Remote sensing small target directed detection model, detection method and construction method

Abstract

The invention relates to a remote sensing small target directed detection model, a detection method and a construction method, and belongs to the technical field of remote sensing image processing. On the basis of feature enhancement, by means of asymmetric image filling operation of kite convolution, an attention feature fusion mechanism is iteratively used to construct a feature fusion refinement module so as to strengthen small target edge feature extraction and detail perception ability, a cascade group attention mechanism is used to fully integrate small target edge information features, and feature diversity is enhanced; designing a multi-scale shared convolution detection head, and realizing high-low layer feature information fusion enhancement and effective integration of semantic features through a multi-level feature fusion architecture and detail enhancement convolution; according to the method, I < n >-I < oU > and S < oU > are fused, an angle-distance-shape joint loss function is designed, the problem of geometric feature blurring caused by pixel missing of a small target and the inconsistency of a metric value and loss caused by a rotation parameter are effectively solved, and the positioning precision and robustness of target directed detection are enhanced.

Description

Technical Field

[0001] The present invention relates to a remote sensing small target directional detection model, a detection method and a construction method, and belongs to the technical field of remote sensing image processing. Background Art

[0002] Remote sensing small object detection is of great significance in fields such as land and resources monitoring, urban planning, and ecological and environmental protection. Remote sensing small object detection technology has experienced rapid development over the past few decades, gradually transitioning from early methods based on handcrafted features to end-to-end solutions driven by deep learning. During this period, different algorithms have continuously evolved to address various challenges.

[0003] In the early days, the HOG+SVM framework, based on handcrafted features, was the primary method for detecting small remote sensing targets. This approach relied on manually designed feature descriptors (such as HOG features) and traditional machine learning classifiers (such as SVM). While this method was able to achieve target detection to a certain extent, its detection accuracy and generalization capabilities were limited because handcrafted features could not fully and accurately represent the complex remote sensing image information.

[0004] Subsequently, the rise of deep learning brought new breakthroughs in small target detection in remote sensing. The two-stage algorithm FasterR-CNN emerged, significantly improving target localization accuracy by introducing a region proposal network (RPN). However, FasterR-CNN suffers from significant limitations when handling small objects, and its performance is limited by the lack of semantic information in shallow features. In convolutional neural networks, shallow features contain more spatial information but relatively less semantic information. Small objects occupy fewer pixels in an image, and their features rely more heavily on shallow features. Therefore, it is difficult to extract sufficient semantic information from these shallow features for accurate target detection. Therefore, the Feature Pyramid Network (FPN), which fuses multi-scale features, was proposed. FPN fuses feature maps of different scales through bottom-up and top-down approaches, enabling the network to acquire rich feature information at different scales. This method achieved multi-scale feature fusion for the first time, significantly improving small target detection accuracy. Subsequent research has continued to improve feature fusion strategies, such as SNIP and FSAF. SNIP alleviates the training conflict between objects of different scales by training and testing the model on images of different scales. FSAF adaptively selects appropriate feature layers for detection of objects of different scales, further optimizing the effect of feature fusion. Other studies have introduced attention mechanisms, such as CBAM, which adaptively adjust the channel and spatial attention of feature maps, enhancing the network's focus on important features and thus improving the accuracy of small object detection. However, it should be noted that most of these methods are designed for horizontal box detection and are difficult to adapt to the common need for oriented object detection in remote sensing scenarios.

[0005] In terms of oriented detection technology, the RRPN, proposed in 2016, was groundbreaking. By generating rotating anchor frames, it enabled the first detection of oriented objects in remote sensing scenarios. However, the discrete preset angles of the anchor frames used by the RRPN resulted in insufficient rotation sensitivity. These discrete angle presets cannot accurately adapt to the actual angles of various rotating objects.

[0006] Therefore, to improve the accuracy of oriented detection, RotBox introduces an angle prediction branch. This branch allows the model to directly predict the object's rotation angle. However, RotBox does not effectively address the problem of feature misalignment caused by small objects occupying fewer pixels, nor does it address the insensitivity of traditional IoU loss to rotation angles.

[0007] After 2020, anchor-free methods have gradually gained popularity. Oriented RepPoints achieves continuous angle prediction through point set regression, avoiding the problem of discretization of anchor box preset angles and enabling the model to more accurately capture the target's rotation angle. However, the problem of feature misalignment caused by small targets occupying fewer pixels has not been effectively solved, affecting the model's training results. The RoI Transformer improves the positioning accuracy of rotating targets in dense scenes by cascading rotational region of interest transformations. However, it also does not effectively solve the existing feature misalignment problem.

[0008] The YOLO family of detectors continues to advance in the field of single-stage object detection, providing new insights for oriented remote sensing detection. YOLOv5-obb improves the rotation anchor box matching strategy, improving detection efficiency and, to a certain extent, accuracy. However, it is still limited by the quantization error of discrete angle encoding and cannot accurately represent the object's rotation angle. YOLOv7-obb introduces an attention mechanism to enhance the features of small objects. However, the traditional GIoU loss remains insensitive to rotation angles, resulting in poor geometric constraint performance for oriented object detection.

[0009] In addition to the problem of feature misalignment caused by fewer small target pixels mentioned above, the traditional IoU loss is insensitive to rotation angles. There are also problems such as the complex background information of remote sensing images, which makes it difficult to extract edge features, resulting in discontinuous extracted features, and the inconsistency between measurement values ​​and losses caused by the introduction of angle parameters in directional detection.

[0010] For example, Chinese patent publication number CN112329721A discloses a method for remote sensing small target detection with a lightweight model design. The method includes: using the single-stage YOLOv4 algorithm as the main framework; clustering anchor boxes using the K-means++ algorithm, and then introducing the SENet channel attention mechanism to construct a detection model with better remote sensing small target detection performance; lightweighting the model; and reducing the number of model parameters and memory consumption through the idea of ​​model lightweighting, so that the lightweight model can maintain the accuracy of complex neural network models while consuming the same memory and computing power as lightweight algorithms, and has better detection results than other mainstream lightweight algorithms. However, remote sensing small target detection still suffers from complex background information, difficulty in extracting edge features, large size variations between different targets, and a small number of pixels on small targets, resulting in discontinuous and misaligned extracted features, as well as inconsistent metric values ​​and losses due to the introduction of angle parameters in directional detection. Summary of the Invention

[0011] The purpose of the present invention is to provide a remote sensing small target directional detection model, detection method and construction method to solve the problems of complex background information, difficult edge feature extraction, large size variation of different targets, and fewer pixels of small targets, which lead to discontinuous and misaligned extracted features in remote sensing small target detection.

[0012] The remote sensing small target directional detection model described in the present invention includes:

[0013] Backbone network: including kite convolution, C3k2_iAFF module and C2PSA_CGA module, used to perform preliminary feature extraction on the input image, providing a basis for subsequent feature fusion and detection;

[0014] Neck network: includes kite convolution and C3k2_iAFF modules, located between the backbone network and the head network, used to fuse the features of different levels extracted by the backbone network to generate feature representation;

[0015] Head network: includes a multi-scale shared convolutional detection head, which is used for multi-scale shared feature fusion;

[0016] Kite convolution, C3k2_iAFF module, and C2PSA_CGA module enhance the feature expression and recognizability of small objects by performing asymmetric image padding feature extraction operations, iteratively using attention mechanisms, and cascading edge information.

[0017] The multi-scale shared convolutional detection head uses multi-scale feature fusion and shared enhancement strategy to construct a detection head with a multi-scale shared detail enhancement structure to enhance the ability to perceive details of small targets.

[0018] By using asymmetric padding to create horizontal and vertical convolution kernels for different areas of the image, we can better adapt to the Gaussian spatial distribution of pixels of small targets, enhance feature extraction, and increase the receptive field; we iteratively apply the multi-scale channel attention mechanism to improve the initial integration of feature maps, effectively enhancing the feature expression and recognizability of small targets; and we use the cascaded group attention mechanism to improve the diversity of attention maps, and gradually integrate edge information during the cascade process, while reducing computational redundancy.

[0019] A detection head with a multi-scale shared detail enhancement structure is constructed. Through multi-dimensional feature fusion, detail enhancement, dynamic weight adjustment, and re-parameterization, the expressive power of features is gradually enhanced, and the multi-scale feature information is fully integrated to enhance the detail perception ability of small targets without increasing the computational cost.

[0020] Preferably, the first layer of the kite convolution adopts four parallel convolution branches to achieve multi-directional feature capture by differentiating padding parameters and kernel sizes; each branch has different padding parameters and convolution kernels to achieve feature extraction in different directions, and its calculation formula is as follows:

[0021]

[0022] in, is the convolution operator, Wi is the corresponding convolution kernel, and the four parameters in P represent the number of padded pixels in the left, right, up, and down directions respectively;

[0023] After the first convolution layer, the height h′, width w′, and number of channels c′ of the output feature map are related to the height h1, width w1, and number of channels c1 of the input feature map. The calculation formula is as follows:

[0024]

[0025] Among them, c2 is the number of channels of the final output feature map of the kite convolution, and s is the convolution step size;

[0026] The concatenated tensor is normalized by a convolution kernel of W(2,2,c2). The height and width of the output feature map are adjusted to the preset values ​​h2 and w2, and the final output is Y(h2,w2,c2). The calculation formula is as follows:

[0027]

[0028] The effective receptive field of the pinwheel-shaped convolution (Pconv) exhibits Gaussian decay characteristics. When the kernel size k = 3, the actual receptive field expands to 25 pixels, and the convolution weight of the central region is significantly higher than that of the periphery. This allows for greater attention to central features, which aligns with the characteristic of small object feature concentration. This allows for better focus on key features in tasks such as small object detection, improving feature extraction accuracy while achieving lightweightness. This is particularly suitable for detail-sensitive tasks such as small object detection.

[0029] Preferably, the C3k2_iAFF module realizes dynamic feature selection and adaptive fusion through a multi-level attention mechanism, dynamically calibrates the semantic association and spatial distribution differences between features, and refines the feature fusion process to enhance the feature expression of small targets; the C3k2_iAFF module first pre-processes the input feature convolution, and then distributes the features to multiple branches through the Split operation; each branch is independently connected to the Bottleneck_iAFF unit to realize parallel iterative optimization of multi-scale features: the deep branch uses the Bottleneck_iAFF unit to enhance the global semantic consistency, and the shallow branch uses the Bottleneck_iAFF unit to retain the detail features; finally, the multi-branch features processed by the Bottleneck_iAFF unit are aggregated through Concat and then unified through convolution.

[0030] Preferably, the Bottleneck_iAFF unit dynamically generates weights for the features of the preliminary interaction by iteratively using the multi-scale channel attention sub-units, and uses the output weights to weight the inputs X and Y element by element, thereby strengthening key feature channels and suppressing redundant information. Then, the features after the initial attention calibration are processed again. Through iterative optimization, the feature fusion weights are gradually refined to avoid information bias in the initial fusion. Each round of iteration dynamically adjusts the fusion ratio of X and Y based on the previous results, and finally outputs the fused features.

[0031] Preferably, the C2PSA_CGA module constructs a new feature fusion processing framework by fusing the C2PSA module and the CGA unit. First, the input features are processed in parallel by multi-scale convolution of the C2PSA module to capture the spatial information of different receptive fields; then the CGA unit is introduced to divide the features into each attention head, and the feature representation is gradually refined through independent self-attention calculation and cross-head cascade operation; finally, cross-layer feature fusion is achieved through convolution.

[0032] Compared with C2PSA, C2PSA_CGA uses CGA's input segmentation strategy to avoid duplicate extraction problems at the feature processing level, uses CGA to reduce QKV channel redundancy at the computational efficiency level, and expands learning capacity through CGA's cross-head cascade at the model depth level. It can significantly improve feature diversity, balance computational efficiency and model expressiveness, and innovate the cascade attention mechanism to strengthen the focus on key information.

[0033] Preferably, the CGA unit divides the input features into different parts, each of which is input to an attention head; each head calculates its own attention map, and then concatenates the outputs of all heads, in this way, the output of each head is added to the input of the next head, thereby gradually refining the feature representation; finally, they are projected back to the input dimension through a linear layer.

[0034] Preferably, the multi-scale shared convolution detection head is equipped with a small target detection layer and a shared enhancement strategy unit to construct a multi-layered feature perception system.

[0035] A multi-scale shared convolutional detection head is designed to achieve effective integration of high- and low-level feature information fusion and enhancement with semantic features through a multi-level feature fusion architecture and detail enhancement convolution.

[0036] Preferably, the bottom detection layer of the small target detection layer focuses on capturing high-resolution details of small targets and retains key information such as edges and textures; the middle detection layer gradually integrates semantic and detail features to balance the expression requirements of different scales; and the high-level detection layer focuses on the strong semantic features of medium and large targets.

[0037] Preferably, the specific workflow of the shared enhancement strategy unit is as follows: the multi-scale input first undergoes detail enhancement convolution, then the response of the feature channel is enhanced by an efficient channel attention mechanism, and finally the detail enhancement convolution is processed again to form an extraction-optimization-re-extraction link. Multi-dimensional feature fusion and re-parameterization operations gradually enhance the expressive power of features, obtain targeted processing of directional detection, bounding box regression, and classification regression output, and improve the detection accuracy of multi-scale targets in complex scenarios. The method for constructing a remote sensing small target directional detection model described in the present invention includes:

[0038] Step 1: Construct the aforementioned remote sensing small target directed detection model;

[0039] Step 2: Construct the Inner_SIoU loss function and optimize the remote sensing small target directed detection model constructed in step 1;

[0040] The Inner_SIoU loss function is as follows:

[0041] Inner_SIoU=(1-α·IoU)+λ·(Angle_Cost+Distance_Cost+Shape_Cost)

[0042] Loss = 1-Inner_SIoU

[0043] Among them, α is the dynamic adjustment factor, λ is the dynamic weight coefficient, Angle_Cost is the angle cost, Distance_Cost is the distance cost, Shape_Cost is the shape cost, and Loss is the loss value.

[0044] The SIoU loss function, which is sensitive to the angle of the rotating frame and contains multi-dimensional constraints, is fused with the Inner-IoU loss function with a dynamic scaling mechanism to obtain the Inner-SIoU loss function, which includes an angle-aware weight adjustment module and dynamically balances angle errors. This enhances the ability to cope with angle constraints and scale changes in oriented target detection.

[0045] The method for detecting a directed remote sensing small target of the present invention adopts the directed remote sensing small target detection model constructed by the method for constructing a directed remote sensing small target detection model to perform detection.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The invention discloses a remote sensing small target directed detection model, detection method and construction method. Based on feature enhancement, the invention uses the asymmetric image filling operation of kite convolution and iteratively uses the attention feature fusion mechanism to construct a feature fusion refinement module to enhance the edge feature extraction and detail perception capabilities of small targets. The invention uses the cascade group attention mechanism to fully integrate the edge information features of small targets and enhance feature diversity. The invention designs a multi-scale shared convolution detection head, realizes the effective integration of high- and low-level feature information fusion and semantic features through a multi-level feature fusion architecture and detail enhancement convolution. The invention also fuses the Inner-IoU and SIoU, and utilizes the Inner-IoU to fully integrate the edge information features of small targets and enhance feature diversity. The dynamic scaling mechanism of Inner-IoU suppresses the interference of background noise on small targets under rotation posture. At the same time, the phased sample weighting strategy is combined to accelerate the convergence process of rotated targets. The rotation-sensitive angle penalty term of SIoU, the center point polar coordinate distance constraint and aspect ratio dynamic compensation are used to construct a multi-dimensional geometric constraint system. Combined with the scale-adaptive weight distribution strategy of Inner-IoU, the Inner-SIoU loss function is obtained, and the angle-distance-shape joint loss function is designed to effectively solve the geometric feature ambiguity problem of small targets caused by pixel missing and the inconsistency between measurement value and loss caused by rotation parameters, thereby enhancing the positioning accuracy and robustness of target directional detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1The results of horizontal detection and directional detection are shown in Figure 1. (a) is the original image, (b) is the horizontal detection, and (c) is the directional detection.

[0049] Figure 2 This is the structural diagram of the YOLO11-obb model;

[0050] Figure 3 This is a structural diagram of a remote sensing small target directional detection model according to the present invention;

[0051] Figure 4 This is a structural diagram of a kite convolution according to the present invention;

[0052] Figure 5 Figure 1 is a structural diagram of a C3k2-iAFF module described in the present invention; (a) is a structural diagram of a multi-scale channel attention subunit, (b) is a structural diagram of a Bottleneck_iAFF unit, and (c) is a structural diagram of a C3k2-iAFF module;

[0053] Figure 6 : This is a structural diagram of a C2PSA_CGA module according to the present invention; (a) is a structural diagram of a CGA unit, and (b) is a structural diagram of a C2PSA_CGA module;

[0054] Figure 7 This is a structural diagram of a multi-scale shared convolutional detection head according to the present invention;

[0055] Figure 8 This is a structural diagram of a detail enhancement convolution according to the present invention;

[0056] Figure 9 A histogram showing the ablation experiment results of the present invention;

[0057] Figure 10 This is a thermal diagram of the DIOR-R comparative experiment described in the present invention;

[0058] Figure 11 This is a histogram of the comparative experiment of the UCAS-AOD dataset described in the present invention;

[0059] Figure 12 The visual comparison results before and after the improvement of the remote sensing image with dense small targets described in the present invention are shown in Figure 1. (a) and (c) are the YOLO11-obb detection results, and (b) and (d) are the FEMS-YOLO detection results for the same images.

[0060] Figure 13This is a visualization comparison of the results before and after improvement when the scale of small targets in different categories of remote sensing images varies greatly according to the present invention; (a) is the YOLO11-obb detection result, and (b) is the FEMS-YOLO detection result for the same image;

[0061] Figure 14 This is the mAP50 curve comparison chart analysis diagram of the present invention;

[0062] Figure 15 This is a PR curve comparison analysis diagram described in the present invention; in the figure, (a) is the separate curves of YOLO11-obb and FEMS-YOLO, and (b) is the comparison curve of YOLO11-obb and FEMS-YOLO. DETAILED DESCRIPTION

[0063] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0064] Example 1

[0065] In the field of remote sensing target detection, oriented object detection aims to accurately describe the predicted target bounding box position, size and rotation angle by rotating the rectangular box. Compared with the traditional horizontal box detection, its core difference is the introduction of angle parameters. Figure 1 The axis-aligned rectangular frame (shown in (b)) uses high computational efficiency, but has significant limitations in remote sensing scenes. The horizontal frame easily contains a large number of background pixels, especially for targets with large aspect ratios or densely arranged targets (such as vehicles and ships), resulting in impure feature extraction and poor geometric adaptability. The horizontal frame is difficult to fit closely to the target edge, affecting positioning accuracy. Figure 1 As shown in c), the compact representation of the rotating frame can effectively reduce background interference, improve the detection robustness in dense scenes, and better conform to the spatial distribution characteristics of the target in the remote sensing image. The horizontal detection and directional detection results are different, so this embodiment chooses the directional detection of the research target, such as Figure 1 shown.

[0066] Among algorithms used for directed target detection, the YOLO series, a typical representative of single-stage target detection, excels in remote sensing detection tasks compared to traditional two-stage algorithms (such as Faster R-CNN) and anchor-free methods (such as RepPoints). Its end-to-end architecture eliminates the redundant computation of region proposals, resulting in significantly faster inference than two-stage algorithms. By integrating a hierarchical feature pyramid with a cross-scale feature fusion mechanism, it balances shallow high-resolution features (facilitating small target localization) with deep semantic features (facilitating classification robustness), overcoming the problem of missed detections in traditional algorithms due to the sparse pixel count of small targets. Furthermore, it employs lightweight designs such as depthwise separable convolution and dynamic convolution, making it more suitable for real-time applications.

[0067] Therefore, this embodiment selects the latest YOLO11-obb as the baseline algorithm, whose structure is as follows Figure 2 As shown in the figure, in response to the above-mentioned problems of feature discontinuity, feature misalignment, and inconsistency between measurement value and loss, improvement measures are proposed. The improved model structure diagram FEMS-YOLO (Feature Enhancement and Multi-Scale Sharing, FEMS) is shown in the figure. Figure 3 shown.

[0068] This embodiment discloses a remote sensing small target directional detection model (FEMS-YOLO model), including:

[0069] Backbone network: including kite convolution, C3k2_iAFF module and C2PSA_CGA module, used to perform preliminary feature extraction on the input image, providing a basis for subsequent feature fusion and detection;

[0070] Neck network: includes kite convolution and C3k2_iAFF modules, located between the backbone network and the head network, used to fuse the features of different levels extracted by the backbone network to generate feature representation;

[0071] Head network: includes a multi-scale shared convolutional detection head, which is used for multi-scale shared feature fusion;

[0072] Kite convolution, C3k2_iAFF module, and C2PSA_CGA module enhance the feature expression and recognizability of small objects by performing asymmetric image padding feature extraction operations, iteratively using attention mechanisms, and cascading edge information.

[0073] The multi-scale shared convolutional detection head uses multi-scale feature fusion and shared enhancement strategy to construct a detection head with a multi-scale shared detail enhancement structure to enhance the ability to perceive details of small targets.

[0074] 1. Enhanced representation of small target features

[0075] 1.1 Pinwheel-shaped Convolution (PConv)

[0076] Traditional convolutional networks typically use symmetrical methods to process directional image features, which inherently impairs edge feature extraction for small remote sensing targets. Their isotropic receptive field makes it difficult to capture structured information such as texture orientation and edge distribution. Furthermore, as network depth increases, the number of parameters increases, leading to a decrease in computational efficiency.

[0077] To address this problem, this embodiment introduces PConv, which provides a solution that combines accuracy and efficiency for detection through asymmetric filling strategy and Gaussian distribution driven kernel optimization design. Its multi-directional feature capture mechanism effectively alleviates the limitations of traditional symmetric convolution in edge texture extraction, and achieves accurate analysis of complex directional features in remote sensing images through differential filling and kernel optimization. The central feature focusing strategy driven by Gaussian distribution strengthens the feature expression of the core area of ​​small targets through dynamic weight allocation, significantly improving the feature misalignment problem caused by few target pixels and background interference. The PConv structure diagram and principle are shown below. Figure 4 shown.

[0078] The first layer of PConv uses four parallel convolution branches to achieve multi-directional feature capture by differentiating padding parameters and kernel sizes. Each branch has different padding parameters and convolution kernels to achieve feature extraction in different directions. The calculation formula is as follows:

[0079]

[0080] in, is the convolution operator, Wi is the corresponding convolution kernel, and the four parameters in P represent the number of padded pixels in the left, right, up, and down directions respectively;

[0081] After the first convolution layer, the height h′, width w′, and number of channels c′ of the output feature map are related to the height h1, width w1, and number of channels c1 of the input feature map. The calculation formula is as follows:

[0082]

[0083] Among them, c2 is the number of channels of the final output feature map of the kite convolution, and s is the convolution step size;

[0084] Finally, the concatenated tensor is normalized by a convolution kernel of W(2,2,c2). The height and width of the output feature map are adjusted to the preset values ​​h2 and w2, and the final output is Y(h2,w2,c2). The calculation formula is as follows:

[0085]

[0086] PConv's effective receptive field exhibits Gaussian decay characteristics. When the kernel size k = 3, the actual receptive field expands to 25 pixels, and the convolution weight of the central region is significantly higher than that of the periphery. This allows for greater focus on central features, which aligns with the characteristic of small object feature concentration. This allows for better focus on key features in tasks such as small object detection, improving feature extraction accuracy while achieving lightweightness. This is particularly suitable for detail-sensitive tasks such as small object detection.

[0087] 1.2 Iterative Attention Feature Fusion

[0088] In object detection tasks driven by deep convolutional neural networks, feature fusion is the core step in integrating multi-scale and multi-semantic information. The quality of this fusion directly determines the model's representational capabilities and detection performance. However, traditional feature fusion methods (such as simple summation and channel concatenation) often overlook the semantic differences and scale mismatches between features at different levels. This is particularly true for small remote sensing target detection, where edge features are difficult to extract, and the large size variations and small pixel count of small targets further exacerbate the difficulty of feature fusion, leading to feature conflicts or information loss during the fusion process.

[0089] To address the above problems, this embodiment proposes to fuse iterative attention feature fusion (iAFF) with the C3k2 module, that is, Figure 2 The internal structure of C3k2-iAFF is as follows: Figure 5 As shown in (c), dynamic feature selection and adaptive fusion are achieved through a multi-level attention mechanism, and the semantic association and spatial distribution differences between features are dynamically calibrated, making the feature fusion process more refined to enhance the feature expression of small targets. The iterative attention feature fusion module (iAFF) uses the multi-scale channel attention sub-unit (Multi-Scale Channel Attention Module, MS-CAM) iteratively, as shown in Figure 5 As shown in (a), dynamic weights are generated for the initial interaction features, and the output weights are used to weight the inputs X and Y element by element, strengthening the key feature channels and suppressing redundant information. Then, the features after the first attention calibration are processed again. Through iterative optimization, the feature fusion weights are gradually refined to avoid information bias in the initial fusion. Each round of iteration dynamically adjusts the fusion ratio of X and Y based on the previous results, and finally outputs the fused features. Based on this, this embodiment segments the input features and performs deep extraction on the segmented single-path features to optimize feature fusion, and constructs the Bottleneck_iAFF unit, whose structure is shown in the figure. Figure 5(b) The C3k2-iAFF module architecture systematically integrates the Bottleneck_iAFF unit to form an efficient feature processing pipeline. The architecture first preprocesses the input features through convolution and then distributes the features to multiple branches through a Split operation. Each branch independently connects to the Bottleneck_iAFF unit to achieve parallel iterative optimization of multi-scale features: Deep branches use the Bottleneck_iAFF unit to enforce global semantic consistency, while shallow branches use the Bottleneck_iAFF unit to preserve detailed features. Finally, the multi-branch features processed by the Bottleneck_iAFF unit are aggregated through Concat and then unified through convolution.

[0090] 1.3 Cascading Group Enhanced Attention

[0091] The C2PSA module is a core component in YOLO11 for enhanced feature extraction. Its design combines the CSP (CrossStage Partial) architecture with the Pyramid Squeeze Attention (PSA) mechanism. Multi-scale convolution kernels are introduced to process input features in parallel, with each branch capturing spatial information with a different receptive field. However, when dealing with small objects, feature extraction is susceptible to background interference, resulting in insufficient capture of detailed information and insufficient global semantic aggregation efficiency, which in turn hinders the full complementarity of multi-scale features.

[0092] To address this issue, this embodiment adopts a strategy of combining the cascaded group attention mechanism (CGA) with the C2PSA module to construct a C2PSA_CGA cascade group enhanced attention module (i.e., C2PSA_CGA module), as shown in the following example: Figure 6 (a) shows the main idea of ​​the CGA unit. The main idea of ​​the CGA unit is to enhance the diversity of features input to the attention head. Unlike previous self-attention, it provides different input splits for each head and cascades the output features across heads. Specifically, the CGA unit divides the input features into different parts, each of which is input to an attention head. Each head calculates its self-attention map and then concatenates the outputs of all heads. In this way, the output of each head is added to the input of the next head, thereby gradually refining the feature representation. Finally, they are projected back to the input dimension through a linear layer. In this way, the CGA unit not only reduces the computational redundancy in multi-head attention, but also improves the model capacity by increasing the network depth.

[0093] A new feature fusion processing framework is constructed by fusing the C2PSA module and the CGA unit, such as Figure 6As shown in (b). First, the input features are processed in parallel by C2PSA's multi-scale convolution to capture spatial information in different receptive fields. Then, the CGA unit is introduced to split the features into attention heads. Through independent self-attention calculations and cross-head cascade operations, the feature representation is gradually refined. Finally, cross-layer feature fusion is achieved through convolution. Compared with C2PSA, the C2PSA_CGA module uses the CGA unit's input splitting strategy to avoid duplicate extraction problems at the feature processing level. In terms of computational efficiency, the CGA unit is used to reduce QKV channel redundancy. In terms of model depth, the cross-head cascade of the CGA unit expands learning capacity. This can significantly improve feature diversity, balance computational efficiency and model expressiveness, and innovate the cascaded attention mechanism to strengthen the focus on key information.

[0094] The C2PSA_CGA module encourages each attention head to focus on the feature details of different areas, gradually integrating edge information in the cascade process, significantly improving the feature discontinuity problem caused by the difficulty in extracting edge features of small targets. At the same time, it enhances feature diversity while maintaining computational efficiency, enabling the network to more accurately capture diverse features such as texture and structure in remote sensing scenes.

[0095] 2. Multi-scale shared convolutional detection head

[0096] In target detection tasks, small targets have low pixel ratios and insufficient feature resolution, making it difficult to retain sufficient edge, texture and other detail information. In addition, the lack of semantic information can easily lead to problems such as feature discontinuity and feature misalignment. Traditional detection heads are limited to single or limited-scale feature processing modes. They are unable to fully utilize the underlying high-resolution details, and are also difficult to efficiently integrate high-level semantic features, resulting in poor detection performance. To address this problem, this embodiment designs a multi-scale shared convolutional detection head (MS-SCD). Its main concept is to add a small target detection layer (P2) and a shared enhancement strategy (SCD-Strategy), such as Figure 7 shown.

[0097] 2.1 Multi-scale Sharing Strategy

[0098] Multi-scale sharing builds a multi-layered feature perception system by adding a small target detection layer. The bottom detection layer focuses on capturing high-resolution details of small targets, preserving key information such as edges and textures. The middle detection layer gradually integrates semantic and detailed features to balance the expression requirements of different scales. The high-level detection layer focuses on the strong semantic features of medium and large targets. This multi-layer architecture, from low-level details to high-level semantics, enables refined feature modeling of small targets, effectively addressing the shortcomings of traditional detection heads in capturing small target features.

[0099] The specific process of the multi-scale shared convolution detection strategy is as follows: the multi-scale input first undergoes detail enhancement convolution (DEConv), then the efficient channel attention mechanism (ECA) enhances the response of the feature channel, and finally undergoes detail enhancement convolution again to form an extraction-optimization-reextraction chain. Multi-dimensional feature fusion and re-parameterization operations gradually enhance the expressive power of features, obtaining outputs for targeted processing of directional detection, bounding box regression, and classification regression, thereby improving the detection accuracy of multi-scale targets in complex scenarios.

[0100] Here, the Detect module is the downstream execution link in the MS-SCD structure. It completes the final detection decision based on the optimized feature map, and completes the output of target classification, boundary regression and rotation angle prediction results.

[0101] In addition, the shared design of DEConv and ECA avoids parameter redundancy, maintains efficient model operation in deep feature processing, and achieves a balanced optimization of detection accuracy and computational cost.

[0102] 2.2 Detail Enhancement Convolution (DEConv)

[0103] In the target detection task, small targets are difficult to detect because of the loss of high-frequency detail features such as edges and contours. Traditional convolution is limited to a single feature extraction mode and is difficult to effectively capture multi-dimensional detail information. Therefore, this embodiment introduces detail enhancement convolution (DEConv) in the shared strategy module. Figure 8 As shown, a new solution is provided to solve this problem.

[0104] Traditional convolution has limited ability to capture high-frequency details (such as edges and contours) in feature extraction, and high-frequency information (such as edges and contours) is crucial for image quality restoration and quality enhancement. Inspired by this, DEConv integrates vanilla convolution (VC) with multiple differential convolutions, including center differential convolution (CDC), angular differential convolution (ADC), horizontal differential convolution (HDC), and vertical differential convolution (VDC), to construct a multi-dimensional feature extraction framework to make up for the shortcomings of a single convolution. This multi-dimensional feature extraction mechanism directly compensates for the lack of detailed representation in target detection, especially for small targets, provides richer feature input for accurate target identification, and significantly improves the accuracy of small target detection.

[0105] To address the parameter redundancy problem caused by multiple parallel convolutions, DEConv uses reparameterization technology to effectively fuse parallel convolution kernels into a single standard convolution kernel. The relationship between its input and output is shown in the following formula:

[0106]

[0107] DEConv's re-parameterized structure enables the detection head to fully extract target details with the help of multiple branches when processing multi-scale features, while also maintaining efficient model operation through re-parameterization. This ultimately improves the detection accuracy and generalization capability of the entire target detection system for small targets, becoming the core supporting module for the performance breakthrough of the multi-scale shared detection head of this embodiment. For the multi-scale shared detection head studied in this embodiment, DEConv's re-parameterization feature perfectly adapts to lightweight requirements, significantly reducing the number of model parameters while ensuring the extraction of detailed features of multi-scale targets (especially small targets), thereby improving detection efficiency.

[0108] Example 2

[0109] The present invention discloses a method for constructing a remote sensing small target directional detection model, including:

[0110] Step 1: construct a remote sensing small target directional detection model described in Example 1;

[0111] Step 2: Construct the Inner_SIoU loss function and optimize the remote sensing small target directed detection model constructed in step 1;

[0112] The Inner_SIoU loss function is as follows:

[0113] Inner_SIoU=(1-α·IoU)+λ·(Angle_Cost+Distance_Cost+Shape_Cost)

[0114] Loss = 1-Inner_SIoU

[0115] Among them, α is the dynamic adjustment factor, λ is the dynamic weight coefficient, Angle_Cost is the angle cost, Distance_Cost is the distance cost, Shape_Cost is the shape cost, and Loss is the loss value.

[0116] The SIoU loss function, which is sensitive to the angle of the rotating frame and contains multi-dimensional constraints, is fused with the Inner-IoU loss function with a dynamic scaling mechanism to obtain the Inner-SIoU loss function, which includes an angle-aware weight adjustment module and dynamically balances angle errors. This enhances the ability to cope with angle constraints and scale changes in oriented target detection.

[0117] In response to the problems of small remote sensing targets often being embedded in complex backgrounds and having scarce pixels, the traditional IoU (Intersection over Union) method is easily affected by background interference, resulting in positioning deviations, making it difficult to meet the requirements of high-precision positioning and direction discrimination. In addition, in target detection scenarios with directions, although the introduction of angle parameters can more accurately describe the target direction, it leads to the problem of inconsistency between the measurement value and the loss. The CIoU in YOLO11-obb is developed based on the IoU (Intersection over Union) loss function. It still has some limitations. For example, when two boxes do not intersect, the IoU is 0, which cannot reflect the distance information between the boxes.

[0118] To address these issues, this application proposes the Inner_SIoU loss function, which combines the advantages of both methods to solve existing problems: First, the Inner_IoU scaling strategy improves the anti-interference ability of small target overlap calculations; second, combined with the SIoU angle penalty mechanism, it accurately constrains target direction deviations while optimizing bounding box spatial alignment using its distance and shape costs. This fusion solution not only addresses the metric and loss inconsistency caused by angle parameters in oriented detection, but also achieves coordinated optimization of anti-interference positioning and directional discrimination, providing a more robust solution for remote sensing oriented small target detection, accelerating model convergence, and promoting multi-dimensional constraint alignment.

[0119] 1. Inner_IoU loss function

[0120] Traditional IoU measures positioning accuracy by calculating the intersection-over-union ratio between the predicted box and the true box, but its sensitivity to background interference leads to significant deviations in the calculation of overlapping areas of small targets. Since small targets are sparsely pixelated and often embedded in complex backgrounds, traditional IoU is easily affected by background noise, resulting in an artificially high intersection-over-union ratio, which in turn causes regression errors. To solve this problem, Inner_IoU proposes adjusting the size of the bounding box by a scaling factor ratio, and then calculating the scaled IoU. This method can reduce background interference, especially in small target detection, and can focus more on the target itself, so that the calculated IoU value can better reflect the actual overlap of the target, providing a more accurate basis for the positioning of small targets. Its loss function is defined as:

[0121]

[0122] Union=w1·h1·ratio 2 +w2·h2·ratio 2 -Inter+eps

[0123] Inter=max(0,min(b1 x2 , b2 x2 )-max(b1x1 b2 x1 ))

[0124] ×max(0,min(b1 y 2b2 y2 )-max(b1 y1 , b2 y1 ))

[0125] By introducing a scaling factor (ratio), the bounding box size is adaptively adjusted to focus on the core area of ​​the target and suppress background interference. The core formula is shown below:

[0126]

[0127] Among them, (x, y, w, h) represents the center coordinates and width and height of the original bounding box, and ratio is the scaling factor.

[0128] Inner-IoU adaptively adjusts the bounding box size by introducing a scaling factor, effectively suppressing the interference of complex background in remote sensing small target detection, focusing the bounding box on the core area of ​​the target, and making the intersection-over-union calculation closer to the actual target overlap.

[0129] 2. SIoU loss function

[0130] Traditional IoU measures positioning accuracy by intersection over union (IoU), but it only focuses on geometric overlap and ignores differences in direction, distance, and shape, resulting in limited detection performance for small directional targets in remote sensing. Therefore, SIoU introduces angle cost (Angle_Cost), distance cost (Distance_Cost), and shape cost (Shape_Cost) based on traditional IoU to construct multi-dimensional geometric constraints and solve the problems of directional sensitivity and spatial alignment in target detection. Angle cost (Angle_Cost), as the core component of the SIoU loss function, realizes quantitative constraints on the directional deviation between the predicted box and the true box through precise mapping of geometric transformation and trigonometric functions, quantifies the angular deviation between the predicted box and the true box through cosine similarity, and constructs directional discriminative loss. Its calculation formula is as follows:

[0131]

[0132] where θ p and θ g are the rotation angles of the predicted box and the true box, respectively. This formula maps the angle difference to the interval [0,1]. The larger the angle deviation, the higher the cost value, forcing the model to strengthen the regression accuracy of the rotation angle during training.

[0133] The distance cost (Distance_Cost) eliminates the interference of rotation on the center point distance measurement through polar coordinate transformation. Traditional Euclidean distance is easily affected by posture in target directional detection, resulting in positioning error. The center point coordinates are converted to polar coordinates and the normalized distance cost is defined. Its calculation formula is as follows:

[0134]

[0135] where c w and c h The width and height of the minimum bounding box between the predicted box and the ground-truth box. This strategy normalizes the center point distance to the target scale to avoid distance measurement failure caused by rotation.

[0136] Shape Cost (Shape_Cost) is a dynamic correction module designed to address the aspect ratio distortion caused by rotation. Traditional IoU is insensitive to the shape changes of rotated objects, resulting in aspect ratio estimation bias. The exponential function is used to strengthen the shape difference penalty. The calculation formula is as follows:

[0137]

[0138] where w max =max(w p , w g ), h max =max(h p , h g The fourth power operation causes the shape difference to be nonlinearly amplified, significantly improving the model's robustness to shape distortion. The final SIoU loss calculation formula is as follows:

[0139] SIoU=1-IoU+Angle_cost+Distance_cost+Shape_cost.

[0140] 3. Fusion of Inner-IoU and SIoU loss functions

[0141] With Inner_IoU as the core, we utilize its bounding box scaling strategy to reduce background interference in remote sensing small target detection, ensure that overlap calculations focus on the target body, and improve positioning benchmark accuracy. SIoU's angle cost, distance cost, and shape cost are introduced through dynamic weight coefficients. This allows the model to not only focus on the overlap of the target core area during optimization, but also enforce constraints on the spatial distance deviation and shape difference between the predicted box and the true box, meeting the multi-dimensional optimization requirements of remote sensing small directional targets for direction, position, and shape. A more rotation-aware dynamic loss function, the Inner_SIoU loss function, is constructed, and its reasoning process is as follows:

[0142] 1) When calculating the overlap of small remote sensing targets, traditional IoU is easily affected by complex background interference, resulting in measurement deviation. This embodiment adopts the Inner-IoU scaling bounding box strategy, suppressing background noise through dynamic adjustment factors to achieve precise focus on the core area of ​​the small target:

[0143]

[0144] Where δ and β are adaptive parameters. The nonlinear characteristics of the exponential function make α approach 0 as the IoU value increases. When the predicted box and the ground-truth box highly overlap, α approaches 0 and the traditional IoU calculation is directly used. When the overlap is low (such as in small object detection scenarios), α approaches 1 and the bounding box is scaled to reduce background interference.

[0145] 2) When dealing with rotated objects, traditional IoU loss often leads to optimization deviations due to inconsistencies between the angle parameters and the loss function. An adaptive weight allocation strategy is used to achieve consistent alignment between the metric value and the loss function, defining a dynamic weight coefficient λ, which is defined as follows:

[0146]

[0147] This formula dynamically associates λ with the IoU value and the angle cost: when the angle deviation is small, λ increases with the IoU, strengthening the geometric constraint; when the angle deviation is large, λ approaches 0, avoiding the loss explosion caused by the angle error.

[0148] 3) Combining Inner-IoU with SIoU, the loss calculation formula for Inner_SIoU is as follows:

[0149] Inner_SIoU=(1-α·IoU)+λ·(Angle_Cost+Distance_CosT_Shape_Cost)

[0150] Loss = 1-Inner_SIoU

[0151] The Inner_SIoU value is converted into a loss value. When the predicted box and the ground-truth box are highly matched in terms of interference-free overlap, distance, and shape (Inner-SIoU approaches 1), the loss value approaches 0. The model training goal is to minimize this loss, forming a clear optimization guide. Inner-SIoU uses a fusion strategy to retain the interference-free performance of Inner-IoU for small objects. It also effectively resolves the inconsistency between metric and loss caused by angle parameters in oriented detection through SIoU's multi-dimensional cost constraints, achieving consistent alignment between metric and loss, providing a more precise optimization criterion for object detection in complex remote sensing scenarios.

[0152] Example 3

[0153] This embodiment discloses a method for detecting directed small remote sensing targets, which uses a directed small remote sensing target detection model constructed by the method for constructing a directed small remote sensing target detection model described in Example 2 for detection.

[0154] The present invention is experimented below, and the experimental results and analysis are as follows:

[0155] 1. Experimental environment and parameter configuration

[0156] The experiments in this paper used Python as a development tool and developed and trained deep learning models based on the PyTorch framework. The specific configuration of the experimental environment is shown in Table 1, and the experimental parameter settings are detailed in Table 2. The DIOR-R and UCAS-AOD datasets were used to evaluate the performance of the model.

[0157] Table 1 Specific configuration of the experimental environment

[0158]

[0159] Table 2 Parameter settings used in the experiment

[0160]

[0161]

[0162] 2. Dataset Introduction

[0163] The present invention uses the DIOR-R and UCAS-AOD datasets equipped with rotated target annotations to test and compare the proposed method. As a derivative of the DIOR dataset, the DIOR-R dataset includes 23,463 images and 190,288 annotation instances, covering a wide range of typical remote sensing environments and 20 daily target categories, including aircraft (APL), airports (APO), baseball fields (BF), and basketball courts (BC). The UCAS-AOD dataset contains 1,510 aerial images and 14,596 rotated annotation instances, focusing on two typical targets in complex aerospace scenes: vehicles (Car) and aircraft (Airplane). As a representative dataset in the early field of directed detection, although UCAS-AOD has fewer target categories, its high-resolution images and refined rotated annotations can provide effective verification for the robustness of small target directed detection algorithms.

[0164] 3. Evaluation Criteria

[0165] The experimental evaluation uses several key indicators, including precision (P), recall (R), average precision (AP), mean average precision (mAP), the core indicator of complexity calculation GFLOPs, and the number of model parameters (Params), aiming to comprehensively compare and analyze the performance of target detection models. Precision is the ratio of the number of positive instances correctly identified by the model to the total number of predicted positive instances. Recall measures the ratio of the number of positive instances correctly identified by the model to the total number of actual positive instances. The parameter count is the sum of the parameters of all network layers during the model training process. The indicator calculation formula is as follows:

[0166]

[0167] Where: true positive (TP), true negative (TN), false positive (FP) and false negative (FN), N is the number of categories in the dataset, and the evaluation unit of P, R, AP, and mAP is "%.

[0168] 4. Ablation Experiment

[0169] In order to verify the optimization effect of each module improved by the present invention and explore the effectiveness of the method, an ablation experiment was conducted on the basis of YOLO11-obb. The results of the ablation experiment on the DIOR-R dataset are shown in Table 3, where A = YOLO11-obb, B = PConv, C = C3k2-iAFF, D = C2PSA-CGA, E = MS-SCD, E1 = P2, E2 = SCD, F = Inner_SIoU, MS-SCD is composed of the P2 small target detection layer and the SCD shared enhancement strategy, "√" indicates that the improved method is used, and "-" indicates that the improved method is not used. The experimental results are shown in Table 3. The improvement effect of the method proposed by the present invention can be more intuitively seen by expressing the ablation experiment results in a bar graph, as shown in Figure 9 shown.

[0170] Table 3 Ablation experiment results

[0171]

[0172] Experimental results show that when using PConv and C3k2-iAFF modules, the model P value is as high as 87.3%, and mAP50 is increased to 85.5%, verifying the module's optimization effect on feature enhancement and fusion; after further superimposing modules such as C2PSA-CGA and MS-SCD (P2+SCD), the R value rises to 82.4%, mAP50 reaches 86.5%, and the number of parameters drops to 2.30M, achieving a balance between detection accuracy and model parameters, verifying the module's role in enhancing the representation of small target features and lightweighting; when all modules (PConv, C3k2-iAFF, C2PSA-CGA, MS-SCD, Inner-SIoU) are enabled in collaboration, the P value reaches 86.6%, the R value reaches 82.7%, and the mAP50 rises to 86.8%, achieving a significant improvement in detection accuracy.

[0173] 5. Comparative Experiment

[0174] To validate the performance of the FEMS-YOLO algorithm on directional detection datasets, comparative experiments were conducted on the DIOR-R dataset and UCAS-AOD. The results are shown in Tables 4 and 5. It is clear that FEMS-YOLO demonstrates significant advantages in detection accuracy compared to other mainstream models. In Table 4, APL, APO, etc. are abbreviations for class labels in the DIOR-R dataset.

[0175] Table 4 Comparative experimental results of DIOR-R dataset

[0176]

[0177] Converting the comparative experimental data in Table 4 into a heat map can more intuitively show the advantages of the improved FEMS-YOLO model. The DIOR-R comparative experiment heat map is as follows: Figure 10 shown.

[0178] Table 5 Comparative experimental results of UCAS-AOD

[0179]

[0180] Converting the comparative experimental data in Table 5 into a histogram can more intuitively show the advantages of the improved FEMS-YOLO model. The comparative experimental histogram of the UCAS-AOD dataset is as follows: Figure 11 shown.

[0181] Experimental results on the DIOR-R and UCAS-AOD datasets show that FEMS-YOLO demonstrates significant advantages in the task of oriented target detection. In the DIOR-R dataset (containing 20 categories of complex remote sensing targets), FEMS-YOLO's mAP50 reached 86.8%, an increase of 2.1% over the baseline model YOLO11-obb (84.7%), and outperformed other mainstream detection models. The average accuracy of 17 of its 20 categories of targets ranked first. In the UCAS-AOD dataset, FEMS-YOLO's mAP50 was 97.2%, surpassing the baseline YOLO11-obb (96.9%). Its AP values ​​for the two categories of targets were 95.6% and 98.9% respectively, further demonstrating the robustness of the model in specific areas. Compared to mainstream single-stage (e.g., RetinaNet-O, ​​82.1%) and two-stage models (e.g., RoI Transformer, 88.9%), the FEMS-YOLO proposed in this embodiment achieves significant detection performance. Comparative experiments on two datasets validate the effectiveness of the proposed FEMS-YOLO: based on feature enhancement and multi-scale sharing strategies, it achieves higher detection stability than the single-stage framework. This fully demonstrates the superiority of the improvements proposed in this embodiment for directional detection tasks, achieving a balance between detection accuracy and parameter count.

[0182] 6. Visualization Results Analysis

[0183] Visual analysis can more intuitively show the detection effect of the model, such as Figure 12-15 As shown in the figure, the visualization results of YOLO11-obb and FEMS-YOLO on the DIOR-R dataset and the comparison results of the Precision-Recal l curve are shown.

[0184] Figure 12 Figure 1 shows a visual comparison of the improved results for a remote sensing image with dense small objects. (a) and (c) show the detection results of YOLO11-obb (the red circles in the figures indicate undetected objects), while (b) and (d) show the detection results of FEMS-YOLO for the same images. Compared to the base model, FEMS-YOLO successfully identifies and locates all objects in the image with no false detection boxes.

[0185] Figure 13The following visual comparison results are obtained for different categories of small objects in remote sensing images, with significant scale variations. (a) shows the detection result using YOLO11-obb, and (b) shows the detection result using FEMS-YOLO for the same image. While the large scale variations in the bounding boxes of ships and ports can lead to missed detections, the improved model does not affect detection due to scale variations, demonstrating the effectiveness of this model in detecting directional small objects.

[0186] like Figure 14 As shown, the model's improvement slowed down in the middle and late stages of training, gradually approaching convergence. However, FEMS-YOLO entered a gentle upward trend earlier and outperformed YOLO11-obb in terms of convergence speed, demonstrating the effectiveness of the loss function improvements proposed in this example. Finally, FEMS-YOLO achieved a higher mAP50 than YOLO11-obb, verifying the effectiveness of this example's feature enhancement improvements.

[0187] Analysis from the PR curve comparison chart: Figure 15 (a) shows that the mAP50 of YOLO11-obb is 84.7% for all categories, while FEMS-YOLO improves to 86.8%. In addition, Figure 15 In (b), the blue curve representing the comprehensive performance of all categories of FEMS-YOLO is generally higher than the YOLO-obb curve, indicating that at the same recall rate, FEMS-YOLO has better precision. For example, when the recall rate is close to 0.8, it still maintains a high precision, while the precision of YOLO-obb has dropped significantly, reflecting that FEMS-YOLO has better control over false detections. In the UCAS-AOD dataset, FEMS-YOLO's mAP50 is 97.2%, further confirming that FEMS-YOLO outperforms the baseline algorithm YOLO11-obb in overall performance, proving the effectiveness of the improvement.

[0188] This application addresses the measurement and loss inconsistencies caused by feature discontinuity, feature misalignment, and directional detection angle parameters in remote sensing small target detection. We propose a FEMS-YOLO method for directional small target detection based on feature enhancement and multi-scale sharing. This method leverages the asymmetric image padding operation of kite-enhanced convolution and iteratively uses an attention feature fusion mechanism to construct a feature fusion refinement module to enhance small target edge feature extraction and detail perception. A cascaded group attention mechanism is used to fully integrate small target edge information features and enhance feature diversity. A multi-scale shared convolutional detection head is designed, and through a multi-layered feature fusion architecture and detail-enhanced convolution, high- and low-level feature information fusion and semantic features are effectively integrated. The SIoU loss function, which is sensitive to the angle of the rotating frame and contains multi-dimensional constraints, is fused with the Inner-IoU loss function, which has a dynamic scaling mechanism. This results in a new Inner-SIoU loss function that includes an angle-aware weight adjustment module and dynamically balances angle error and penalty coefficients. This strengthens the joint optimization of angle constraints and scale changes in directional target detection. Experimental results show that the improved model achieved 86.6% precision, 82.7% recall and 86.8% mAP50 on the DIOR-R dataset, which are 0.7%, 2.7% and 2.1% higher than the baseline model YOLO11-obb respectively. At the same time, the number of parameters was reduced by 13%. It outperformed other mainstream detection models in terms of average precision and achieved 97.2% mAP50 on the UCAS-AOD dataset, verifying the effectiveness of the method proposed in this application.

[0189] Of course, the above contents are only preferred embodiments of the present invention and should not be considered as limiting the scope of the embodiments of the present invention. The present invention is not limited to the above examples. Equivalent changes and improvements made by ordinary technicians in this technical field within the essential scope of the present invention should all fall within the scope of the patent of the present invention.

Claims

1. A remote sensing small target directional detection model, characterized by: include: Backbone network: including kite convolution, C3k2_iAFF module and C2PSA_CGA module, used to perform preliminary feature extraction on the input image, providing a basis for subsequent feature fusion and detection; Neck network: includes kite convolution and C3k2_iAFF modules, located between the backbone network and the head network, used to fuse the features of different levels extracted by the backbone network to generate feature representation; Head network: includes a multi-scale shared convolutional detection head, which is used for multi-scale shared feature fusion; Kite convolution, C3k2_iAFF module, and C2PSA_CGA module enhance the feature expression and recognizability of small objects by performing asymmetric image padding feature extraction operations, iteratively using attention mechanisms, and cascading edge information. The multi-scale shared convolutional detection head uses multi-scale feature fusion and shared enhancement strategy to construct a detection head with a multi-scale shared detail enhancement structure to enhance the ability to perceive details of small targets.

2. A remote sensing small target directional detection model according to claim 1, characterized in that: The first layer of the kite convolution uses four parallel convolution branches to achieve multi-directional feature capture by differentiating padding parameters and kernel sizes. Each branch has different padding parameters and convolution kernels to achieve feature extraction in different directions. The calculation formula is as follows: in, is the convolution operator, Wi is the corresponding convolution kernel, and the four parameters in P represent the number of padded pixels in the left, right, up, and down directions respectively; After the first convolution layer, the height h′, width w′, and number of channels c′ of the output feature map are related to the height h1, width w1, and number of channels c1 of the input feature map. The calculation formula is as follows: Among them, c2 is the number of channels of the final output feature map of the kite convolution, and s is the convolution step size; The concatenated tensor is normalized by a convolution kernel of W(2,2,c2). The height and width of the output feature map are adjusted to the preset values ​​h2 and w2, and the final output is Y(h2,w2,c2). The calculation formula is as follows:

3. A remote sensing small target directional detection model according to claim 1, characterized in that: The C3k2_iAFF module first pre-processes the input features through convolution, and then distributes the features to multiple branches through the Split operation; each branch is independently connected to the Bottleneck_iAFF unit to achieve parallel iterative optimization of multi-scale features: the deep branches use the Bottleneck_iAFF unit to enhance global semantic consistency, and the shallow branches use the Bottleneck_iAFF unit to retain detailed features; finally, the multi-branch features processed by the Bottleneck_iAFF unit are aggregated through Concat and then unified through convolution.

4. A remote sensing small target directional detection model according to claim 3, characterized in that: The Bottleneck_iAFF unit dynamically generates weights for the features of the preliminary interaction by iteratively using the multi-scale channel attention sub-units, and uses the output weights to weight the inputs X and Y element by element, strengthening the key feature channels and suppressing redundant information. Then, the features after the first attention calibration are processed again; through iterative optimization, the feature fusion weights are gradually refined to avoid information bias in the initial fusion; each round of iteration dynamically adjusts the fusion ratio of X and Y based on the previous results, and finally outputs the fused features.

5. A remote sensing small target directional detection model according to claim 1, characterized in that: The C2PSA_CGA module constructs a new feature fusion processing framework by fusing the C2PSA module with the CGA unit. First, the input features are processed in parallel by multi-scale convolution of the C2PSA module to capture the spatial information of different receptive fields. Then, the CGA unit is introduced to split the features into each attention head. After independent self-attention calculation and cross-head cascade operation, the feature representation is gradually refined. Finally, cross-layer feature fusion is achieved through convolution.

6. A remote sensing small target directional detection model according to claim 5, characterized in that: The CGA unit divides the input features into different parts, each of which is input to an attention head; each head calculates its own attention map, and then concatenates the outputs of all heads. In this way, the output of each head is added to the input of the next head, thereby gradually refining the feature representation; finally, a linear layer is used to project them back to the input dimension.

7. A remote sensing small target directional detection model according to claim 1, characterized in that: The multi-scale shared convolution detection head adds a small target detection layer and a shared enhancement strategy unit to build a multi-level feature perception system.

8. A remote sensing small target directional detection model according to claim 7, characterized in that: The bottom detection layer of the small target detection layer focuses on capturing high-resolution details of small targets and retaining key information such as edges and textures; the middle detection layer gradually integrates semantic and detail features to balance the expression requirements of different scales; the high-level detection layer focuses on the strong semantic features of medium and large targets; The specific workflow of the shared enhancement strategy unit is as follows: the multi-scale input first undergoes detail enhancement convolution, then the response of the feature channel is enhanced by an efficient channel attention mechanism, and finally it is processed again by detail enhancement convolution to form an extraction-optimization-reextraction chain. Multi-dimensional feature fusion and re-parameterization operations gradually enhance the expressive power of features, obtain targeted processing of directional detection, bounding box regression, and classification regression output, and improve the detection accuracy of multi-scale targets in complex scenarios.

9. A method for constructing a remote sensing small target directional detection model, characterized in that: include: Step 1: construct a remote sensing small target directional detection model as described in any one of claims 1 to 8; Step 2: Construct the Inner_SIoU loss function and optimize the remote sensing small target directed detection model constructed in step 1; The Inner_SIoU loss function is as follows: Inner_SIoU=(1-α·IoU)+λ·(Angle_Cost+Distance_Cost+Shape_Cost) Loss = 1-Inner_SIoU Among them, α is the dynamic adjustment factor, λ is the dynamic weight coefficient, Angle_Cost is the angle cost, Distance_Cost is the distance cost, Shape_Cost is the shape cost, and Loss is the loss value.

10. A method for detecting directional small remote sensing targets, characterized in that: Detection is performed using the remote sensing small target directional detection model constructed by the method for constructing a remote sensing small target directional detection model described in claim 9.

Images