steps of a remote sensing target detection method according to any one of claims 1-7.

Images