Dynamic alignment and edge feature calibration method for small and medium-sized target detection in remote sensing image

Abstract

A dynamic alignment and edge feature calibration method for small and medium-sized target detection in a remote sensing image comprises the following steps: S100, in a feature extraction stage, extracting multi-scale texture information by adopting multi-scale texture sensing (MSTP), capturing a relationship among pixels, and simultaneously amplifying homogeneous details of an object to represent important textures; s200, in the feature processing stage, useful multi-scale features are reserved by selecting a BS-FPN composed of boundary enhancement DBS and C2f, and redundant features are further filtered; and S300, inputting the processed features into a task perception detection head TAD-Head for processing, and comprehensively integrating supervision information from a classification task and a positioning task through a cross-task perception mechanism to realize consistency detection of classification and positioning. According to the method, the problem of weak and small targets in the remote sensing image is solved by exploring high-frequency detail information characteristics of an object and enhancing the cross-task perception interaction capability.

Description

Technical Field

[0001] This disclosure belongs to the field of remote sensing image processing technology, and specifically relates to a dynamic alignment and edge feature calibration method for small target detection in remote sensing images. Background Technology

[0002] Aerial remote sensing imagery (ARSI) is a valuable data source for remote sensing target detection (RSOD). Its high-resolution images have significant application value in areas such as real-time urban traffic monitoring, urban disaster prevention and control, military applications, and target tracking. Target detection technology, which accurately identifies and locates key scenes and targets in remote sensing images, has become one of the key technologies for intelligent interpretation of remote sensing images.

[0003] Currently, in remote sensing image detection research, similar to visible light target detection methods, the remote sensing image detection pipeline framework can be divided into two categories: "two-stage" and "one-stage." The "two-stage" algorithm mainly combines coarse and fine detection through a region proposal network to achieve high-precision prediction. The "one-stage" algorithm directly detects targets through dense prediction, and its inference speed far exceeds that of the "two-stage" method; moreover, its speed and accuracy have made significant progress in recent years. Although these remote sensing target detection (RSOD) methods have achieved significant improvements in accuracy, the detection of small targets in high-altitude remote sensing images still faces many challenges due to the significant differences between high-altitude remote sensing images and visible light images. In the same scene, remote sensing images exhibit large variations in object scale, low resolution of most targets, significant changes in high-frequency features, and susceptibility to imaging conditions and background noise interference. Summary of the Invention

[0004] To address the aforementioned problems, this disclosure provides a dynamic alignment and edge feature calibration method for small target detection in remote sensing images, comprising: S100: In the feature extraction stage, multi-scale texture perception MSTP is used to extract multi-scale texture information, capture the relationship between pixels, and at the same time magnify the homogeneous details of the object to represent important textures. S200: In the feature processing stage, useful multi-scale features are preserved by selecting BS-FPN composed of boundary enhancement DBS and C2f, and redundant features are further filtered out. S300: The processed features are input into the Task Aware Detection Head (TAD-Head) for processing. Through a cross-task awareness mechanism, the supervision information from both classification and localization tasks is fully integrated to achieve consistency detection between classification and localization.

[0005] Furthermore, this invention also discloses a dynamic alignment and edge feature calibration device for small target detection in remote sensing images, comprising: An apparatus for extracting multi-scale texture information using a multi-scale texture-aware MSTP during the feature extraction stage, capturing the relationships between pixels, and simultaneously magnifying the homogeneous details of an object to characterize important textures. An apparatus for retaining useful multi-scale features and further filtering redundant features by selecting a BS-FPN composed of boundary enhancement DBS and C2f during the feature processing stage. This device is used to input the processed features into the Task Aware Detection Head (TAD-Head) for processing, and to fully integrate the supervision information from both classification and localization tasks through a cross-task awareness mechanism to achieve consistency detection between classification and localization.

[0006] Furthermore, the present invention also discloses a computer storage medium, wherein the storage medium includes computer instructions that, when run on a computer, cause the computer to perform the method described thereon.

[0007] Furthermore, the present invention also discloses an electronic device, wherein the electronic device comprises: Memory, processor, and computer programs stored in memory and executable on the processor, wherein, The processor implements the method when executing the program.

[0008] This paper proposes a novel one-stage detector, DAEC, specifically designed to address the challenge of small target detection in remote sensing images. First, to explore the texture details of objects, MSTP is developed to improve the key texture representation of fundamental features. MSTP captures pixel correlations in the image, enhancing the internal pixel relationships between objects. Second, a DBS method is proposed to adaptively aggregate boundary and semantic information by learning spatial weighted modulated features, achieving finer-grained recalibration of target contours and object locations. Furthermore, this paper proposes TAD-Head to facilitate interaction between classification and localization tasks, introducing additional supervision information for each task, thereby generating more consistent predictions. Attached Figure Description

[0009] Figure 1 This is a flowchart of a method for dynamic alignment and edge feature calibration of small targets in remote sensing images provided in one embodiment of this disclosure; Figure 2 This is a diagram of the overall architecture of the DAEC-Detector provided in one embodiment of this disclosure; Figure 3 This is an architecture diagram of MSTP provided in one embodiment of this disclosure; Figure 4 This is a DBS architecture diagram provided in one embodiment of this disclosure; Figure 5This is a schematic diagram of a parallel head and a TAD-head provided in one embodiment of this disclosure. Detailed Implementation

[0010] To further describe the present invention, the following description is in conjunction with the accompanying drawings. Figures 1 to 5 Further explanation is needed.

[0011] In one embodiment, such as Figure 1 As shown, this disclosure provides a dynamic alignment and edge feature calibration method for small target detection in remote sensing images, which includes: S100: In the feature extraction stage, multi-scale texture perception MSTP is used to extract multi-scale texture information, capture the relationship between pixels, and at the same time magnify the homogeneous details of the object to represent important textures. S200: In the feature processing stage, useful multi-scale features are preserved by selecting BS-FPN composed of boundary enhancement DBS and C2f, and redundant features are further filtered out. S300: The processed features are input into the Task Aware Detection Head (TAD-Head) for processing. Through a cross-task awareness mechanism, the supervision information from both classification and localization tasks is fully integrated to achieve consistency detection between classification and localization.

[0012] In this embodiment, the overall framework of the proposed Dynamic Alignment and Edge Feature Correction Detector (DAEC-Detector) based on the YOLOv8 model is as follows: Figure 2 As shown, DAEC consists of three basic components: Multi-Scale Texture Awareness (MSTP), a novel neck boundary selection pyramid network (BS-FPN) composed of boundary selection enhancement (DBS), and a task-aware detection head (TAD-Head). In the feature extraction stage, because target region features are easily interfered with by background information after a series of convolutional layers, resulting in the loss of target feature information, MSTP is used to extract multi-scale texture information, capture the relationships between pixels, and simultaneously amplify homogeneous details of the object to represent important textures. In the feature processing stage, BS-FPN, composed of DBS and C2f, retains useful multi-scale features and further filters redundant features. DBS is used to aggregate boundary features from the semantic information of low-level and high-level features to better preserve boundary detail features and recalibrate the object. Finally, the processed features are input into the TAD-Head for further processing, comprehensively integrating supervision information from two tasks through a cross-task awareness mechanism to achieve consistent detection between classification and localization.

[0013] Compared to the serial feature processing architecture, the MSTP proposed in this paper introduces a parallel structure to process multi-scale features independently. This design avoids information attenuation during layer-by-layer downsampling, ensuring that each scale branch can directly extract information from the original feature map, thus preserving the gradient integrity at different scales to the maximum extent. However, while this parallel fusion mechanism enhances effective features, it may exacerbate interference from redundant information in the background region. To address this, this method integrates a standard spatial attention branch into the parallel feature fusion module, achieving dual optimization of noise filtering and key information focusing by adaptively suppressing the response of non-critical regions and enhancing the feature representation of the target region.

[0014] Compared to the traditional FPN structure, this method replaces the simple concatenation operation in the original baseline Neck with a DBS module, thus constructing a novel Boundary Selection Pyramid Network (BS-FPN). DBS introduces a bidirectional fusion mechanism between high-resolution and low-resolution features, enabling more thorough information transfer between features and further improving the multi-scale feature fusion effect. Furthermore, by learning spatial weight modulation features, it adaptively aggregates boundary and semantic information to depict finer-grained object contours and recalibrate object positions.

[0015] To address the issues of feature conflict and insufficient interaction in traditional detection heads during multi-task processing, this method proposes a novel TAD-Head detection head. This head aims to achieve efficient interaction between classification and localization tasks and optimizes prediction consistency by introducing additional supervisory information. The dynamic alignment in TAD-Head is mainly reflected in two aspects: the Task Decomposition Module (TDM) uses a layer attention mechanism to dynamically allocate features, allowing classification and localization tasks to selectively focus on features from different layers as needed, avoiding conflicts and achieving initial alignment of task features; the prediction alignment step dynamically adjusts the bounding box position and classification confidence by generating a spatial offset map (O) and a spatial probability map (M), and then performs element-wise multiplication of M / O with the task interaction features (Z) to achieve fine spatial and semantic alignment of classification and localization results. These two aspects work together to achieve dynamic alignment of classification and localization tasks, improving detection performance.

[0016] The input features in S100 are generated by the deep Backbone layer (Backbone's C4 layer). Multi-Scale Texture Perception (MSTP) is used to extract multi-scale texture information of the target and background in the input image, capturing the relationships between target-related pixels (including pixels in the target itself and the surrounding background that may affect target differentiation), while magnifying regions or details within the target objects that have similar attributes (such as grayscale, texture, and color) that need to be identified and located in the detection task.

[0017] The input to the BS-FPN in S200 is the feature extraction stage (outputting multi-scale feature maps from different layers of the backbone network). A boundary selection pyramid network (BS-FPN) composed of Dynamic Boundary Selection (DBS) and C2f (CSPBottleneck with 2 Convolutions) preserves useful multi-scale features and further filters redundant features. In the YOLOv8 network architecture, the C2f module is a key component for implementing Cross Stage Partial Fusion.

[0018] The input to S300 is a multi-scale feature map processed by BS-FPN (composed of Dynamic Boundary Selection DBS and C2f). The processed features are then input into the Task-Aware Detection Head (TAD-Head) for further processing. Through a cross-task awareness mechanism, it comprehensively integrates supervision information from classification and localization tasks to achieve consistent detection between classification and localization.

[0019] In another embodiment, the MSTP consists of a multi-scale feature extraction branch, an edge information enhancer (EIE), and a spatial attention mechanism (SAM).

[0020] In this embodiment, texture information, as a key visual element, reflects the organizational structure and arrangement pattern of an object's surface. These features often exhibit a gradual or periodic variation. However, when processing small objects, the presence of numerous irrelevant signals in the image and the small pixel count of the object itself greatly suppress subtle texture information. Extracting details directly at the input resolution, however, would result in excessive computation, thus reducing inference speed.

[0021] Therefore, inspired by MDSAM, this method proposes a multi-scale texture-aware (MSTP) approach to amplify the texture representation of objects. MSTP consists of a multi-scale feature extraction branch, an edge information enhancer (EIE), and a spatial attention mechanism.

[0022] In another embodiment, S100 further includes: S101: Multi-scale texture features are extracted using four parallel branches through adaptive average pooling, convolution, and upsampling; S102: The low-frequency background is then separated by the Edge Information Enhancer (EIE) to enhance the edge information; S103: The spatial attention mechanism is used to suppress the background and highlight the correlation of target pixels. Finally, the features are fused by channel splicing and 1×1 convolution.

[0023] In this embodiment, the method first utilizes four parallel branches to extract multi-scale texture features through adaptive average pooling, convolution, and upsampling. Then, it separates low-frequency background using EIE to enhance edge information. Next, it employs a spatial attention mechanism to suppress background and highlight target pixel correlation. Finally, it fuses features through channel concatenation and 1×1 convolution, achieving the integration of fine-grained details and multi-scale edge information, thus solving the problem of missing information for weak targets. The MSTP network structure is as follows: Figure 3 As shown, the specific process is as follows: Assuming input features These are features generated by Backbone deep processing, where C represents the dimension of the feature channel, and W and H represent the width and height of the feature, respectively. MSTP first processes the features... The data is fed into a multi-scale branch and undergoes adaptive average pooling to extract local information of different sizes, capturing multi-level features of the image. The workflow is formatted as follows: in, This indicates an average pooling operation using four convolutional kernels of different sizes. This represents the features obtained after adaptive average pooling, i.e., the multi-scale features.

[0024] Then, the obtained multi-scale features The input is fed into a 1×1 convolution and a 3×3 convolution to extract detailed information, which is then upsampled to restore the output dimension before being input into the next convolution. Figure 2 The EIE module further enhances object edge features and improves the model's accuracy in locating target boundaries. Its operation process is formatted as follows: in This represents a j×j convolutional layer with batch normalization and a sigmoid function. This is represented as upsampling. At this point, we obtain the scale features at time t+1. .

[0025] To enhance the detail information at each scale, we introduce an edge information enhancer. The features in each branch are enhanced as shown in the following formula: in, These are features enhanced with detail. The Edge Information Enhancer (EIE) structure is as follows: Figure 3As shown, the formula is as follows: in, This represents the intermediate features of EIE. EIE improves the accuracy of boundary localization by separating low-frequency background information and reducing interference from non-edge regions, thereby enhancing the edge information in the feature map.

[0026] Furthermore, to highlight the correlation between target pixels and other pixels while suppressing the relationship matrix of background pixels, we introduce a spatial attention mechanism (SAM). First, the input features... Global average pooling along the channel dimension Then, after two 3×3 convolutions, the attention weights are calculated using the Sigmoid function and reshaped into a one-dimensional vector. After another 1 1. Convolution restored to It is expressed as follows: in, This indicates a global average pooling operation. Represents the ReLU activation function. This represents the Hadamard product. These features are then fused using channel concatenation and a 1×1 convolutional layer, as shown below: in, It is the feature output of Multi-Scale Texture-Aware (MSTP). In this way, It contains both fine-grained details and multi-scale edge information, solving the problem of missing information in small targets.

[0027] In another embodiment, S200 further includes: S201: DBS utilizes two recalibration attention units (RAUs) for high-level semantic features. and low-level spatial features The mutual representations are extracted and refined through channel reshaping, point-by-point multiplication and reverse operation to make up for the missing spatial boundary information in the high layer and the missing semantic information in the low layer. S202: The Attention Selection Mechanism (ASM) is used to calculate the spatial weights of feature modulation. By using skip connections to preserve input features, the gradient vanishing effect is mitigated, thereby refining the boundaries of coarse features and enhancing their semantics. S203: The final features are obtained by projecting and fusing features through a 1×1 convolutional layer.

[0028] In this embodiment, shallow features (such as edges and contours) play a crucial role in the detection of small objects in visual tasks, but these features gradually weaken as network depth increases. Shallow features have less semantic meaning but are rich in detail, have more distinct boundaries, and less distortion; deep features contain rich semantic information. In previous studies, element-wise addition has been widely used as the most direct feature fusion method. However, such methods are limited by the inherent difference in the receptive fields of shallow and deep features. Conventional element-wise addition, concatenation, and other fusion operations are difficult to achieve semantic alignment, easily leading to feature redundancy and noise interference, and reducing fusion efficiency.

[0029] To alleviate this problem, we further propose a Dynamic Boundary Selection (DBS) method, which achieves complementarity between the boundaries and semantic information of high- and low-level features by introducing Recalibrated Attention Units (RAUs) and an Aspect-Sensitive Attention (ASM) mechanism. First, the proposed DBS utilizes two RAUs to complement the high-level semantic features. and low-level spatial features The mutual representations are extracted, and the features are refined through channel reshaping, point-by-point multiplication, and inverse operations to compensate for the missing spatial boundary information in the higher layers and the missing semantic information in the lower layers. The RAU block function process can be represented as follows: in , The input features are used as linear mappings by applying two convolutions with kernel sizes of 1×1 and 3×3. , , The sigmoid function is used. The channel dimension is reshaped to 32 to obtain the feature map. , . It is point-by-point multiplication, subtracting the feature through the reverse operation. This refines inaccurate and coarse estimates into accurate feature maps.

[0030] Then, to enhance the network's ability to focus on the most relevant spatial context regions of the detected target, an attention selection mechanism (ASM) is proposed in DBS to compute the spatial weights of feature modulation. Essentially, it uses spatial attention (SA) to generate spatial selection masks in shallow layers and channel attention (CA) in deeper layers to filter irrelevant semantic information to highlight small target regions and suppress background information. Then, skip connections are used to preserve input features to alleviate gradient vanishing, achieving boundary refinement and semantic enhancement of coarse features.

[0031] like Figure 4 As shown, ASM first concatenates the feature maps obtained from i RAUs to obtain... Then, by applying max pooling and average pooling (using respectively) and To efficiently extract spatial relationships, as shown in the formula: in and These are the spatial characteristics of average pooling and max pooling.

[0032] To facilitate information exchange between different spatial features, ASM concatenates spatial pooling features in shallow layers and uses convolutional layers. The pooling feature (with 2 channels) is converted into N spatial attention maps. For each spatial attention map... Apply activation functions to obtain spatial selection masks. As shown below: in, This represents the sigmoid function. Meanwhile, ASM uses CA (Cartesian Algorithm) at a deeper level to help filter out irrelevant semantic information, making the fused features more conducive to classification and localization. CA is expressed as the following formula: in, The ReLU activation function is used. Furthermore, skip connections are added to the input features to mitigate the vanishing gradient problem and simplify the learning process. Finally, features are projected and fused through a 1×1 convolutional layer to obtain the final features (i.e., Its expression is as follows: in It is a 1×1 convolution with batch normalization and ReLU activation layers.

[0033] In another embodiment, the attention selection mechanism employs spatial attention (SA) to generate a spatial selection mask at a shallow layer, and channel attention (CA) at a deep layer to filter irrelevant semantic information to highlight small target regions and suppress background information.

[0034] In another embodiment, the TAD-Head consists of a shared convolutional module (SC), a task decomposition module (TDM), and a feature interaction branch (FIB).

[0035] In this embodiment, current object detection models, which widely employ parallel detection heads, are limited by convolutional operations and parallel structure design, making it difficult to effectively address the two core challenges of detection tasks: object diversity and the interactivity between classification and localization tasks. Existing methods typically learn task-related features (such as...) in two independent branches by stacking convolutional layers. Figure 5 As shown in (a), this design has significant limitations, especially the lack of cross-task perception capabilities required by an ideal detector, which makes it difficult to coordinate and optimize classification and localization tasks.

[0036] To address the issues of feature conflicts and insufficient interaction in traditional detection heads during multi-task processing, this method proposes a novel TAD-Head detection head. This head aims to achieve efficient interaction between classification and localization tasks and optimizes prediction consistency by introducing additional supervisory information. The TAD-Head consists of three core components: shared convolution (SC), task decomposition module (TDM), and feature interaction branch (FIB), as follows... Figure 5 As shown in (b), the three work together to significantly improve the performance of the detection model.

[0037] In another embodiment, S300 further includes: S301: The Task Decomposition Module (TDM) uses a layer attention mechanism for dynamic feature allocation, enabling each task to selectively focus on layer-specific features. S302: In the localization branch, task interaction features The input is fed into the spatial attention mechanism to dynamically calculate task-specific features, thereby achieving task decomposition; in the classification branch, dynamic feature selection is performed through interactive features to complete task interaction and feature fusion.

[0038] In this example, the features output by Neck at different k stages The inputs are fed into convolutional layers with a kernel size of 1×1 for channel dimension adjustment, and denoted as the output of the k-th stage convolutional layer, which is then used as the input features of TAD-Head. .Will First, we enter the SC module. This module significantly reduces the number of parameters in the model through a parameter sharing mechanism, making the model more suitable for resource-constrained devices. The specific calculation process of the SC module can be represented as follows: in, It is a 3×3 convolutional layer with set normalization and sigmoid function, and [ , ] represents the concatenation operation.

[0039] To resolve the feature conflict between classification and localization tasks in single-branch designs, D-Head introduces TDM (Tracked Memory Management) to learn task interaction features through multi-layer convolution. TDM employs a layer attention mechanism for dynamic feature allocation, enabling each task to selectively focus on layer-specific features, thereby avoiding conflicts in shared representations. Furthermore, to better utilize spatial information and achieve effective inter-task interaction, we designed a FIB module: in the localization branch, task interaction features... The input is fed into the spatial attention mechanism to dynamically calculate task-specific features, thereby achieving task decomposition. In the classification branch, dynamic feature selection is performed through interactive features to complete task interaction and feature fusion. Specifically, the correlation between classification and localization is calculated using the following formula: in, There are two task branches outputting. , yes After average pooling and max pooling respectively, the final result will be... The classification score and location offset of the target are predicted separately.

[0040] In another embodiment, this experiment uses YOLOv8 as the baseline model to train and validate the proposed small target detection method (DAEC) in remote sensing images. The specific process and parameter settings are as follows: The hardware platform uses an AMD Ryzen Threadripper central processing unit and two NVIDIA RTX3080 graphics cards to ensure the efficiency of model training and inference; In terms of datasets, the Visdrone dataset was selected as the main training dataset (6471 images were used for training and 548 images were used for validation), and the ROSD and VEDAI datasets were used to verify the robustness of the method. During the training phase, the batch size of the Visdrone dataset was set to 8, the number of training epochs was set to 200, and data augmentation strategies were enabled (initial mosaic augmentation and initial hybrid augmentation parameters were both set to 0.5). The model optimization adopted the stochastic gradient descent (SGD) optimizer, with an initial learning rate of 0.0001, weight decay of 0.0005, and initial momentum of 0.8. The learning rate decay method was cosine annealing, and the model parameters were optimized in stages using cross-entropy loss (for classification tasks) and CIoU loss (for regression tasks). During the validation phase, the validation confidence and intersection-over-union (IoU) of the Visdrone validation set were set to 0.001 and 0.5, respectively. When validating on the ROSD and VEDAI datasets, only the batch size was adjusted to 4 and the input images were uniformly scaled to 640×640 pixels, while the other hyperparameters remained consistent with those in the Visdrone training phase.

[0041] The impact of MSTP, DBS, and TAD-Head on model performance: To clarify the importance of each component in the proposed DAEC, we progressively validated the effectiveness of MSTP, DBS, and TAD-Head on the VisDrone2019 dataset through ablation experiments in baseline (YOLOv8). The experimental results are shown in Table 1 (in the table, "√" indicates that the corresponding module is enabled, and "-" indicates that the corresponding module is not enabled).

[0042] Table 1 As shown in Table 1, the introduction of the MSTP module significantly improved all evaluation metrics of the model, particularly in accuracy, compared to the baseline model. and Gains of 6.0% and 5.1% were achieved, respectively. This improvement stems from MSTP. This improvement effectively expands the semantic cues within small targets by extracting the internal texture information of the object, thereby improving target saliency and classification robustness, and solving the problem that the texture information of small targets in remote sensing images is weak and easily interfered with by the background.

[0043] When integrating the DBS module separately into the baseline model, The performance index improved by 5.4%, demonstrating that DBS achieves mutual representation of high-level semantic features and low-level spatial features through two recalibrated attention units (RAUs), compensating for the lack of spatial boundary information in high-level features and the insufficient semantic information in low-level features. Simultaneously, by combining the attention selection mechanism (ASM) to calculate feature modulation spatial weights, it refines the boundaries and enhances the semantics of coarse features, accurately calibrating the spatial location of small targets and providing the model with more precise localization information, thus adapting to the detection needs of remote sensing images with large target scale variations and blurred boundaries. Furthermore, after collaborating with MSTP, DAEC achieved 50.0% (↑7.8%) on VisDrone. This indicates that MSTP provides DBS with richer semantic clues about the target, assisting it in accurately identifying target boundary regions; while DBS assigns more precise spatial location attributes to the texture features extracted by MSTP, reducing the interference of background texture on target judgment. Together, they optimize the representation quality of target features, further verifying the necessity of collaboration between the two modules in the feature extraction and processing stages.

[0044] Introducing TAD-Head alone into the baseline model achieved 49.6% (↑7.4%) of the desired performance on the dataset. TAD-Head reduces parameter redundancy through shared convolution (SC) and achieves dynamic feature allocation for classification and localization tasks using the layer attention mechanism of the Task Decomposition Module (TDM), avoiding feature conflicts between tasks. Simultaneously, it integrates supervision information from both tasks through a Feature Interaction Branch (FIB), achieving consistent optimization of classification confidence and bounding box localization. This effectively solves the problems of insufficient interaction between classification and localization tasks and prediction bias in traditional parallel detection heads, improving the model's collaborative accuracy in classifying and localizing small targets. Finally, the complete DAEC model integrating MSTP, DBS, and TAD-Head outperforms the YOLOv8 baseline model in... and The results improved by 6.1% and 7.6%. This fully demonstrates that the three core components proposed in this disclosure support each other in function and synergistically enhance performance, together forming an efficient architecture adapted to the needs of small target detection in remote sensing images. It effectively solves key technical problems such as small target scale, weak texture, strong background interference, and inconsistent classification and localization in high-altitude remote sensing images, and significantly improves the detection performance and robustness of the model.

[0045] To further verify the robustness of DECA, we conducted comparative experiments on two representative remote sensing datasets, RSOD and VEDAI. These datasets represent typical challenges in remote sensing target detection, including different scale distributions and small target detection, respectively, and can more comprehensively verify the model's robustness and small target detection capabilities in complex remote sensing scenarios. Compared to baseline algorithms, our method significantly improves the overall detection accuracy. Table 2 compares the baseline algorithms and the proposed method. Experimental results show that the final model incorporating the three modules achieves good results, with mAP improvements of 1.8% and 3.8% on the RSOD and VEDAI datasets, respectively. The accuracy rates were 0.4% and 4.8% respectively. In particular, the results obtained from the VEDAI dataset show an overall improvement in detection accuracy, which proves the effectiveness of our model in identifying small targets on remote sensing datasets.

[0046] Table 2 In another embodiment, a dynamic alignment and edge feature calibration device for small target detection in remote sensing images includes: An apparatus for extracting multi-scale texture information using a multi-scale texture-aware MSTP during the feature extraction stage, capturing the relationships between pixels, and simultaneously magnifying the homogeneous details of an object to characterize important textures. An apparatus for retaining useful multi-scale features and further filtering redundant features by selecting a BS-FPN composed of boundary enhancement DBS and C2f during the feature processing stage. This device is used to input the processed features into the Task Aware Detection Head (TAD-Head) for processing, and to fully integrate the supervision information from both classification and localization tasks through a cross-task awareness mechanism to achieve consistency detection between classification and localization.

[0047] In this embodiment, a unique detector called DAEC is proposed based on the YOLOv8 detection model. It addresses the problem of weak targets in remote sensing images by exploring high-frequency detail features of objects and enhancing cross-task perceptual interaction capabilities. The DAEC detector extracts deep information through MSTP, amplifying the differences between the object and the background while exploring the internal texture details of the object. It also aggregates boundary and semantic information through DBS, thereby depicting finer-grained target contours and recalibrating object positions. Furthermore, to alleviate task defocus caused by the entanglement between classification and regression, TAD-Head is employed for cross-task interaction, enabling specialized predictions for classification and localization and improving performance.

[0048] Furthermore, the present invention discloses a computer storage medium comprising computer instructions that, when executed on a computer, cause the computer to perform any of the methods described above.

[0049] Furthermore, the present invention also discloses an electronic device, wherein the electronic device comprises: Memory, processor, and computer programs stored in memory and executable on the processor, wherein, When the processor executes the program, it implements any of the methods described above.

[0050] Furthermore, the technical effects produced by the key technical means of the present invention are summarized as follows: 1. Through the bidirectional fusion mechanism of the DBS module, the interaction between features of different resolutions is closer, which not only enhances the comprehensiveness and accuracy of feature information, but also helps to more accurately describe the object outline and recalibrate the object position.

[0051] 2. The MSTP module, through its multi-scale parallel branches and Edge Information Enhancer (EIE), effectively extracts and enhances fine-grained detail information, especially the edge sharpness of objects, which is crucial for improving the accuracy of small object detection. Its multi-scale parallel structure, combined with a spatial attention mechanism, effectively filters background noise and enhances the feature representation of target regions, allowing the model to focus more on the object of interest, thereby reducing the false detection rate caused by background interference.

[0052] 3. TAD-Head effectively solves the feature conflict problem between classification and localization tasks through the dynamic feature allocation mechanism of the Task Decomposition Module (TDM). With the help of the additional supervision information of spatial offset map and spatial probability map in the feature interaction branch, it enhances the prediction consistency and reliability of classification and localization tasks, and ultimately improves the performance of the overall detection model.

[0053] 4. Although MSTP adds a step for extracting edge information, it does not significantly increase the computational burden and maintains high inference efficiency due to the use of efficient algorithm design (such as 3×3 average pooling and 1×1 convolution).

[0054] Furthermore, we can further analyze and infer some technical effects: 1. By introducing the bidirectional fusion mechanism in the DBS module and optimizing the design of TAD-Head, information at different scales can be better utilized during training, reducing training redundancy caused by loss or redundancy of multi-scale information, thereby accelerating model convergence, reducing training costs, and potentially improving the generalization ability of the final model.

[0055] 2. Using a multi-scale processing architecture (such as MSTP) combined with a spatial attention mechanism helps the model maintain stable detection accuracy when facing complex scenes such as changes in lighting and occlusion. This improves the model's adaptability and reliability under various environmental conditions and reduces performance fluctuations caused by changes in external factors.

[0056] 3. Although this technical solution is primarily designed for the detection of small targets in remote sensing images, its core concept—improving target recognition accuracy through efficient information fusion and feature extraction—is also applicable to other visual tasks, such as the detection of minute lesions in medical images and the recognition of long-range targets in autonomous driving. It provides new ideas and technical support for research in related fields, promoting the cross-disciplinary application and development of the technology.

[0057] 4. Employing Boundary Selection Pyramid Network (BS-FPN) makes the entire system more modular and easier to integrate into existing computer vision frameworks. This reduces the difficulty of integrating new technologies into existing systems, accelerates product iteration cycles, and improves development efficiency.

[0058] 5. With more accurate small target detection results, users can obtain more precise geographic information data or other key information in detection scenarios, thereby making better decisions. This enhances user trust and satisfaction with the system, which is especially important in critical application scenarios that require high-precision data support.

[0059] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A method for dynamic alignment and edge feature calibration of small targets in remote sensing images, comprising the following steps: S100: In the feature extraction stage, multi-scale texture perception MSTP is used to extract multi-scale texture information, capture the relationship between pixels, and at the same time magnify the homogeneous details of the object to represent important textures. S200: In the feature processing stage, useful multi-scale features are preserved and redundant features are further filtered out by using the boundary selection pyramid network BS-FPN composed of boundary enhancement DBS and C2f. S300: The processed features are input into the Task Aware Detection Head (TAD-Head) for processing. Through a cross-task awareness mechanism, the supervision information from both classification and localization tasks is fully integrated to achieve consistency detection between classification and localization.

2. The method according to claim 1, preferably, the MSTP consists of a multi-scale feature extraction branch, an edge information enhancer (EIE), and a spatial attention mechanism (SAM).

3. The method according to claim 2, wherein step S100 further comprises: S101: Multi-scale texture features are extracted using four parallel branches through adaptive average pooling, convolution, and upsampling; S102: The low-frequency background is then separated by the Edge Information Enhancer (EIE) to enhance the edge information; S103: The spatial attention mechanism is used to suppress the background and highlight the correlation of target pixels. Finally, the features are fused by channel splicing and 1×1 convolution.

4. The method according to claim 1, wherein step S200 further comprises: S201: DBS utilizes two recalibrated attention units (RAUs) for high-level semantic features. and low-level spatial features The mutual representations are extracted and refined through channel reshaping, point-by-point multiplication and reverse operation to make up for the missing spatial boundary information in the high layer and the missing semantic information in the low layer. S202: The attention selection mechanism ASM is used to calculate the spatial weights of feature modulation. Skip connections are used to preserve input features to alleviate gradient vanishing, thereby refining the boundaries of coarse features and enhancing semantics. S203: The final features are obtained by projecting and fusing features through a 1×1 convolutional layer.

5. The method according to claim 4, wherein the attention selection mechanism uses spatial attention (SA) to generate a spatial selection mask at a shallow layer, and channel attention (CA) to filter irrelevant semantic information at a deep layer to highlight small target regions and suppress background information.

6. The method according to claim 1, wherein the TAD-Head is composed of a shared convolutional module (SC), a task decomposition module (TDM), and a feature interaction branch (FIB).

7. The method according to claim 6, wherein step S300 further comprises: S301: The Task Decomposition Module (TDM) uses a layer attention mechanism for dynamic feature allocation, enabling each task to selectively focus on layer-specific features. S302: In the localization branch, task interaction features The input is fed into the spatial attention mechanism to dynamically calculate task-specific features, thereby achieving task decomposition; in the classification branch, dynamic feature selection is performed through interactive features to complete task interaction and feature fusion.

8. A dynamic alignment and edge feature calibration device for small target detection in remote sensing images, comprising: An apparatus for extracting multi-scale texture information using a multi-scale texture-aware MSTP during the feature extraction stage, capturing the relationships between pixels, and simultaneously magnifying the homogeneous details of an object to characterize important textures. An apparatus for retaining useful multi-scale features and further filtering redundant features by selecting a BS-FPN composed of boundary enhancement DBS and C2f during the feature processing stage. This device is used to input the processed features into the Task Aware Detection Head (TAD-Head) for processing, and to fully integrate the supervision information from both classification and localization tasks through a cross-task awareness mechanism to achieve consistency detection between classification and localization.

9. A computer storage medium, wherein, The storage medium includes computer instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 7.

10. An electronic device, wherein, The electronic device includes: Memory, processor, and computer programs stored in memory and executable on the processor, wherein, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

Images