Remote sensing small target detection method and device, computer device and storage medium

Abstract

The embodiment of the present application relates to the field of artificial intelligence technology, and discloses a remote sensing small target detection method and device, computer equipment and storage medium, the method comprises: obtaining an image to be identified;The image to be identified is input into a remote sensing target detection model to obtain a remote sensing detection target in the image to be identified;The remote sensing target detection model comprises a super-resolution branch and a target detection branch;The super-resolution branch and the target detection branch share a parameter sharing encoder;The target detection branch is used for receiving the image to be identified and performing target detection on the image to be identified;The super-resolution branch is used for iteratively training the target detection branch together with the target detection branch in the remote sensing target detection model training stage to adjust the parameters of the target detection branch;Output target detection result.The embodiment of the present application can effectively share the feature parameters of the super-resolution branch, the structure is efficient, the small target area can be accurately focused, and the accuracy of target detection is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method, apparatus, computer device, and storage medium for detecting small targets in remote sensing. Background Technology

[0002] Currently, remote sensing target detection aims to identify and locate targets from aerial and satellite imagery, and is a core task in remote sensing image analysis. With the continuous development of platforms such as satellites and drones, acquiring remote sensing images has become more convenient, providing more possibilities for Earth observation. However, target detection in remote sensing images still faces significant challenges, particularly in complex backgrounds, weak feature responses, large scale variations, and the identification of small targets.

[0003] In recent years, super-resolution technology has been introduced into remote sensing image processing to enhance details and provide clear input for detection. Multi-task learning combining super-resolution and object detection has become a research hotspot. Some studies have also embedded super-resolution tasks into detection frameworks to replace downsampling and other methods, mitigating information loss with low computational cost. However, current mainstream multi-task frameworks mostly adopt a sequential paradigm of "super-resolution reconstruction → detection," which has significant drawbacks: the super-resolution stage overemphasizes pixel-level indicators such as PSNR and SSIM, which can easily conflict with the preservation of semantic features required for detection. Furthermore, the detection stage cannot guide the super-resolution process in reverse, and the reconstruction results may introduce artifacts or detail distortion. Independent modeling in two stages also leads to computational redundancy, limiting its application in resource-constrained scenarios. At the same time, even though the query mechanism of the DETR series models overcomes the limitations of CNN's local perception through self-attention and can capture global context, its encoding process still does not effectively distinguish key image information. When lacking saliency guidance, it is necessary to uniformly model the entire image feature space, which can easily introduce redundant and irrelevant background features, leading to a decrease in the accuracy of small target localization. This problem is more prominent in complex backgrounds or extremely small target scenes.

[0004] To address the aforementioned issues, current research is gradually shifting towards multi-task learning frameworks to achieve synergistic optimization of image reconstruction and object detection. However, two major challenges remain: (1) Inefficient feature sharing mechanisms: Gradient conflicts easily arise during network training for super-resolution and detection tasks, leading to mutual interference between the two tasks and inconsistent optimization objectives; (2) Insufficient dynamic reasoning capabilities: Although multi-task learning for super-resolution and object detection has improved the usability of low-quality images to some extent, effectively focusing on and enhancing salient regions remains a key challenge in improving the accuracy of small object detection. In practical applications, super-resolution reconstruction may lead to a shift in feature distribution, resulting in insufficient adaptability of the query mechanism during dynamic reasoning, making it unable to effectively cope with changes in image features, thereby affecting detection accuracy.

[0005] From a theoretical perspective, the root of these problems lies in two points: first, the differences in task characteristics and requirements. For example...Figure 1 As shown, the activation maps of the two tasks differ. Super-resolution focuses more on frequency domain information recovery, improving resolution by enhancing details; while object detection relies on spatial semantic representation, emphasizing object localization and recognition. Both lack a unified alignment mechanism at the feature level. Secondly, the blurring of salient regions in low-quality images makes it difficult for models to accurately focus on small target regions. Convolutional neural network (CNN)-based methods, due to their local perception characteristics, often fail to accurately focus on and identify these faint targets when image quality is low and background interference is strong. The DETR series of object detection methods based on Transformer can dynamically focus on target regions through an attention mechanism and effectively capture long-range contextual information. However, traditional Top-k filtering query selection strategies suffer from increased false positive rates and struggle to accurately focus on small target regions. Therefore, there is an urgent need to explore a multi-task fusion strategy that is feature-shared, semantically consistent, and structurally efficient. Summary of the Invention

[0006] In view of the above problems, embodiments of the present invention provide a remote sensing small target detection method, apparatus, computer equipment and storage medium to solve the problems of non-shareable features, semantic inconsistency and structural inefficiency in the prior art.

[0007] According to one aspect of the present invention, a method for detecting small targets in remote sensing is provided, the method comprising:

[0008] Acquire the image to be recognized;

[0009] The image to be identified is input into a remote sensing target detection model to obtain the remote sensing target in the image to be identified; the remote sensing target detection model includes a super-resolution branch and a target detection branch; the super-resolution branch and the target detection branch share a parameter-shared encoder; the target detection branch is used to receive the image to be identified and perform target detection on the image to be identified; the super-resolution branch is used to perform iterative training together with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch;

[0010] Output the target detection results.

[0011] In an alternative approach, before inputting the image to be identified into the remote sensing target detection model to obtain the remote sensing target in the image to be identified, the method further includes:

[0012] Freeze the super-resolution branch, input the training image into the target detection branch for iterative training, and obtain the pre-trained target detection branch;

[0013] The super-resolution branch and the pre-trained target detection branch are trained together iteratively to obtain a trained remote sensing target detection model.

[0014] In one alternative approach, the super-resolution branch includes a parameter-shared encoder, a super-resolution decoder, and an image reconstruction module; the super-resolution branch and the pre-trained target detection branch are iteratively trained together to obtain a trained remote sensing target detection model, including:

[0015] The training images are input into the parameter-sharing encoder to obtain the sample feature extraction results;

[0016] The sample feature extraction results are input into the super-resolution decoder and image reconstruction module to obtain the sample image after super-resolution;

[0017] The feature extraction results are input into the pre-trained target detection branch to obtain sample target boxes;

[0018] Based on the ultra-high resolution sample image and sample target bounding box, and the preset multi-task joint loss function, the parameters are adjusted, and the process continues to execute the steps of inputting the training image into the parameter sharing encoder to obtain the feature extraction result, inputting the feature extraction result into the super-resolution decoder and the pre-trained target detection branch respectively to obtain the ultra-high resolution sample image and sample target bounding box, and adjusting the parameters based on the ultra-high resolution sample image and sample target bounding box, and the preset multi-task joint loss function, until a trained remote sensing target detection model is obtained.

[0019] In one optional approach, the parameter adjustment based on the ultra-high resolution sample image and sample target bounding box, and a preset multi-task joint loss function, includes:

[0020] Set the target detection loss as follows:

[0021] ;

[0022] in, These are the losses from conventional detection. For each weight, For significance constraint loss, Loss weights are used to constrain significance.

[0023] Set super-resolution loss The super-resolution loss Pixel reconstruction constraints are applied:

[0024] ;

[0025] in, Constraint weights for pixel reconstruction Measure the pixel difference between the reconstructed image and the high-resolution label;

[0026] The multi-task joint loss function Defined as:

[0027] .

[0028] In one alternative approach, the parameter-sharing encoder includes a multi-layer Swin module, which divides the image to be recognized into multiple non-overlapping windows through convolutional layers, performs self-attention calculation within each window, and employs a shift window strategy to capture multi-scale features when crossing windows, thereby obtaining multi-scale feature extraction results.

[0029] The super-resolution decoder includes a multi-level patch extension module and a residual Swin module, which are used to restore spatial resolution through deconvolution and normalization operations during feature amplification, and to adaptively adjust the feature channel dimension to obtain multi-scale features.

[0030] The image reconstruction module includes a convolutional layer, an activation function, and an upsampling module, which are used to map the multi-scale features output by the super-resolution decoder to the image domain to obtain a sample image with ultra-high resolution.

[0031] In one alternative approach, the feature amplification process restores spatial resolution through deconvolution and normalization operations, and adaptively adjusts the feature channel dimension to obtain multi-scale features, including:

[0032] ;

[0033] The input features are represented by 'Reshape'; 'LN' represents the normalization operation; 'TansConv' represents the deconvolution operation, used to amplify features in the spatial dimension; and 'Reshape' represents restoring the result to a sequence form.

[0034] In one alternative approach, inputting the feature extraction results into the pre-trained target detection branch to obtain sample target boxes includes:

[0035] The global features in the multi-scale feature extraction results output by the parameter-sharing encoder are subjected to sequence average pooling and then broadcast expansion and concatenated with local features to obtain fused features.

[0036] The fused features are input into a three-layer feedforward network for layer-by-layer dimensionality reduction to obtain the saliency score map of each channel;

[0037] An initial saliency map is generated based on the saliency score map corresponding to the top-level feature map, and then passed down to lower levels to obtain a cross-scale saliency map.

[0038] Calculate the original target image coordinates corresponding to the query vector of the target location in each feature map layer;

[0039] The saliency score of the query vector is calculated based on the original target image coordinates and the ground truth bounding box.

[0040] Based on the saliency score, a hierarchical filtering mechanism is used to filter the query vectors at each level and scale to obtain the target query vector set.

[0041] Attention is updated based on the target query vector set, and the final sample target box is generated through the detection head.

[0042] According to another aspect of the present invention, a remote sensing target detection device is provided, comprising:

[0043] The acquisition module is used to acquire the image to be recognized;

[0044] The target detection module is used to input the image to be identified into a remote sensing target detection model to obtain the remote sensing target in the image to be identified; the remote sensing target detection model includes a super-resolution branch and a target detection branch; the super-resolution branch and the target detection branch share a parameter-shared encoder; the target detection branch is used to receive the image to be identified and perform target detection on the image to be identified; the super-resolution branch is used to perform iterative training together with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch;

[0045] The output module is used to output the target detection results.

[0046] According to another aspect of the present invention, a computer device is provided, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus;

[0047] The memory is used to store at least one executable instruction that causes the processor to perform the operation of the remote sensing small target detection method.

[0048] According to another aspect of the present invention, a computer-readable storage medium is provided, the storage medium storing at least one executable instruction, which, when executed on a computer device, causes the computer device to perform the operation of the remote sensing small target detection method.

[0049] This invention provides an embodiment of the method for acquiring an image to be identified; inputting the image to be identified into a remote sensing target detection model to obtain remote sensing targets in the image; the remote sensing target detection model includes a super-resolution branch and a target detection branch; the super-resolution branch and the target detection branch share a parameter-sharing encoder; the target detection branch is used to receive the image to be identified and perform target detection on the image; the super-resolution branch is used to perform iterative training together with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch; and outputs target detection results. This method effectively shares the feature parameters of the super-resolution branch, has an efficient structure, and can accurately focus on small target regions, effectively improving the detection accuracy, especially for small targets.

[0050] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0051] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0052] Figure 1 This diagram illustrates the differences in activation maps of existing object detection models for different tasks.

[0053] Figure 2 A flowchart illustrating the remote sensing small target detection method provided in an embodiment of the present invention is shown;

[0054] Figure 3 This diagram illustrates the structure of remote sensing target detection in the remote sensing small target detection method provided in this embodiment of the invention.

[0055] Figure 4 A schematic diagram of the structure of the remote sensing target detection device provided in an embodiment of the present invention is shown;

[0056] Figure 5 A schematic diagram of the structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0057] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0058] CNN-Based: Since the introduction of SRCNN, super-resolution (SR) methods based on convolutional neural networks (CNNs) have made significant progress in image reconstruction tasks. These methods typically enhance feature representation through attention mechanisms, such as channel attention and global attention to improve feature selection capabilities, while self-similar attention and non-local sparse attention further introduce global dependencies to enhance contextual modeling. However, non-local feature modeling incurs high computational costs, and CNNs, limited by their local receptive field, struggle to simultaneously address detail reconstruction and global semantic modeling. Therefore, a trade-off between efficiency and performance remains in remote sensing image processing scenarios.

[0059] Transformer-Based: Leveraging its self-attention mechanism's advantages in long-range dependency modeling, the Transformer demonstrates stronger feature representation capabilities in super-resolution tasks. The SwinIR model, combining hierarchical feature extraction and reconstruction modules, achieves tasks such as super-resolution, denoising, and compression artifact removal. The TTST model effectively reduces redundant computation and improves reconstruction accuracy through dynamic token selection and multi-scale feedforward layer design. ESTNet significantly reduces computational complexity while maintaining performance through efficient channel attention and residual group mechanisms. Overall, the Transformer architecture achieves a better balance between global feature modeling and efficiency, providing a new approach for remote sensing image super-resolution tasks.

[0060] This invention addresses the problems existing in current technologies by proposing a saliency-driven multi-task collaborative framework, SDCoNet, for small target detection in low-quality remote sensing images. This framework breaks away from the traditional sequential optimization model of "image enhancement – ​​target detection," achieving deep integration of super-resolution and detection tasks through an implicit feature sharing mechanism. This allows the two tasks to collaboratively optimize within a shared semantic feature space, effectively alleviating the problems of feature fragmentation and optimization conflicts. On the NWPU VHR-10-Split dataset, this framework achieves significant performance improvements in detection for small target categories such as storage-tank and vehicle, with mAP as the core evaluation metric, compared to the DINO baseline model and the classic Faster RCNN method. Specifically, during the encoding stage, shared semantic features facilitate mutual optimization between super-resolution and target detection tasks. To enhance the robustness of the framework, this embodiment introduces a saliency guidance mechanism into the network, using a saliency prior map to guide the selection of query tokens in the detection network, helping the model focus on potential target regions in low-quality images. In addition, to mitigate gradient interference during multi-task training, the framework introduces a gradient routing strategy, which separates and guides the gradient flow between tasks during training to maintain the stability and consistency of the overall optimization.

[0061] In summary, the core of the embodiments of the present invention is reflected in the following aspects:

[0062] (1) An implicit feature sharing paradigm is proposed to break through the traditional serial multi-task learning and realize the semantically consistent feature collaboration between super-resolution and detection tasks;

[0063] (2) A saliency-guided query token selection mechanism is introduced, which enables the detection network to dynamically focus on small target regions in complex backgrounds, effectively suppressing feature pollution caused by multi-task coupling;

[0064] (3) Design a gradient routing strategy to alleviate optimization conflicts between multiple tasks, guide the super-resolution branch to generate high-frequency information that is beneficial to the detection task, and thus achieve stable joint optimization between tasks;

[0065] (4) A saliency-driven multi-task collaborative optimization framework was constructed and its effectiveness was verified on multiple public remote sensing datasets. The results showed that the method significantly outperformed existing mainstream algorithms in small target detection of low-quality images, while maintaining good computational efficiency and generalization performance.

[0066] Figure 2 A flowchart of a remote sensing small target detection method provided in an embodiment of the present invention is shown. This method is executed by a computer device. The computer device can be a computer, a smart terminal, a distributed device, a wearable device, a robot, etc., and the embodiments of the present invention do not impose specific limitations. Figure 2 As shown, the method includes the following steps:

[0067] Step 110: Obtain the image to be recognized.

[0068] The image to be identified can be any captured image. This image can be aerial photographs taken by satellites or drones, or satellite bird's-eye views. Remote sensing target detection involves performing AI visual recognition on aerial photographs taken by satellites or drones, or satellite bird's-eye views, to accurately locate, outline, and identify various targets within the image, and to provide the target's location and category.

[0069] Step 120: Input the image to be identified into the remote sensing target detection model to obtain the remote sensing target in the image to be identified.

[0070] Remote sensing target detection is constrained by a fundamental trade-off between restoring small target details and suppressing complex background interference. Traditional sequential paradigms treat super-resolution (SR) and detection as independent stages, inherently resulting in task disconnect. In this separated optimization process, the super-resolution module often generates semantically meaningless pseudo-textures, and subsequent detectors cannot provide feedback to correct these artifacts. To overcome these limitations, constructing a unified framework that deeply couples low-level detail restoration with high-level semantic abstraction is crucial. Therefore, this invention proposes a saliency-guided multi-task collaborative detection framework (SDCoNet), i.e., a remote sensing target detection model. This framework establishes a dual-branch architecture centered on a cross-task parameter-sharing encoder, achieving simultaneous optimization of pixel-level reconstruction and target-level recognition without introducing redundant inference overhead. For example, Figure 3 As shown, the remote sensing target detection model includes a super-resolution branch and a target detection branch; the super-resolution branch and the target detection branch share a parameter-shared encoder; the target detection branch is used to receive the image to be identified and perform target detection on the image; the super-resolution branch is used to perform iterative training together with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch. In this embodiment of the invention, the super-resolution branch includes a parameter-shared encoder, a super-resolution decoder, and an image reconstruction module. The target detection branch also includes the parameter-shared encoder.

[0071] Specifically, the parameter-sharing encoder includes a multi-layer Swin module, which divides the image to be recognized into multiple non-overlapping windows through convolutional layers, performs self-attention calculation within each window, and uses a shift window strategy to capture multi-scale features when crossing windows, thereby obtaining multi-scale feature extraction results.

[0072] In this embodiment of the invention, a parameter-sharing encoder for both the super-resolution branch (SR) and the object detection branch (Det) is constructed using Swin-T (Swin Transformer Tiny). The alternating W-MSA and SW-MSA strategy of Swin-T provides support for feature collaboration between the two tasks. Figure 3As shown in another embodiment, the shared encoder is composed of four stacked Swin modules at different scales, each integrating a patch processing module and an attention calculation unit. First, a convolutional layer processes the input low-resolution image data (the image to be recognized). This parameter-sharing encoder divides the low-resolution image into multiple non-overlapping windows, performs self-attention calculation within each window, and employs a shifting window strategy when crossing windows to capture multi-scale features. Specifically, the Swin-T encoder progressively maps the input image into a high-dimensional feature representation, generating multiple feature maps at different scales with sizes of 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image size, and channel numbers of 96, 192, 384, and 768, respectively. The core of this sharing mechanism lies in achieving feature collaboration between SR and Det: the SR task focuses on details and texture, with its gradient guiding the shallow layers of the encoder to capture edges and local information; the Det task focuses on semantics and category discrimination, with its constraints ensuring that the deep layers of the encoder retain the target structure and semantic features. During training, the network adaptively adjusts the feature extraction tendencies at different levels, enabling the two tasks to complement and collaborate in a unified feature space, effectively alleviating the problem of feature fragmentation in traditional sequential methods.

[0073] The super-resolution decoder includes a multi-level patch expanding module and a residual Swin module. These modules restore spatial resolution through deconvolution and normalization operations during feature amplification and adaptively adjust the feature channel dimension to obtain multi-scale features. In this embodiment, the super-resolution decoder is structurally aligned with the shared encoder, continuing the hierarchical feature modeling approach of the Swin Transformer and combining a U-Net-style layer-by-layer upsampling path to achieve high-precision reconstruction. Specifically, the super-resolution decoder includes a multi-level patch expanding module and a residual Swin module. Each layer restores spatial resolution through deconvolution and normalization operations during feature amplification and adaptively reshapes the feature channel dimension, achieving gradual restoration of fine-grained features and semantic enhancement.

[0074] The super-resolution decoder restores spatial resolution through deconvolution and normalization operations during feature amplification and adaptively adjusts the feature channel dimension to obtain multi-scale features. The specific process is described below:

[0075] ;

[0076] in, Input features; Reshape This indicates that the token sequence is reconstructed into a spatial structure; LN (Layer Normalization) represents the normalization operation; TansConv represents the deconvolution operation, which is used to amplify features in the spatial dimension; the outermost Reshape represents restoring the result to a sequence form.

[0077] In each Swin Transformer block, the input features are first calculated using a self-attention mechanism. The input and output are then represented by an attention function, followed by further nonlinear processing through an MLP layer, and finally, a residual connection is established between the input and output. The specific processing procedure is described below:

[0078] ;

[0079] in, The input features of the block, Block output characteristics This refers to a window-based multi-head self-attention mechanism (including W-MSA or SW-MSA). The MLP is a feedforward neural network module, consisting of two layers of linear transformation and the GELU activation function. This module performs a residual connection after attention calculation and MLP, fusing the original input and the output of each stage by direct addition, thereby achieving deep feature modeling and information preservation.

[0080] The image reconstruction module includes convolutional layers, activation functions, and an upsampling module. These are used to map the multi-scale features output by the super-resolution decoder to the image domain, obtaining a super-resolution sample image. The reconstruction module further maps the multi-scale features output by the decoder to the image domain, achieving high-quality reconstruction. This image reconstruction module extracts low-level spatial features through lightweight convolutional layers and combines them with LeakyReLU activation to enhance nonlinear representation capabilities. Subsequently, it utilizes the PixelShuffle upsampling module for spatial magnification, supporting resolution improvements of 2×, 4×, or 8×. Compared to traditional upsampling methods, this structure performs better in restoring high-frequency details and texture continuity. Finally, a high-resolution output image is generated through convolutional fusion operations, completing the mapping from low-quality input to clear reconstruction. The entire reconstruction process can be represented as follows:

[0081] ;

[0082] in, Activation function This indicates the pixel rearrangement upsampling module. These are the input features for the image reconstruction module. This is the final high-resolution image output.

[0083] In this embodiment of the invention, before inputting the image to be identified into the remote sensing target detection model to obtain the remote sensing target in the image, the remote sensing target detection model is first trained. To resolve the optimization conflict between the two tasks, a "gradient routing strategy" is designed. This strategy divides the training process into two stages: a "semantic stabilization stage" (first stage) and a "synergistic enhancement stage" (second stage). In the first stage, the super-resolution branch is frozen first, and the target detection branch is trained to establish a robust semantic prior. Subsequent joint training routes the gradients of the two tasks to a shared encoder, ensuring that the recovered details are semantically discriminative. Through staged gradient control, synergistic learning of the two tasks is achieved in the training process. The training process in the two stages is as follows: the super-resolution branch is frozen, and the training image is input into the target detection branch for iterative training to obtain a pre-trained target detection branch; the super-resolution branch and the pre-trained target detection branch are jointly iteratively trained to obtain a trained remote sensing target detection model.

[0084] The second phase specifically includes the following steps in the training process:

[0085] Step 001: Input the training image into the parameter sharing encoder to obtain the sample feature extraction result.

[0086] During training, the training images are input into the parameter-shared encoder. The training images are divided into multiple non-overlapping windows through convolutional layers. Self-attention is calculated within each window, and a shift-window strategy is used to capture multi-scale features when crossing windows. This yields multi-scale sample feature extraction results for the sample images, including multiple sample feature maps at multiple scales. The multi-scale features of the extracted sample feature maps are represented as follows: ,in Indicates the first l The feature maps of the layers have different spatial resolutions and number of channels, and contain rich semantic information to support multi-size target detection, thereby enabling accurate detection of small targets in the target detection process.

[0087] After obtaining the sample feature extraction results, the sample feature extraction results are simultaneously input into the super-resolution branch and the pre-trained target detection branch, and steps 002-004 are executed.

[0088] Step 002: Input the sample feature extraction results into the super-resolution decoder and image reconstruction module to obtain the sample image after super-resolution.

[0089] In the process of amplifying the features in the sample feature extraction results, the super-resolution decoder restores the spatial resolution through deconvolution and normalization operations, and performs adaptive adjustments on the feature channel dimension to obtain multi-scale sample features.

[0090] The image reconstruction module further maps the multi-scale sample features output by the super-resolution decoder to the image domain, thereby obtaining the ultra-high resolution sample image.

[0091] Step 003: Input the feature extraction results into the pre-trained target detection branch to obtain sample target boxes.

[0092] In this embodiment of the invention, the DINO framework is selected as the baseline detector for the target detection branch because of its robust convergence characteristics. This target detection branch employs a saliency-guided mechanism to dynamically filter query tokens. This filtering strategy focuses computational resources on key regions, thereby enhancing the perception capability of small targets.

[0093] Specifically, the feature extraction results are input into the pre-trained target detection branch through the following steps to obtain sample target boxes:

[0094] a. The global features in the multi-scale feature extraction results output by the parameter-shared encoder are subjected to sequence average pooling, then broadcasted and expanded, and concatenated with local features to obtain fused features.

[0095] The input multi-scale feature extraction results are split into local and global features along the channel dimension. Local features are retained in their original form, while global features are broadcast expanded and concatenated after sequence average pooling.

[0096] b. Input the fused features into a three-layer feedforward network for layer-by-layer dimensionality reduction to obtain the saliency score map of each channel.

[0097] The fused features are input into a three-layer feedforward network (including linear transformation and GELU activation), and dimensionality is reduced layer by layer to output a single-channel saliency score map, where each value represents the retention probability or attention intensity at the corresponding location. The specific process is as follows:

[0098] ;

[0099] in, The saliency score map corresponding to the layer feature map. It is a lightweight feedforward network module (including layer normalization, linear mapping and GELU). This indicates a local-to-global fusion operation. It is a three-layer dimensionality reduction network.

[0100] c. Generate an initial saliency map based on the saliency score map corresponding to the top-level feature map, and pass it down to the lower layers to obtain a cross-scale saliency map.

[0101] In this embodiment of the invention, the first l The saliency score of each layer feature map is denoted as... ,

[0102] A top-down saliency propagation mechanism is constructed: top-level features generate an initial saliency map, which is then propagated to lower levels, forming a cross-scale enhancement path. For l-1 The formula for calculating the saliency plot of a layer across scales is as follows:

[0103] ;

[0104] in, These are learnable modulation coefficients. This indicates a bilinear interpolation upsampling operation, which provides foreground confidence guidance for multi-level query selection.

[0105] d. Calculate the original target image coordinates corresponding to the query vector of the target location in each feature map layer.

[0106] In this embodiment of the invention, a scale-independent saliency confidence score based on relative geometric position is used as a supervision signal to measure the spatial relevance between the saliency query vector Query and the true target.

[0107] Among them, the l The query vector Query for the position (i,j) in the layer feature map corresponds to the original image coordinates, and is represented as:

[0108] .

[0109] e. Calculate the saliency score of the query vector based on the original target image coordinates and the ground truth bounding box.

[0110] The significance score of the query vector, also known as the significance confidence score, is defined as:

[0111] ;

[0112] in, Represents the true bounding box. To query the relative distance between the vector and the center of the true bounding box, The scale parameter is adjustable. This continuous, scale-independent supervision label uses a Gaussian decay function to provide a smooth distance-sensitive confidence distribution, avoiding numerical instability (such as negative value problems) that may be caused by the original linear decay. It also better captures boundary gradations through exponential form, improving the robustness and training stability of small target detection. At the same time, it is significantly different from traditional discrete foreground / background labeling, avoiding bias towards large targets.

[0113] f. Based on the saliency score, the query vectors at each level and scale are filtered through a hierarchical filtering mechanism to obtain the target query vector set.

[0114] To reduce computational overhead, this embodiment of the invention employs a hierarchical query vector filtering mechanism, which integrates layer-wise and level-wise filtering strategies. This fully leverages the characteristics of low resolution and rich semantics of high-level features, and by dynamically adjusting the query filtering ratio, prioritizes high-salience regions, thereby improving attention efficiency.

[0115] The t-th Transformer encoding layer and the t-th Each feature layer introduces a filtering ratio. Only the first one is retained. Attention calculation is performed on N queries:

[0116] ;

[0117] in, For the set of query vectors to be filtered, For attention update operations, The position of the query vector is encoded. When using Deformable Attention, the computational complexity decreases from O(NHK) to... This reduces the number of effective query vectors to the original number. .

[0118] In one specific implementation of the present invention, the top scores are selected based on the salience scores. N tokens constitute Other tokens are blocked. This invention, through a hierarchical filtering mechanism, optimizes the utilization of high semantic features and significantly improves detection performance.

[0119] To mitigate the semantic misalignment of unselected query vectors, this embodiment of the invention designs a background embedding compensation mechanism, defining a learnable row / column embedding matrix. For a position (i,j) in the feature map, the background embedding can be represented as:

[0120] ;

[0121] The embedding mechanism operates only on unselected query vectors (Query), fusing them with the original output through element-by-element addition to ensure semantic consistency. This background embedding compensation mechanism improves detection accuracy and optimizes computational efficiency through positional anchoring and semantic compensation.

[0122] g. After performing attention calculation based on the target query vector set, perform saliency prediction to obtain the sample target box.

[0123] In this process, attention calculation is performed based on the filtered target query vector set, and the target detection results of the samples are generated through the prediction module.

[0124] Step 004: Based on the ultra-high resolution sample image and sample target bounding box, and the preset multi-task joint loss function, adjust the parameters, and continue to execute the steps of inputting the training image into the parameter sharing encoder to obtain the feature extraction result, inputting the feature extraction result into the super-resolution decoder and the pre-trained target detection branch respectively to obtain the ultra-high resolution sample image and sample target bounding box, and adjusting the parameters based on the ultra-high resolution sample image and sample target bounding box, and the preset multi-task joint loss function, until a trained remote sensing target detection model is obtained.

[0125] During joint training, collaborative learning is performed on the super-resolution branch and the pre-trained object detection branch. The parameter adjustment based on the super-resolution sample images and sample bounding boxes, as well as the preset multi-task joint loss function, includes:

[0126] Set the target detection loss as follows:

[0127] ;

[0128] in, These are the losses from conventional detection. For each weight, For significance constraint loss, Loss weights are used to constrain significance.

[0129] Set super-resolution loss The super-resolution loss Pixel reconstruction constraints are applied:

[0130] ;

[0131] in, Constraint weights for pixel reconstruction Measure the pixel difference between the reconstructed image and the high-resolution label;

[0132] The multi-task joint loss function Defined as:

[0133] .

[0134] The embodiments of the present invention, through the above-mentioned joint optimization framework and through the above-mentioned multi-task joint loss function and gradient staged control during training, enable low-level reconstruction features and high-level detection semantics to promote each other during training, realize feature complementarity and dynamic collaboration, and finally obtain a well-trained remote sensing target detection model.

[0135] Step 130: Output the target detection results.

[0136] Based on the aforementioned remote sensing target detection model, the target detection results in the image to be identified are output.

[0137] This invention provides an embodiment of the method for acquiring an image to be identified; inputting the image to be identified into a remote sensing target detection model to obtain remote sensing targets in the image; the remote sensing target detection model includes a super-resolution branch and a target detection branch; the super-resolution branch and the target detection branch share a parameter-sharing encoder; the target detection branch is used to receive the image to be identified and perform target detection on the image; the super-resolution branch is used to perform iterative training together with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch; and outputs target detection results. This method effectively shares the feature parameters of the super-resolution branch, has an efficient structure, and can accurately focus on small target regions, effectively improving the accuracy of target detection, especially for small targets.

[0138] Figure 4 A schematic diagram of the structure of the remote sensing small target detection device provided in an embodiment of the present invention is shown. Figure 4 As shown, the device 300 includes:

[0139] The acquisition module 310 is used to acquire the image to be recognized;

[0140] The target detection module 320 is used to input the image to be identified into a remote sensing target detection model to obtain the remote sensing target in the image to be identified; the remote sensing target detection model includes a super-resolution branch and a target detection branch; the super-resolution branch and the target detection branch share a parameter-shared encoder; the target detection branch is used to receive the image to be identified and perform target detection on the image to be identified; the super-resolution branch is used to perform iterative training together with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch;

[0141] Output module 330 is used to output the target detection results.

[0142] The working process of each module in the remote sensing small target detection device of this invention is largely the same as the specific implementation of the steps in the method of the foregoing embodiments, and will not be repeated here.

[0143] This invention provides an embodiment of the method for acquiring an image to be identified; inputting the image to be identified into a remote sensing target detection model to obtain remote sensing targets in the image; the remote sensing target detection model includes a super-resolution branch and a target detection branch; the super-resolution branch and the target detection branch share a parameter-sharing encoder; the target detection branch is used to receive the image to be identified and perform target detection on the image; the super-resolution branch is used to perform iterative training together with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch; and outputs target detection results. This method effectively shares the feature parameters of the super-resolution branch, has an efficient structure, and can accurately focus on small target regions, effectively improving the accuracy of target detection, especially for small targets.

[0144] Figure 5 The diagram shows a structural schematic of a computer device provided in an embodiment of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computer device.

[0145] like Figure 5 As shown, the computer device may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0146] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408. Communication interface 404 is used to communicate with other network elements such as clients or other servers. The processor 402 executes program 410, specifically performing the relevant steps described in the embodiment of the remote sensing small target detection method.

[0147] Specifically, program 410 may include program code, which includes computer-executable instructions.

[0148] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0149] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0150] Specifically, program 410 can be called by processor 402 to cause the computer device to perform the following operations:

[0151] Acquire the image to be recognized;

[0152] The image to be identified is input into a remote sensing target detection model to obtain the remote sensing target in the image to be identified; the remote sensing target detection model includes a super-resolution branch and a target detection branch; the super-resolution branch and the target detection branch share a parameter-shared encoder; the target detection branch is used to receive the image to be identified and perform target detection on the image to be identified; the super-resolution branch is used to perform iterative training together with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch;

[0153] Output the target detection results.

[0154] In one alternative approach, the super-resolution branch includes a parameter-sharing encoder, a super-resolution decoder, and an image reconstruction module; before inputting the image to be identified into the remote sensing target detection model to obtain the remote sensing target in the image to be identified, the method further includes:

[0155] Freeze the super-resolution branch, input the training image into the target detection branch for iterative training, and obtain the pre-trained target detection branch;

[0156] The super-resolution branch and the pre-trained target detection branch are trained together iteratively to obtain a trained remote sensing target detection model.

[0157] In one alternative approach, the super-resolution branch includes a parameter-shared encoder, a super-resolution decoder, and an image reconstruction module; the super-resolution branch and the pre-trained target detection branch are iteratively trained together to obtain a trained remote sensing target detection model, including:

[0158] The training images are input into the parameter-sharing encoder to obtain the sample feature extraction results;

[0159] The sample feature extraction results are input into the super-resolution decoder and image reconstruction module to obtain the sample image after super-resolution;

[0160] The feature extraction results are input into the pre-trained target detection branch to obtain sample target boxes;

[0161] Based on the ultra-high resolution sample image and sample target bounding box, and the preset multi-task joint loss function, the parameters are adjusted, and the process continues to execute the steps of inputting the training image into the parameter sharing encoder to obtain the feature extraction result, inputting the feature extraction result into the super-resolution decoder and the pre-trained target detection branch respectively to obtain the ultra-high resolution sample image and sample target bounding box respectively, and adjusting the parameters based on the ultra-high resolution sample image and sample target bounding box, and the preset multi-task joint loss function, until a trained remote sensing target detection model is obtained.

[0162] In one optional approach, the parameter adjustment based on the ultra-high resolution sample image and sample target bounding box, and a preset multi-task joint loss function, includes:

[0163] Set the target detection loss as follows:

[0164] ;

[0165] in, These are the losses from conventional detection. For each weight, For significance constraint loss, Loss weights are used to constrain significance.

[0166] Set super-resolution loss The super-resolution loss Pixel reconstruction constraints are applied:

[0167] ;

[0168] in, Constraint weights for pixel reconstruction Measure the pixel difference between the reconstructed image and the high-resolution label;

[0169] The multi-task joint loss function Defined as:

[0170] .

[0171] In one alternative approach, the parameter-sharing encoder includes a multi-layer Swin module, which divides the image to be recognized into multiple non-overlapping windows through convolutional layers, performs self-attention calculation within each window, and employs a shift window strategy to capture multi-scale features when crossing windows, thereby obtaining multi-scale feature extraction results.

[0172] The super-resolution decoder includes a multi-level patch extension module and a residual Swin module, which are used to restore spatial resolution through deconvolution and normalization operations during feature amplification, and to adaptively adjust the feature channel dimension to obtain multi-scale features.

[0173] The image reconstruction module includes a convolutional layer, an activation function, and an upsampling module, which are used to map the multi-scale features output by the super-resolution decoder to the image domain to obtain a sample image with ultra-high resolution.

[0174] In one alternative approach, the feature amplification process restores spatial resolution through deconvolution and normalization operations, and adaptively adjusts the feature channel dimension to obtain multi-scale features, including:

[0175] ;

[0176] The input features are represented by 'Reshape'; 'LN' represents the normalization operation; 'TansConv' represents the deconvolution operation, used to amplify features in the spatial dimension; and 'Reshape' represents restoring the result to a sequence form.

[0177] In one alternative approach, inputting the feature extraction results into the pre-trained target detection branch to obtain sample target boxes includes:

[0178] The global features in the multi-scale feature extraction results output by the parameter-sharing encoder are subjected to sequence average pooling and then broadcast expansion and concatenated with local features to obtain fused features.

[0179] The fused features are input into a three-layer feedforward network for layer-by-layer dimensionality reduction to obtain the saliency score map of each channel;

[0180] An initial saliency map is generated based on the saliency score map corresponding to the top-level feature map, and then passed down to lower levels to obtain a cross-scale saliency map.

[0181] Calculate the original target image coordinates corresponding to the query vector of the target location in each feature map layer;

[0182] The saliency score of the query vector is calculated based on the original target image coordinates and the ground truth bounding box.

[0183] Based on the saliency score, a hierarchical filtering mechanism is used to filter the query vectors at each level and scale to obtain the target query vector set.

[0184] After performing attention calculation based on the target query vector set, saliency prediction is performed to obtain the sample target box.

[0185] This invention provides an embodiment of the method for acquiring an image to be identified; inputting the image to be identified into a remote sensing target detection model to obtain remote sensing targets in the image; the remote sensing target detection model includes a super-resolution branch and a target detection branch; the super-resolution branch and the target detection branch share a parameter-sharing encoder; the target detection branch is used to receive the image to be identified and perform target detection on the image; the super-resolution branch is used to perform iterative training together with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch; and outputs target detection results. This method effectively shares the feature parameters of the super-resolution branch, has an efficient structure, and can accurately focus on small target regions, effectively improving the accuracy of target detection, especially for small targets.

[0186] This invention provides a computer-readable storage medium storing at least one executable instruction that, when executed on a computer device, causes the computer device to perform the remote sensing small target detection method in any of the above method embodiments.

[0187] Executable instructions can be used to cause a computer device to perform the following operations:

[0188] In one alternative approach, the super-resolution branch includes a parameter-sharing encoder, a super-resolution decoder, and an image reconstruction module; before inputting the image to be identified into the remote sensing target detection model to obtain the remote sensing target in the image to be identified, the method further includes:

[0189] Freeze the super-resolution branch, input the training image into the target detection branch for iterative training, and obtain the pre-trained target detection branch;

[0190] The super-resolution branch and the pre-trained target detection branch are trained together iteratively to obtain a trained remote sensing target detection model.

[0191] In one alternative approach, the super-resolution branch includes a parameter-shared encoder, a super-resolution decoder, and an image reconstruction module; the super-resolution branch and the pre-trained target detection branch are iteratively trained together to obtain a trained remote sensing target detection model, including:

[0192] The training images are input into the parameter-sharing encoder to obtain the sample feature extraction results;

[0193] The sample feature extraction results are input into the super-resolution decoder and image reconstruction module to obtain the sample image after super-resolution;

[0194] The feature extraction results are input into the pre-trained target detection branch to obtain sample target boxes;

[0195] Based on the ultra-high resolution sample image and sample target bounding box, and the preset multi-task joint loss function, the parameters are adjusted, and the process continues to execute the steps of inputting the training image into the parameter sharing encoder to obtain the feature extraction result, inputting the feature extraction result into the super-resolution decoder and the pre-trained target detection branch respectively to obtain the ultra-high resolution sample image and sample target bounding box respectively, and adjusting the parameters based on the ultra-high resolution sample image and sample target bounding box, and the preset multi-task joint loss function, until a trained remote sensing target detection model is obtained.

[0196] In one optional approach, the parameter adjustment based on the ultra-high resolution sample image and sample target bounding box, and a preset multi-task joint loss function, includes:

[0197] Set the target detection loss as follows:

[0198] ;

[0199] in, These are the losses from conventional detection. For each weight, For significance constraint loss, Loss weights are used to constrain significance.

[0200] Set super-resolution loss The super-resolution loss Pixel reconstruction constraints are applied:

[0201] ;

[0202] in, Constraint weights for pixel reconstruction Measure the pixel difference between the reconstructed image and the high-resolution label;

[0203] The multi-task joint loss function Defined as:

[0204] .

[0205] In one alternative approach, the parameter-sharing encoder includes a multi-layer Swin module, which divides the image to be recognized into multiple non-overlapping windows through convolutional layers, performs self-attention calculation within each window, and employs a shift window strategy to capture multi-scale features when crossing windows, thereby obtaining multi-scale feature extraction results.

[0206] The super-resolution decoder includes a multi-level patch extension module and a residual Swin module, which are used to restore spatial resolution through deconvolution and normalization operations during feature amplification, and to adaptively adjust the feature channel dimension to obtain multi-scale features.

[0207] The image reconstruction module includes a convolutional layer, an activation function, and an upsampling module, which are used to map the multi-scale features output by the super-resolution decoder to the image domain to obtain a sample image with ultra-high resolution.

[0208] In one alternative approach, the feature amplification process restores spatial resolution through deconvolution and normalization operations, and adaptively adjusts the feature channel dimension to obtain multi-scale features, including:

[0209] ;

[0210] The input features are represented by 'Reshape'; 'LN' represents the normalization operation; 'TansConv' represents the deconvolution operation, used to amplify features in the spatial dimension; and 'Reshape' represents restoring the result to a sequence form.

[0211] In one alternative approach, inputting the feature extraction results into the pre-trained target detection branch to obtain sample target boxes includes:

[0212] The global features in the multi-scale feature extraction results output by the parameter-sharing encoder are subjected to sequence average pooling and then broadcast expansion and concatenated with local features to obtain fused features.

[0213] The fused features are input into a three-layer feedforward network for layer-by-layer dimensionality reduction to obtain the saliency score map of each channel;

[0214] An initial saliency map is generated based on the saliency score map corresponding to the top-level feature map, and then passed down to lower levels to obtain a cross-scale saliency map.

[0215] Calculate the original target image coordinates corresponding to the query vector of the target location in each feature map layer;

[0216] The saliency score of the query vector is calculated based on the original target image coordinates and the ground truth bounding box.

[0217] Based on the saliency score, a hierarchical filtering mechanism is used to filter the query vectors at each level and scale to obtain the target query vector set.

[0218] After performing attention calculation based on the target query vector set, saliency prediction is performed to obtain the sample target box.

[0219] This invention provides an embodiment of the method for acquiring an image to be identified; inputting the image to be identified into a remote sensing target detection model to obtain remote sensing targets in the image; the remote sensing target detection model includes a super-resolution branch and a target detection branch; the super-resolution branch and the target detection branch share a parameter-sharing encoder; the target detection branch is used to receive the image to be identified and perform target detection on the image; the super-resolution branch is used to perform iterative training together with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch; and outputs target detection results. This method effectively shares the feature parameters of the super-resolution branch, has an efficient structure, and can accurately focus on small target regions, effectively improving the detection accuracy, especially for small targets.

[0220] This invention provides a remote sensing target detection device for performing the above-described remote sensing small target detection method.

[0221] This invention provides a computer program that can be called by a processor to enable a computer device to execute the remote sensing small target detection method in any of the above method embodiments.

[0222] This invention provides a computer program product, which includes a computer program stored on a computer-readable storage medium. The computer program includes program instructions, which, when executed on a computer, cause the computer to perform the remote sensing small target detection method in any of the above method embodiments.

[0223] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of the present invention are not directed to any particular programming language. It should be understood that the content of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0224] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0225] Similarly, it should be understood that, in order to streamline the invention and aid in understanding one or more of the various aspects of the invention, features of the embodiments of the invention are sometimes grouped together in a single embodiment, figure, or description thereof in the above description of exemplary embodiments of the invention. However, this disclosure should not be construed as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim.

[0226] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0227] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

Claims

1. A method for detecting small targets in remote sensing, characterized in that, The method includes: Acquire the image to be recognized; The method involves inputting the image to be identified into a remote sensing target detection model to obtain remote sensing targets in the image. The remote sensing target detection model includes a super-resolution branch and a target detection branch. The super-resolution branch and the target detection branch share a parameter-sharing encoder. The target detection branch receives the image to be identified and performs target detection on it. The super-resolution branch iterates with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch. A Swin-T parameter-sharing encoder for the dual tasks of the super-resolution branch and the target detection branch is constructed. The method further includes: inputting the training image into the parameter-sharing encoder to obtain sample feature extraction results; inputting the sample feature extraction results into a super-resolution decoder and an image reconstruction module to obtain a super-resolution sample image; inputting the feature extraction results into a pre-trained target detection branch to obtain sample target boxes; adjusting parameters based on the super-resolution sample image and sample target boxes, and a preset multi-task joint loss function, and continuing iterative training until a well-trained remote sensing target detection model is obtained. The step of inputting the feature extraction results into the pre-trained target detection branch to obtain sample target boxes includes: performing sequence average pooling on the global features in the multi-scale feature extraction results output by the parameter-shared encoder, then broadcasting and expanding them and concatenating them with local features to obtain fused features; inputting the fused features into a three-layer feedforward network for layer-by-layer dimensionality reduction to obtain saliency score maps for each channel; generating an initial saliency map based on the saliency score map corresponding to the top-level feature map and passing it to lower layers to obtain cross-scale saliency maps; calculating the original target image coordinates corresponding to the query vector of the target position in each layer of the feature map; calculating the saliency score of the query vector based on the original target image coordinates and the ground truth bounding box; filtering the query vectors of each layer and each scale based on the saliency score through a hierarchical filtering mechanism to obtain a target query vector set; performing attention updates based on the target query vector set and generating the final sample target boxes through the detection head; Output the target detection results.

2. The method according to claim 1, characterized in that, The super-resolution branch includes a parameter-shared encoder, a super-resolution decoder, and an image reconstruction module; before inputting the image to be identified into the remote sensing target detection model to obtain the remote sensing target in the image to be identified, the method further includes: Freeze the super-resolution branch, input the training image into the target detection branch for iterative training, and obtain the pre-trained target detection branch; The super-resolution branch and the pre-trained target detection branch are trained together iteratively to obtain a trained remote sensing target detection model.

3. The method according to claim 1, characterized in that, The parameter adjustment based on the ultra-high resolution sample image and sample target bounding box, and the preset multi-task joint loss function includes: Set the target detection loss as follows: ; in, These are the losses from conventional detection. For each weight, For significance constraint loss, Loss weights are used to constrain significance. Set super-resolution loss The super-resolution loss Pixel reconstruction constraints are applied: ； in, Constraint weights for pixel reconstruction Measure the pixel difference between the reconstructed image and the high-resolution label; The multi-task joint loss function Defined as: .

4. The method according to claim 1, characterized in that, The parameter-sharing encoder includes a multi-layer Swin module, which divides the image to be recognized into multiple non-overlapping windows through convolutional layers, performs self-attention calculation within each window, and uses a shift window strategy to capture multi-scale features when crossing windows, thereby obtaining multi-scale feature extraction results. The super-resolution decoder includes a multi-level patch extension module and a residual Swin module, which are used to restore spatial resolution through deconvolution and normalization operations during feature amplification, and to adaptively adjust the feature channel dimension to obtain multi-scale features. The image reconstruction module includes a convolutional layer, an activation function, and an upsampling module, which are used to map the multi-scale features output by the super-resolution decoder to the image domain to obtain a sample image with ultra-high resolution.

5. The method according to claim 4, characterized in that, The feature amplification process restores spatial resolution through deconvolution and normalization operations, and adaptively adjusts the feature channel dimension to obtain multi-scale features, including: ; For input features; This indicates that the labeled sequence is reconstructed into a spatial structure; LN represents the normalization operation; and TansConv represents the deconvolution operation, which is used to amplify features in the spatial dimension.

6. A remote sensing small target detection device, characterized in that, The device includes: The acquisition module is used to acquire the image to be recognized; A target detection module is used to input the image to be identified into a remote sensing target detection model to obtain remote sensing targets in the image to be identified. The remote sensing target detection model includes a super-resolution branch and a target detection branch. The super-resolution branch and the target detection branch share a parameter-sharing encoder. The target detection branch is used to receive the image to be identified and perform target detection on the image. The super-resolution branch is used to perform iterative training together with the target detection branch during the training phase of the remote sensing target detection model to adjust the parameters of the target detection branch. A Swin-T parameter-sharing encoder for the dual tasks of the super-resolution branch and the target detection branch is constructed. The training of the remote sensing target detection model includes: inputting the training image into the parameter-sharing encoder to obtain sample feature extraction results; inputting the sample feature extraction results into the super-resolution decoder and image reconstruction module to obtain a super-resolution sample image; inputting the feature extraction results into the pre-trained target detection branch to obtain sample target boxes; adjusting parameters based on the super-resolution sample image and sample target boxes, and a preset multi-task joint loss function, and continuing iterative training until a trained remote sensing target detection model is obtained. The step of inputting the feature extraction results into the pre-trained target detection branch to obtain sample target boxes includes: performing sequence average pooling on the global features in the multi-scale feature extraction results output by the parameter-shared encoder, then broadcasting and expanding them and concatenating them with local features to obtain fused features; inputting the fused features into a three-layer feedforward network for layer-by-layer dimensionality reduction to obtain saliency score maps for each channel; generating an initial saliency map based on the saliency score map corresponding to the top-level feature map and passing it to lower layers to obtain cross-scale saliency maps; calculating the original target image coordinates corresponding to the query vector of the target position in each layer of the feature map; calculating the saliency score of the query vector based on the original target image coordinates and the ground truth bounding box; filtering the query vectors of each layer and each scale based on the saliency score through a hierarchical filtering mechanism to obtain a target query vector set; performing attention updates based on the target query vector set and generating the final sample target boxes through the detection head; The output module is used to output the target detection results.

7. A computer device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation of the remote sensing small target detection method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The storage medium stores at least one executable instruction, which, when executed on a computer device, causes the computer device to perform the operation of the remote sensing small target detection method as described in any one of claims 1-5.

Images