A directional detection model, detection method, and construction method for small remote sensing targets

Abstract

This invention relates to a directed detection model, detection method, and construction method for small remote sensing targets, belonging to the field of remote sensing image processing technology. Based on feature enhancement, it utilizes the asymmetric image filling operation of kite convolution and iteratively uses an attention feature fusion mechanism to construct a refined feature fusion module to enhance the ability to extract edge features and perceive details of small targets. A cascaded group attention mechanism is used to fully integrate edge information features of small targets, enhancing feature diversity. A multi-scale shared convolutional detection head is designed, and a multi-level feature fusion architecture and detail-enhancing convolution are used to achieve the fusion enhancement of high and low-level feature information and the effective integration of semantic features. Inner-IoU and SIoU are fused, and an angle-distance-shape joint loss function is designed to effectively solve the problem of geometric feature blurring caused by pixel loss of small targets and the inconsistency between the metric value and loss caused by rotation parameters, thereby enhancing the positioning accuracy and robustness of directed target detection.

Description

Technical Field

[0001] This invention relates to a directional detection model, detection method, and construction method for small remote sensing targets, belonging to the field of remote sensing image processing technology. Background Technology

[0002] Remote sensing small target detection is of great significance in fields such as land and resources monitoring, urban fine planning, and ecological environmental protection. Remote sensing small target detection technology has undergone rapid development over the past few decades, gradually transitioning from early methods based on handcrafted features to end-to-end solutions dominated 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 main method for detecting small targets in remote sensing. This method relies on manually designed feature descriptors (such as HOG features) and traditional machine learning classifiers (such as SVM). Although it can achieve target detection to a certain extent, its detection accuracy and generalization ability are limited because handcrafted features cannot fully and accurately represent complex remote sensing image information.

[0004] Subsequently, the rise of deep learning brought new breakthroughs to remote sensing small target detection. The two-stage algorithm Faster R-CNN emerged, significantly improving target localization accuracy by introducing a Region Proposal Network (RPN). However, Faster R-CNN has significant limitations when handling small targets, as its small target detection performance is constrained by the semantic deficiencies of shallow features. In convolutional neural networks, shallow features contain a lot of spatial information but relatively little semantic information. Small targets occupy fewer pixels in an image, and their features rely more heavily on shallow features, making it difficult to extract sufficient semantic information from the shallow features corresponding to small targets for accurate detection. Therefore, the Feature Pyramid Network (FPN) for multi-scale feature fusion was proposed. FPN fuses feature maps of different scales through bottom-up and top-down paths, enabling the network to acquire rich feature information at different scales. This method achieved multi-scale feature fusion for the first time, significantly improving the accuracy of small target detection. Subsequent research continued to develop in the direction of improving feature fusion strategies, such as SNIP and FSAF. SNIP alleviates the training conflict between targets of different scales by training and testing the model on images of different scales; FSAF adaptively selects appropriate feature layers for detection of targets of different scales, further optimizing the effect of feature fusion. Some studies have also introduced attention mechanisms, such as CBAM, which can 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 target detection. However, it should be noted that most of these methods are designed for horizontal bounding box detection and are difficult to adapt to the directional target detection requirements that are prevalent in remote sensing scenarios.

[0005] In the field of directed target detection technology, the RRPN proposed in 2016 was groundbreaking. It achieved, for the first time, the detection of directed targets in remote sensing scenes through a rotating anchor frame generation mechanism. However, the anchor frame preset angles used by RRPN are discrete, resulting in insufficient rotation sensitivity. Discrete angle presets cannot accurately adapt to the true angles of various rotating targets.

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

[0007] After 2020, anchor-free methods gradually gained popularity. Oriented RepPoints achieves continuous angle prediction through point set regression, avoiding the problem of pre-discretized anchor-bound angles, allowing the model to capture the target's rotation angle more accurately. However, the feature misalignment problem caused by small targets occupying fewer pixels remains unresolved, affecting the model's training performance. RoI Transformer improves the localization accuracy of rotating targets in dense scenes through cascaded rotational region of interest transformations. However, it also fails to effectively solve the feature misalignment problem.

[0008] The YOLO series of detectors has continuously evolved in the field of single-stage target detection, providing new approaches for directed remote sensing detection tasks. YOLOv5-obb improved the rotating anchor box matching strategy, enhancing detection efficiency and accuracy to some extent. However, it remains limited by the quantization error of discrete angle encoding, unable to accurately represent the target's rotation angle. YOLOv7-obb introduced an attention mechanism to enhance small target features; however, the insensitivity of traditional GIoU loss to rotation angles persists, resulting in poor geometric constraint performance in directed target detection.

[0009] In addition to the feature misalignment problem caused by the small number of pixels in small targets mentioned above, the traditional IoU loss is not sensitive to rotation angles. There are also problems such as the difficulty in extracting edge features due to the complex background information of remote sensing images, resulting in discontinuous extracted features, and the inconsistency between the metric value and the loss due to the introduction of angle parameters in directional detection.

[0010] For example, Chinese Patent Publication No. CN112329721A discloses a lightweight model design for remote sensing small target detection. This 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 good performance for remote sensing small targets. The model is then lightweighted. This lightweighting approach reduces the number of model parameters and memory consumption, allowing the lightweight model to maintain the accuracy of complex neural network models while having the same memory and computational power consumption as lightweight algorithms, resulting in better detection performance compared to other mainstream lightweight algorithms. However, remote sensing small target detection still faces challenges such as complex background information, difficulty in extracting edge features, large variations in target size, and a small number of pixels in small targets, leading to discontinuous and misaligned features. Furthermore, the introduction of angle parameters in directed detection results in inconsistencies between the metric and the loss. Summary of the Invention

[0011] The purpose of this invention is to provide a directed detection model, detection method, and construction method for small remote sensing targets, which solves the problems of complex background information, difficulty in extracting edge features, large variation in target size, and small target pixel count in small remote sensing target detection, leading to discontinuous and misaligned extracted features.

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

[0013] The backbone network includes the kite convolution, C3k2_iAFF module, and C2PSA_CGA module, which are used to perform preliminary feature extraction on the input image, providing a foundation for subsequent feature fusion and detection.

[0014] The neck network, consisting of kite convolution and C3k2_iAFF modules, is located between the backbone network and the head network. It is used to fuse features extracted from different levels by the backbone network to generate feature representations.

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

[0016] The kite convolution, C3k2_iAFF module, and C2PSA_CGA module enhance the feature representation and recognizability of small targets by using asymmetric image filling feature extraction operations, iteratively using attention mechanisms, and cascading integration of edge information.

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

[0018] By employing asymmetric padding to create horizontal and vertical convolution kernels for different regions of the image, the pixel Gaussian spatial distribution of small targets is better adapted, enhancing feature extraction and increasing the receptive field. Iteratively applying a multi-scale channel attention mechanism improves the initial integration of feature maps, effectively strengthening the feature representation and recognizability of small targets. A cascaded group attention mechanism is used to improve the diversity of attention maps and gradually integrate edge information during the cascading process, while reducing computational redundancy.

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

[0020] Preferably, the first layer of the kite convolution employs four parallel convolution branches, achieving multi-directional feature capture through differentiated padding parameters and kernel size; 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, For convolution, Wi is the corresponding convolution kernel, and the four parameters in P represent the number of fill pixels in the left, right, top, and bottom directions, respectively.

[0023] After the first convolutional layer, the relationship between the height h′, width w′, and number of channels c′ of the output feature map and the height h1, width w1, and number of channels c1 of the input feature map is calculated using the following formula:

[0024]

[0025] Where c2 is the number of channels in the final output feature map of the kite convolution, and s is the convolution stride;

[0026] The concatenated tensor is normalized using a W(2,2,c2) convolution kernel. The height and width of the output feature map are adjusted to preset values ​​h2 and w2, respectively. The final output is Y(h2,w2,c2), calculated 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 weights in the central region are significantly higher than those in the periphery. This allows for greater attention to central features, aligning with the characteristic of concentrated features in small objects. In tasks such as small object detection, it can better focus on key features, improving feature extraction accuracy while achieving lightweight design, making it particularly suitable for detail-sensitive tasks such as small object detection.

[0029] Preferably, the C3k2_iAFF module achieves dynamic feature selection and adaptive fusion through a multi-level attention mechanism, dynamically calibrating the semantic associations and spatial distribution differences between features, thus refining the feature fusion process and enhancing the expression of features of small targets. The C3k2_iAFF module 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 strengthen global semantic consistency, while 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.

[0030] Preferably, the Bottleneck_iAFF unit iteratively uses multi-scale channel attention subunits to dynamically generate weights for the features of the initial interaction, and uses the output weights to weight the inputs X and Y element-wise to strengthen key feature channels and suppress redundant information. Then, the features after the first attention calibration are processed again. Through iteration, the feature fusion weights are gradually refined to avoid information deviation in the initial fusion. Each iteration is based on the previous results and dynamically adjusts the fusion ratio of X and Y, finally outputting 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 by the multi-scale convolution parallel processing of the C2PSA module to capture the spatial information of different receptive fields. Then, the CGA unit is introduced to segment the features into each attention head. Through independent self-attention calculation and cross-head cascade operation, the feature representation is gradually refined. Finally, cross-layer feature fusion is achieved through convolution.

[0032] Compared to C2PSA, C2PSA_CGA avoids the problem of repeated extraction at the feature processing level by using CGA's input segmentation strategy, reduces QKV channel redundancy at the computational efficiency level by using CGA, and expands the learning capacity at the model depth level by using CGA's cross-head cascading. It can significantly improve feature diversity, balance computational efficiency and model expressive power, and innovate the cascaded attention mechanism to strengthen the focus of key information.

[0033] Preferably, the CGA unit divides the input features into different parts, with each part input to an attention head; each head computes its self-attention map, and then the outputs of all heads are concatenated. In this way, the output of each head is added to the input of the next head, thereby progressively refining the feature representation; finally, they are projected back to the input dimension through a linear layer.

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

[0035] We designed a multi-scale shared convolutional detection head, which achieves the fusion and enhancement of high and low-level feature information and the effective integration of semantic features through a multi-level feature fusion architecture and detail-enhancing 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 needs 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 channels is enhanced by an efficient channel attention mechanism, and finally it undergoes detail enhancement convolution again to form an extraction-optimization-re-extraction link. Multi-dimensional feature fusion and reparameterization operations gradually enhance the expressive power of the features, resulting in targeted processing for directional detection, bounding box regression, and classification regression outputs, thus improving the detection accuracy of multi-scale targets in complex scenes. The present invention provides a method for constructing a directional detection model for small remote sensing targets, comprising:

[0038] Step 1: Construct the above-mentioned directed detection model for small remote sensing targets;

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

[0044] By fusing SIoU, which is sensitive to the rotation box angle and contains multidimensional constraints, with Inner-IoU, which has a dynamic scaling mechanism, we obtain Inner-SIoU, a loss function that includes an angle-aware weight adjustment module and dynamically balances angle errors, in order to enhance the ability to cope with angle constraints and scale changes in directional target detection.

[0045] The present invention provides a directed detection method for small remote sensing targets, which uses a directed detection model for small remote sensing targets constructed by the aforementioned method for constructing such a model.

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

[0047] This invention discloses a directed detection model, method, and construction method for small remote sensing targets. Based on feature enhancement, it utilizes asymmetric image filling operations via kite convolution and iterative attention-based feature fusion mechanisms to construct a refined feature fusion module, thereby strengthening the extraction of edge features and detail perception of small targets. A cascaded group attention mechanism is used to fully integrate edge information features of small targets, enhancing feature diversity. A multi-scale shared convolutional detection head is designed, achieving effective integration of high- and low-level feature information fusion and semantic features through a multi-layered feature fusion architecture and detail-enhancing convolution. Inner-IoU and SIoU are fused, utilizing Inner-IoU... The dynamic scaling mechanism of SIoU suppresses the interference of background noise on small targets under rotational posture. At the same time, it combines a phased sample weighting strategy to accelerate the convergence process of rotating targets. It uses the rotation-sensitive angle penalty term of SIoU, the polar coordinate distance constraint of the center point, and the dynamic aspect ratio compensation to construct a multi-dimensional geometric constraint system. Combined with the scale-adaptive weight allocation strategy of Inner-IoU, the Inner-SIoU loss function is obtained. An angle-distance-shape joint loss function is designed to effectively solve the problem of geometric feature blurring caused by pixel loss of small targets and the inconsistency between the metric value and loss caused by rotation parameters, thereby enhancing the positioning accuracy and robustness of directional target detection. Attached Figure Description

[0048] Figure 1The images show the results of horizontal and directed detection; (a) is the original image, (b) is the horizontal detection result, and (c) is the directed detection result.

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

[0050] Figure 3 This is a structural diagram of a directional detection model for small remote sensing targets as described in this invention;

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

[0052] Figure 5 The figure shows a structural diagram of a C3k2-iAFF module according to 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 the CGA unit, and (b) is a structural diagram of the 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 an enhanced detail convolution according to the present invention;

[0056] Figure 9 This is a bar chart showing the ablation experiment results described in this invention;

[0057] Figure 10 This is a thermogram of the DIOR-R comparative experiment described in this invention;

[0058] Figure 11 This is a bar chart showing the comparative experiments using the UCAS-AOD dataset described in this invention.

[0059] Figure 12 This is a visualization comparison of the results before and after the improvement in the case of dense small targets in remote sensing images described in this invention; in the figure, (a) and (c) are the YOLO11-obb detection results, and (b) and (d) are the FEMS-YOLO detection results of the same image;

[0060] Figure 13This is a visualization comparison of the results before and after the improvement of the remote sensing image small target with different categories and large time scale changes as described in this invention; in the figure, (a) is the YOLO11-obb detection result, and (b) is the FEMS-YOLO detection result of the same image;

[0061] Figure 14 This is a comparative analysis chart of the mAP50 curve described in this invention;

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

[0063] The present invention will be further described in detail below with reference to 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's position, size, and rotation angle using a rotated bounding box. Compared to traditional horizontal detection, its core difference lies in the introduction of an angle parameter. Horizontal detection (such as...) Figure 1 As shown in Figure b), using axis-aligned rectangular boxes, while computationally efficient, has significant limitations in remote sensing scenarios. The horizontal boxes tend to include a large number of background pixels, especially for targets with significant aspect ratio differences or densely packed elements (such as vehicles and ships), leading to impure feature extraction, poor geometric fit, and difficulty in tightly fitting the target edges, thus affecting positioning accuracy. Directed detection (such as...) Figure 1 As shown in (c), the compact representation of the rotating frame effectively reduces background interference, improves detection robustness in dense scenes, and better reflects the spatial distribution characteristics of targets in remote sensing images. Since there is a difference between horizontal and directional detection results, this embodiment chooses directional detection of the target to be studied. Figure 1 As shown.

[0066] Among algorithms for directed object detection, the YOLO series, as a typical representative of single-stage object detection, offers significantly faster inference speeds compared to traditional two-stage algorithms (such as Faster R-CNN) and anchor-free methods (such as RepPoints) in remote sensing tasks. Its end-to-end architecture eliminates redundant computation in region proposal, resulting in significantly faster inference speeds. Through a hierarchical feature pyramid and cross-scale feature fusion mechanism, it balances shallow high-resolution features (beneficial for small object localization) with deep semantic features (beneficial for classification robustness), overcoming the missed detection problem caused by the sparse pixel density of small objects in traditional algorithms. Furthermore, its lightweight design, employing depthwise separable convolution and dynamic convolution, makes it more suitable for real-time applications.

[0067] Therefore, this embodiment selects the latest YOLO11-obb as the baseline algorithm, and its structure is as follows: Figure 2 As shown, to address the issues mentioned above such as feature discontinuity, feature misalignment, and inconsistency between metric and loss, improvement measures are proposed. The improved model structure diagram is FEMS-YOLO (Feature Enhancement and Multi-Scale Sharing, FEMS), as shown below. Figure 3 As shown.

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

[0069] The backbone network includes the kite convolution, C3k2_iAFF module, and C2PSA_CGA module, which are used to perform preliminary feature extraction on the input image, providing a foundation for subsequent feature fusion and detection.

[0070] The neck network, consisting of kite convolution and C3k2_iAFF modules, is located between the backbone network and the head network. It is used to fuse features extracted from different levels by the backbone network to generate feature representations.

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

[0072] The kite convolution, C3k2_iAFF module, and C2PSA_CGA module enhance the feature representation and recognizability of small targets by using asymmetric image filling feature extraction operations, iteratively using attention mechanisms, and cascading integration of edge information.

[0073] The multi-scale shared convolutional detection head uses a multi-scale feature fusion and sharing enhancement strategy to construct 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 methods, which typically use symmetrical forms to process directional features in images, have inherent limitations. They fail to fully extract edge features from small remote sensing targets, and their isotropic receptive fields struggle to capture structured information such as texture direction and edge distribution. Furthermore, as network depth increases, the surge in parameters leads to a decrease in computational efficiency.

[0077] To address this issue, this embodiment introduces PConv, which provides a solution for detection that combines accuracy and efficiency through an 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, achieving accurate analysis of complex directional features in remote sensing images through differential filling and kernel optimization. The Gaussian distribution-driven center feature focusing strategy enhances the feature representation of the core region of small targets through dynamic weight allocation, significantly improving the feature misalignment problem caused by few target pixels and background interference. The PConv structure and principle are shown in the figure below. Figure 4 As shown.

[0078] The first layer of PConv employs four parallel convolutional branches, achieving multi-directional feature capture through differentiated padding parameters and kernel size. Each branch has different padding parameters and convolutional kernels to extract features in different directions, and its calculation formula is as follows:

[0079]

[0080] in, For convolution, Wi is the corresponding convolution kernel, and the four parameters in P represent the number of fill pixels in the left, right, top, and bottom directions, respectively.

[0081] After the first convolutional layer, the relationship between the height h′, width w′, and number of channels c′ of the output feature map and the height h1, width w1, and number of channels c1 of the input feature map is calculated using the following formula:

[0082]

[0083] Where c2 is the number of channels in the final output feature map of the kite convolution, and s is the convolution stride;

[0084] Finally, the concatenated tensor is normalized using a W(2,2,c2) convolution kernel, and the height and width of the output feature map are adjusted to preset values ​​h2 and w2, respectively. The final output Y(h2,w2,c2) is calculated 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 weights in the central region are significantly higher than those in the periphery. This allows the central features to receive more attention, which aligns with the characteristic of concentrated features in small objects. In tasks such as small object detection, it can better focus on key features, improving feature extraction accuracy while achieving lightweight design, making it particularly suitable for detail-sensitive tasks such as small object detection.

[0087] 1.2 Iterative Attention Feature Fusion

[0088] In target detection tasks driven by deep convolutional neural networks, feature fusion, as the core link in integrating multi-scale and multi-semantic level information, directly determines the model's representation ability and detection performance. However, traditional feature fusion methods (such as simple summation and channel concatenation) often ignore the semantic differences and scale mismatches between features at different levels. Especially in remote sensing small target detection, the difficulty of extracting edge features, large target size variations, and small target pixel counts further exacerbate the difficulty of feature fusion, leading to feature conflicts or information loss during the information fusion process.

[0089] To address the aforementioned issues, this embodiment proposes fusing iterative attentional feature fusion (iAFF) with the C3k2 module, i.e. Figure 2 The internal structure of C3k2-iAFF in the image is as follows: Figure 5 As shown in (c), a multi-level attention mechanism is used to achieve dynamic feature selection and adaptive fusion, dynamically calibrating the semantic relationships and spatial distribution differences between features, thus refining the feature fusion process and enhancing the representation of features of small targets. The Iterative Attention Feature Fusion Module (iAFF) iteratively uses a multi-scale channel attention module (MS-CAM), such as... Figure 5 As shown in (a), dynamic weights are generated for the features of the initial interaction, and the output weights are used to weight the inputs X and Y element-wise to strengthen key feature channels and suppress redundant information. Then, the features after the initial attention calibration are processed again. Optimization is achieved through iteration, gradually refining the feature fusion weights to avoid information bias in the initial fusion. Each iteration dynamically adjusts the fusion ratio of X and Y based on the previous results, ultimately outputting the fused features. This embodiment segments the input features based on this, and performs deep extraction on the segmented single-path features to optimize feature fusion, constructing a Bottleneck_iAFF unit, the structure of which is shown in the figure below. Figure 5As shown in (b), the C3k2-iAFF module architecture integrates the Bottleneck_iAFF unit system to form a highly efficient feature processing pipeline. The architecture first preprocesses the input features through convolution, then distributes the features to multiple branches via a Split operation. Each branch independently connects to the Bottleneck_iAFF unit, enabling parallel iterative optimization of multi-scale features: deeper branches utilize the Bottleneck_iAFF unit to enhance global semantic consistency, while shallower branches retain detailed features with the help of the Bottleneck_iAFF unit. Finally, the multi-branch features processed by the Bottleneck_iAFF unit are aggregated through Concat and then unified through convolution.

[0090] 1.3 Cascaded Groups Enhance Attention

[0091] The C2PSA module is a core component in YOLOv11 used to enhance feature extraction. Its design integrates the CSP (CrossStage Partial) structure and the Pyramid Squeeze Attention (PSA) mechanism. It introduces multi-scale convolutional kernels to process input features in parallel, with each branch capturing spatial information from different receptive fields. However, when dealing with small targets, feature extraction is easily affected by background interference, resulting in insufficient capture of detailed information and inadequate global semantic aggregation efficiency, thus failing to fully leverage the complementarity of multi-scale features.

[0092] To address this issue, this embodiment adopts a strategy that combines Cascaded Group Attention (CGA) with a C2PSA module to construct a C2PSA_CGA cascaded group enhanced attention module (i.e., the C2PSA_CGA module), as follows: Figure 6 As shown in (a), the main idea of ​​the CGA unit is to enhance the diversity of features input to the attention heads. Unlike previous self-attention methods, it provides different input splits for each head and cascades output features across heads. Specifically, the CGA unit divides the input features into different parts, each part being input to an attention head. Each head computes its self-attention map, and then the outputs of all heads are cascaded together. In this way, the output of each head is added to the input of the next head, thereby progressively 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 computational redundancy in multi-head attention but also increases model capacity by increasing network depth.

[0093] A new feature fusion processing framework was constructed by integrating the C2PSA module and the CGA unit, such as... Figure 6As shown in (b), the input features are first processed in parallel using C2PSA multi-scale convolution to capture spatial information from different receptive fields. Then, a CGA unit is introduced to segment the features into attention heads. Through independent self-attention computation and cross-head cascading operations, the feature representation is gradually refined. Finally, cross-layer feature fusion is achieved through convolution. Compared with C2PSA, the C2PSA_CGA module avoids the problem of repeated extraction at the feature processing level by leveraging the input segmentation strategy of the CGA unit. At the computational efficiency level, the CGA unit reduces QKV channel redundancy. At the model depth level, the learning capacity is expanded through cross-head cascading of the CGA unit. This can significantly improve feature diversity, balance computational efficiency and model expressive power, and innovate the cascaded attention mechanism to strengthen the focus of key information.

[0094] The C2PSA_CGA module encourages each attention point to focus on the feature details of different regions, gradually integrating edge information during the cascading process. This significantly improves 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 object detection tasks, small objects, due to their low pixel ratio and insufficient feature resolution, struggle to retain sufficient edge and texture details. Coupled with a lack of semantic information, this easily leads to problems such as feature discontinuity and misalignment. Traditional detection heads, limited by single or finite-scale feature processing modes, cannot fully utilize low-level high-resolution details or efficiently integrate high-level semantic features, resulting in poor detection performance. To address this issue, this embodiment designs a multi-scale shared convolutional detection head (MS-SCD). Its main concept involves adding a small object detection layer (P2) and a shared enhancement strategy (SCD-Strategy), such as... Figure 7 As shown.

[0097] 2.1 Multi-scale sharing strategy

[0098] Multi-scale sharing constructs 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 needs of different scales; and the high-level detection layer focuses on the strong semantic features of medium and large targets. This multi-layered architecture, from bottom-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 convolutional detection strategy is as follows: The multi-scale input first passes through the detail enhancement convolution (DEConv), then through the efficient channel attention (ECA) mechanism to enhance the response of the feature channels, and finally through the detail enhancement convolution again to form an extraction-optimization-re-extraction link. Multi-dimensional feature fusion and reparameterization operations gradually enhance the expressive power of the features, resulting in targeted processing of directional detection, bounding box regression, and classification regression outputs, thereby improving the detection accuracy of multi-scale targets in complex scenes.

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

[0101] Furthermore, the design that shares DEConv and ECA avoids parameter redundancy, maintains efficient model operation in deep feature processing, and achieves a balance between detection accuracy and computational cost.

[0102] 2.2 Enhanced Detail Convolution (DEConv)

[0103] In object detection tasks, small objects are difficult to detect because high-frequency details such as edges and contours are easily lost. Traditional convolution is limited by a single feature extraction mode and cannot effectively capture multi-dimensional detail information. Therefore, this embodiment introduces a detail-enhancing convolution (DEConv) in the shared strategy module, such as... Figure 8 As shown, a new solution is provided to address this problem.

[0104] Traditional convolutions have limited ability to capture high-frequency details (such as edges and contours) in feature extraction, while high-frequency information (such as edges and contours) is crucial for image quality restoration and enhancement. Inspired by this, DEConv integrates vanilla convolution (VC) with various differential convolutions such as central difference convolution (CDC), angular difference convolution (ADC), horizontal difference convolution (HDC), and vertical difference convolution (VDC) to construct a multi-dimensional feature extraction framework, compensating for the shortcomings of single convolutions. This multi-dimensional feature extraction mechanism directly compensates for the lack of detail representation in object detection, especially for small objects, providing richer feature inputs for accurate object identification and significantly improving the accuracy of small object detection.

[0105] To address the parameter redundancy issue caused by multiple parallel convolutions, DEConv employs a reparameterization technique to effectively merge parallel convolution kernels into a single standard convolution kernel. The input-output relationship is shown in the following formula:

[0106]

[0107] DEConv employs a reparameterized structure, enabling the detection head to fully extract target details through multiple branches while maintaining efficient model operation through reparameterization when processing multi-scale features. This ultimately improves the detection accuracy and generalization ability 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 in this embodiment. For the multi-scale shared detection head studied in this embodiment, DEConv's reparameterized characteristics perfectly meet the lightweight requirements, significantly reducing the number of model parameters and improving detection efficiency while ensuring the extraction of detailed features of multi-scale targets (especially small targets).

[0108] Example 2

[0109] This embodiment discloses a method for constructing a directed detection model for small remote sensing targets, including:

[0110] Step 1: Construct a remote sensing small target directed detection model as 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] Where α is the dynamic adjustment factor, λ is the dynamic weight coefficient, Angle_Cost is the in-angle cost, Distance_Cost is the distance cost, Shape_Cost is the shape cost, and Loss is the loss value.

[0116] By fusing SIoU, which is sensitive to the rotation box angle and contains multidimensional constraints, with Inner-IoU, which has a dynamic scaling mechanism, we obtain Inner-SIoU, a loss function that includes an angle-aware weight adjustment module and dynamically balances angle errors, in order to enhance the ability to cope with angle constraints and scale changes in directional target detection.

[0117] To address the challenges of small remote sensing targets often being embedded in complex backgrounds and having scarce pixels, traditional IoU (Intersection over Union) is susceptible to background interference, leading to positioning errors and making it difficult to simultaneously meet the requirements of high-precision positioning and orientation determination. Furthermore, in directional target detection scenarios, while introducing angle parameters can more accurately describe the target orientation, it also introduces inconsistencies between the metric and the loss function. CIoU in YOLO11-obb is developed based on the IoU (Intersection over Union) loss function. However, it still has some limitations; for example, when two bounding boxes do not intersect, the IoU is 0, failing to reflect the distance information between the boxes.

[0118] To address the aforementioned issues, this application proposes the Inner_SIoU loss function, which combines the advantages of both methods to solve existing problems: on the one hand, the scaling strategy of Inner_IoU enhances the anti-interference capability of small target overlap calculation; on the other hand, combined with the angle penalty mechanism of SIoU, it accurately constrains target orientation deviation, while utilizing its distance and shape costs to optimize bounding box spatial alignment. This fusion scheme not only solves the problem of inconsistency between metric and loss caused by angle parameters in directed detection, but also achieves collaborative optimization of anti-interference localization and orientation discrimination, providing a more robust solution for remote sensing directed small target detection, helping the model to converge faster and promoting multi-dimensional constraint alignment.

[0119] 1. Inner_IoU loss function

[0120] Traditional IoU measures localization accuracy by calculating the intersection-union ratio (IoU) between predicted and ground truth bounding boxes. However, its sensitivity to background interference leads to significant calculation errors in overlapping areas of small targets. Because small targets are sparsely pixelated and often embedded in complex backgrounds, traditional IoU is susceptible to background noise, resulting in artificially high IoU values ​​and consequently regression errors. To address this issue, Inner_IoU proposes adjusting the bounding box size using a scaling factor (ratio) before calculating the scaled IoU. This approach reduces background interference, especially in small target detection, allowing for a better focus on the target itself. The calculated IoU value more accurately reflects the actual overlap of targets, providing a more accurate basis for small target localization. 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) to adaptively adjust the bounding box size, the core target region is focused on, and background interference is suppressed. The core formula is shown below:

[0126]

[0127] Where (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 effectively suppresses the interference of complex backgrounds in remote sensing small target detection by introducing a scaling factor to adaptively adjust the bounding box size, focusing the bounding box on the core area of ​​the target, making the intersection-union ratio calculation closer to the actual overlap of the target.

[0129] 2. SIoU Loss Function

[0130] Traditional IoU measures localization accuracy using the intersection-union ratio (IoU), but it only focuses on geometric overlap while ignoring differences in direction, distance, and shape, limiting its performance in detecting small, oriented remote sensing targets. Therefore, SIoU introduces angle cost (Angle_Cost), distance cost (Distance_Cost), and shape cost (Shape_Cost) to build multi-dimensional geometric constraints, addressing the issues of orientation sensitivity and spatial alignment in target detection. Angle cost (Angle_Cost), as the core component of the SIoU loss function, uses precise mapping of geometric transformations and trigonometric functions to quantify the directional deviation between the predicted and ground truth bounding boxes. It then uses cosine similarity to quantify the angular deviation between the predicted and ground truth bounding boxes, constructing an orientation-discriminative loss. Its calculation formula is as follows:

[0131]

[0132] Where θ p and θ g These are the rotation angles of the predicted bounding box and the ground truth bounding 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 improve the regression accuracy of the rotation angle during training.

[0133] Distance cost eliminates the interference of rotation on the distance measurement of the center point through polar coordinate transformation. Traditional Euclidean distance is easily affected by attitude in directional target detection, leading to positioning errors. The normalized distance cost is defined by converting the center point coordinates to polar coordinates, and its calculation formula is as follows:

[0134]

[0135] Where c w and c h This refers to the width and height of the minimum bounding box between the predicted bounding box and the ground truth bounding box. This strategy normalizes the center point distance to the target scale range, avoiding distance metric failure caused by rotation.

[0136] Shape Cost (Shape_Cost) is a dynamically corrected module designed to address aspect ratio distortion caused by rotation. Traditional IoU is insensitive to shape changes in the rotated target, leading to inaccurate aspect ratio estimation. An exponential function is used to strengthen the penalty for shape discrepancies; its 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 nonlinearly amplifies shape differences, significantly improving the model's robustness to shape distortion. The final SIoU loss calculation formula is shown below:

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

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

[0141] Using Inner_IoU as the core, its bounding box scaling strategy reduces background interference in remote sensing small target detection, ensuring that overlap calculation focuses on the target subject and improving localization accuracy. Angle, distance, and shape costs of SIoU are introduced through dynamic weighting coefficients. This allows the model to not only focus on the overlap of the core target region during optimization but also to enforce constraints on the spatial distance deviation and shape differences between the predicted and ground truth bounding boxes, meeting the multi-dimensional optimization needs of remote sensing oriented small 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, leading to measurement bias. This embodiment adopts an Inner-IoU scaling bounding box strategy, which suppresses background noise by dynamically adjusting factors to achieve accurate focusing on the core region of the small target:

[0143]

[0144] Here, δ and β are adaptive parameters. Through the nonlinear characteristics of the exponential function, α approaches 0 as the IoU value increases. When the predicted bounding box and the ground truth bounding box highly overlap, α approaches 0, and the traditional IoU calculation is used directly. When the overlap is low (such as in small object detection scenarios), α approaches 1, and background interference is reduced by scaling the bounding box.

[0145] 2) Traditional IoU loss functions often lead to optimization bias when dealing with rotating targets 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 and the loss function. The dynamic weight coefficient λ is defined by the following formula:

[0146]

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

[0148] 3) Combining Inner-IoU and 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 bounding box and the ground truth bounding box are highly matched in terms of anti-interference overlap, distance, and shape (Inner-SIoU approaches 1), the loss value approaches 0. The model training objective is to minimize this loss, forming a clear optimization direction. Through a fusion strategy, Inner-SIoU retains the anti-interference ability of Inner-IoU for small targets, and effectively solves the inconsistency between metric and loss caused by angle parameters in directed detection through the multi-dimensional cost constraint of SIoU. This achieves consistent alignment between metric and loss, providing a more accurate optimization criterion for target detection in complex remote sensing scenarios.

[0152] Example 3

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

[0154] The following experiments were conducted on this invention, and the results and analysis are as follows:

[0155] 1. Experimental Environment and Parameter Configuration

[0156] This invention uses Python as the development tool and the PyTorch framework for developing and training deep learning models. The specific configuration of the experimental environment is shown in Table 1, and the parameter settings used in the experiment are detailed in Table 2. The DIOR-R dataset and the UCAS-AOD dataset were selected for the experiment to evaluate the model's performance.

[0157] Table 1 Specific configuration of the experimental environment

[0158]

[0159] Table 2. Parameter settings used in the experiment

[0160]

[0161]

[0162] 2. Introduction to the dataset

[0163] This invention uses the DIOR-R and UCAS-AOD datasets, equipped with rotated target annotations, to test and evaluate the proposed method. The DIOR-R dataset, a derivative of the DIOR dataset, encompasses 23,463 images and 190,288 annotation instances, broadly covering typical remote sensing environments and 20 everyday 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 object detection, although UCAS-AOD has fewer target categories, its high-resolution images and refined rotated annotations provide effective validation of the robustness of directed object detection algorithms for small targets.

[0164] 3. Evaluation Criteria

[0165] The experiment evaluated several key metrics, including precision (P), recall (R), average precision (AP), mean average precision (mAP), the core metric for complexity calculation, GFLOPs, and the number of model parameters (Params). The aim was to comprehensively compare and analyze the performance of object detection models. Precision is the proportion of correctly identified positive instances out of the total predicted positive instances. Recall measures the proportion of correctly identified positive instances out of the total number of actual positive instances. Parameters are the sum of parameters across all network layers during model training. The formulas for calculating these metrics are shown below:

[0166]

[0167] Wherein: 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 for P, R, AP, and mAP is "%".

[0168] 4. Ablation test

[0169] To verify the optimization effect of each improved module in this invention and to explore the effectiveness of the method, ablation experiments were conducted based on YOLO11-obb. The ablation experiment results 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 SCD sharing an enhancement strategy. "√" indicates the use of the improved method, and "-" indicates the absence of the improved method. The experimental results are shown in Table 3. Presenting the ablation experiment results in a bar chart provides a more intuitive view of the improvement effect of the proposed method. Figure 9 As shown.

[0170] Table 3 Ablation Experiment Results

[0171]

[0172] Experimental results show that when using PConv and C3k2-iAFF modules, the model's P-value reaches 87.3%, and mAP50 improves to 85.5%, verifying the optimization effect of the modules on feature enhancement and fusion. After further stacking C2PSA-CGA and MS-SCD (P2+SCD) modules, the R-value increases to 82.4%, mAP50 reaches 86.5%, and the number of parameters decreases to 2.30M, achieving a balance between detection accuracy and model parameters, verifying the role of the modules in enhancing and representing small target features and lightweighting. When all modules (PConv, C3k2-iAFF, C2PSA-CGA, MS-SCD, Inner-SIoU) are used in synergy, the P-value reaches 86.6%, the R-value reaches 82.7%, and mAP50 increases to 86.8%, achieving a significant improvement in detection accuracy.

[0173] 5. Comparative Experiment

[0174] To verify the performance of the FEMS-YOLO algorithm on directed detection datasets, comparative experiments were conducted on the DIOR-R dataset and UCAS-AOD dataset. The results are shown in Tables 4 and 5. Based on the data in the tables, it is evident that FEMS-YOLO exhibits a significant advantage 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 the DIOR-R dataset

[0176]

[0177] Converting the comparative experimental data in Table 4 into a heatmap makes the advantages of the improved FEMS-YOLO model more intuitive. The DIOR-R comparative experimental heatmap is shown below. Figure 10 As shown.

[0178] Table 5 Comparative Experimental Results of UCAS-AOD

[0179]

[0180] Converting the comparative experimental data in Table 5 into a bar chart makes the advantages of the improved FEMS-YOLO model more intuitive. The bar chart for the UCAS-AOD dataset comparison experiment is shown below. Figure 11 As shown.

[0181] Experimental results on the DIOR-R and UCAS-AOD datasets demonstrate that FEMS-YOLO exhibits significant advantages in directed target detection tasks. On the DIOR-R dataset (containing 20 classes of complex remote sensing targets), FEMS-YOLO achieves an mAP50 of 86.8%, a 2.1% improvement over the baseline model YOLO11-obb (84.7%), and outperforms other mainstream detection models. Its average accuracy ranks first in 17 of the 20 target classes. On the UCAS-AOD dataset, FEMS-YOLO achieves an mAP50 of 97.2%, surpassing the baseline YOLO11-obb (96.9%). Its AP values ​​for the two target classes are 95.6% and 98.9%, respectively, further demonstrating the model's robustness in specific domains. Compared to mainstream single-stage models (such as RetinaNet-O with 82.1%) and two-stage models (such as RoI Transformer with 88.9%), the FEMS-YOLO proposed in this embodiment exhibits significantly superior detection performance. Comparative experiments on two datasets validate the effectiveness of the proposed FEMS-YOLO model, which utilizes feature enhancement and multi-scale sharing strategies to achieve higher detection stability within a single-stage framework. This fully demonstrates the superiority of the proposed improvements in directed detection tasks, achieving a balance between detection accuracy and parameter quantity.

[0182] 6. Visualized Results Analysis

[0183] Visual analysis can more intuitively demonstrate the detection performance of the model, such as... Figure 12-15 The figures show the visualization results of YOLO11-obb and FEMS-YOLO on the DIOR-R dataset, as well as the results of the Precision-Recall curve comparison.

[0184] Figure 12 This image presents a visual comparison of the improved YOLO11-obb detection results for images with dense small targets. (a) and (c) show the results from YOLO11-obb (objects circled in red are undetected), while (b) and (d) show the results from FEMS-YOLO for the same image. Compared to the base model, FEMS-YOLO successfully identified and located all targets in the image without any incorrect bounding boxes.

[0185] Figure 13The image presents a visual comparison of the results before and after the improvement, considering the significant scale changes of different categories of small targets in remote sensing images. (a) shows the detection results of YOLO11-obb, and (b) shows the detection results of FEMS-YOLO for the same image. Because the target bounding boxes of ships and ports undergo large scale changes, leading to missed detections, the improved model was not affected by scale changes, demonstrating the effectiveness of the model in the field of directed detection of small targets.

[0186] like Figure 14 As shown, the model's improvement rate slows down during the mid-to-late training stages, gradually converging. However, FEMS-YOLO enters a gradual upward trend earlier and outperforms YOLO11-obb in convergence speed, demonstrating the effectiveness of the proposed loss function improvement in this embodiment. Ultimately, FEMS-YOLO's mAP50 is higher than YOLO11-obb, verifying the effectiveness of the feature enhancement improvement in this embodiment.

[0187] Analysis based on the PR curve comparison chart: Figure 15 (a) shows that the mAP50 for all categories of YOLO11-obb is 84.7%, while FEMS-YOLO improves to 86.8%. Furthermore, Figure 15 In (b), the blue curve representing the overall performance of FEMS-YOLO across all categories is higher than the curve of YOLO-obb, indicating that FEMS-YOLO has better precision at the same recall rate. For example, when the recall rate is close to 0.8, it still maintains a high precision rate, while the precision rate of YOLO-obb has dropped significantly, demonstrating that FEMS-YOLO has better control over false positives. In the UCAS-AOD dataset, the mAP50 of FEMS-YOLO 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 issues of feature discontinuity, feature misalignment, and inconsistency between measurement and loss caused by directional detection angle parameters in remote sensing small target detection. It proposes FEMS-YOLO, a directional detection method for small remote sensing targets based on feature enhancement and multi-scale sharing. This method leverages asymmetric image filling operations using kite-enhanced convolution and iteratively employs an attention-based feature fusion mechanism to construct a refined feature fusion module, enhancing the extraction of small target edge features and detail perception. A cascaded group attention mechanism is used to fully integrate small target edge information features, enhancing feature diversity. A multi-scale shared convolutional detection head is designed, achieving enhanced fusion of high- and low-level feature information and effective integration of semantic features through a multi-layered feature fusion architecture and detail-enhanced convolution. Finally, SIoU, which is sensitive to the rotation box angle and contains multi-dimensional constraints, is fused with Inner-IoU, which has a dynamic scaling mechanism, to obtain a new loss function, Inner-SIoU, which includes an angle-aware weight adjustment module and dynamically balances angle error and penalty coefficient, thereby strengthening the joint optimization of angle constraints and scale changes in directional target detection. Experimental results show that the improved model achieves 86.6% precision, 82.7% recall, and 86.8% mAP50 on the DIOR-R dataset, which are improvements of 0.7%, 2.7%, and 2.1% respectively compared to the baseline model YOLO11-obb. At the same time, the number of parameters is reduced by 13%. It outperforms other mainstream detection models in terms of average precision and achieves an mAP50 of 97.2% on the UCAS-AOD dataset, verifying the effectiveness of the method proposed in this application.

[0189] Of course, the above description is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the embodiments of the present invention. The present invention is also not limited to the above examples, and all equivalent changes and improvements made by those skilled in the art within the scope of the present invention should fall within the patent coverage of the present invention.

Claims

1. A directional detection model for small remote sensing targets, characterized in that, include: The backbone network includes the kite convolution, C3k2_iAFF module, and C2PSA_CGA module, which are used to perform preliminary feature extraction on the input image, providing a foundation for subsequent feature fusion and detection. The neck network, consisting of kite convolution and C3k2_iAFF modules, is located between the backbone network and the head network. It is used to fuse features extracted from different levels by the backbone network to generate feature representations. Head network: Includes a multi-scale shared convolutional detection head, which is used for multi-scale shared feature fusion; The kite convolution, C3k2_iAFF module, and C2PSA_CGA module enhance the feature representation and recognizability of small targets by using asymmetric image filling feature extraction operations, iteratively using attention mechanisms, and cascading integration of edge information. The multi-scale shared convolutional detection head uses a multi-scale feature fusion and sharing enhancement strategy to construct a multi-scale shared detail enhancement structure to enhance the ability to perceive details of small targets; The C3k2_iAFF module first preprocesses the input features through convolution, and then distributes the features to multiple branches through a Split operation. Each branch is independently connected to the Bottleneck_iAFF unit to achieve parallel iterative optimization of multi-scale features: deep branches use the Bottleneck_iAFF unit to enhance 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. The Bottleneck_iAFF unit iteratively uses multi-scale channel attention subunits to dynamically generate weights for the features of the initial interaction, 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; optimization is carried out through iteration, and the feature fusion weights are gradually refined to avoid information bias in the initial fusion; each iteration is based on the previous results, and the fusion ratio of X and Y is dynamically adjusted to finally output the fused features; 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 by the C2PSA module through multi-scale convolution parallel processing to capture spatial information of different receptive fields. Then, the CGA unit is introduced to segment the features into each attention head. Through independent self-attention calculation and cross-head cascade operation, the feature representation is gradually refined. Finally, cross-layer feature fusion is achieved through convolution. The CGA unit divides the input features into different parts, and each part is input into an attention head; Each head computes its self-attention map, and then the outputs of all heads are concatenated. In this way, the output of each head is added to the input of the next head, thereby progressively refining the feature representation. Finally, they are projected back to the input dimension through a linear layer.

2. The directed detection model for small remote sensing targets according to claim 1, characterized in that, The first layer of the kite convolution employs four parallel convolution branches, achieving multi-directional feature capture through differentiated padding parameters and kernel size. Each branch has different padding parameters and convolution kernels to extract features in different directions, and its calculation formula is as follows: in, For convolution, Wi is the corresponding convolution kernel, and the four parameters in P represent the number of fill pixels in the left, right, top, and bottom directions, respectively. After the first convolutional layer, the relationship between the height h′, width w′, and number of channels c′ of the output feature map and the height h1, width w1, and number of channels c1 of the input feature map is calculated using the following formula: Where c2 is the number of channels in the final output feature map of the kite convolution, and s is the convolution stride; The concatenated tensor is normalized using a W(2,2,c2) convolution kernel. The height and width of the output feature map are adjusted to preset values ​​h2 and w2, respectively. The final output is Y(h2,w2,c2), calculated as follows: 。 3. The directed detection model for small remote sensing targets according to claim 1, characterized in that, The multi-scale shared convolutional detection head adds a small target detection layer and a shared enhancement strategy unit to construct a multi-layered feature perception system.

4. The directed detection model for small remote sensing targets according to claim 3, characterized in that, The bottom layer of the small target detection layer focuses on capturing high-resolution details of small targets, preserving key information such as edges and textures; the middle layer gradually integrates semantic and detail features to balance the expression needs of different scales; and 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 passes through detail enhancement convolution, then through an efficient channel attention mechanism to enhance the response of the feature channels, and finally through detail enhancement convolution again to form an extraction-optimization-re-extraction link. Multi-dimensional feature fusion and reparameterization operations gradually enhance the expressive power of the features, resulting in targeted processing of directional detection, bounding box regression, and classification regression outputs, thereby improving the detection accuracy of multi-scale targets in complex scenes.

5. A method for constructing a directed detection model for small remote sensing targets, characterized in that, include: Step 1: Construct a remote sensing small target directed detection model as described in any one of claims 1-4; 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: in, As a dynamic adjustment factor, , where is the dynamic weighting coefficient, Angle_Cost is the in-angle cost, Distance_Cost is the distance cost, Shape_Cost is the shape cost, and Loss is the loss value.

6. A directional detection method for small remote sensing targets, characterized in that, The remote sensing small target directed detection model constructed using the method described in claim 5 is used for detection.

Images