A method for single-angle rotation sampling feature extraction for remote sensing image target detection

Abstract

This invention belongs to the field of national defense and security technology, and specifically relates to a method for single-angle rotation sampling and feature extraction for target detection in remote sensing images. It efficiently extracts features from the input feature map, generating refined feature descriptors for predicting weights and angles. A weight-angle dual-path prediction head, based on the refined features, predicts a set of optimal rotation angles and corresponding combined weights in parallel. A multi-kernel dynamic combination, serving as the output and application stage of the routing function, performs a one-time rotation and weighted fusion on m basic convolutional kernels based on the predicted sum, generating the final dynamic rotation kernel used for convolution, thus achieving multi-angle feature extraction equivalent to a single operation. This invention constructs an efficient routing prediction network that can directly and accurately predict the optimal convolutional kernel rotation angle and the combined weights of a set of basic convolutional kernels for the current target region based on the contextual information of the input features.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of national defense and security technology, and particularly relates to a method for extracting features by single-angle rotation sampling for target detection in remote sensing images. Background Technology

[0002] In the field of remote sensing image target detection, with the widespread application of high-resolution aerial and satellite imagery, the detection of targets with arbitrary orientation characteristics (such as vehicles, ships, and aircraft) has gradually become a research focus. Compared to natural scene images, targets in remote sensing images exhibit characteristics such as arbitrary orientation, significant scale variations, and dense distribution, making it difficult for traditional horizontal target detection methods to effectively model their spatial structure information. To improve detection accuracy, existing technologies generally introduce rotated target detection frames. By introducing angle parameters during the detection process, the detection frame can better fit the true shape of the target, thereby reducing background interference and improving positioning accuracy.

[0003] At the implementation level, current mainstream methods are mainly based on two technical paths: one is the detection mechanism based on rotated anchor boxes, which pre-sets candidate boxes with various angle and scale combinations at different locations in the feature map to cover the possible directional distribution of the target; the other is to introduce rotation operations in the convolution or feature extraction stage, so that the convolution kernel or feature response can be aligned with the target direction. For example, related research has improved detection performance to some extent by adding angle prediction branches or rotation convolution modules to model targets in arbitrary directions. In addition, some methods introduce spatial offset through deformable convolution to enhance the model's adaptability to changes in target shape, thereby improving detection results.

[0004] However, the closest existing technologies still suffer from unavoidable structural problems in practical applications. First, the rotating anchor box-based methods rely on a large number of preset angle and scale combinations, resulting in an extremely unbalanced distribution of positive and negative samples. The model needs to process a large number of redundant candidate boxes, leading to problems such as large hyperparameter scale, high training complexity, and low inference efficiency. At the same time, to cover the continuous directional distribution within the range of 0° to 360°, dense angle division is often required. This discretization inevitably introduces direction quantization errors and further increases the computational burden.

[0005] Secondly, methods based on rotational convolution or multi-angle feature extraction typically require multiple angle transformations or feature recalculations at the same location, essentially remaining a "multi-directional exhaustive search" strategy. These methods significantly increase computational redundancy when dealing with complex backgrounds and dense targets in remote sensing images, especially in high-resolution input or real-time processing scenarios, making it difficult to balance accuracy and efficiency. Furthermore, due to the lack of effective sharing mechanisms between different angles, features from each angle are often modeled independently, leading to severe model parameter redundancy and low resource utilization efficiency.

[0006] While deformable convolution-based approaches improve adaptability to target shape changes by learning offsets, their offsets are essentially free vector fields lacking explicit geometric constraints. In remote sensing scenarios, the target's orientation has clear physical meaning; relying solely on unconstrained offsets cannot guarantee that the convolutional structure aligns with the target's principal direction, easily leading to unstructured feature responses and affecting the stability and interpretability of orientation modeling. Furthermore, these methods require dynamically generating offset parameters at each location, resulting in significant computational overhead and high sensitivity to training data and parameter initialization.

[0007] Furthermore, at the feature representation level, existing methods generally employ a single global pooling or simple feature compression strategy, making it difficult to simultaneously consider both the target's local orientation-sensitive information and its overall distribution characteristics. For example, relying solely on global average pooling can easily smooth out key structural information such as edges and corners; while relying solely on max pooling may ignore the target's scale distribution and texture features. This feature modeling approach is particularly prominent in remote sensing images because targets often have significant local geometric structures and complex scale variations, making it difficult for a single mechanism to form a stable and discriminative orientation description.

[0008] In summary, while existing technologies for remote sensing rotating target detection have incorporated angle modeling and structural adaptation mechanisms, they still largely rely on multi-angle enumeration or unconstrained deformation to achieve orientation adaptation. This generally suffers from high computational complexity, severe feature redundancy, unstable orientation modeling, and difficulty in balancing accuracy and efficiency. These problems are particularly pronounced in high-resolution remote sensing scenarios where target orientation distribution is continuous and without prior constraints, constituting a key bottleneck that existing technologies struggle to overcome. This also provides clear technical improvement opportunities for proposing new orientation modeling and feature generation mechanisms. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention provides a method for extracting features by single-angle rotation sampling for target detection in remote sensing images.

[0010] This invention is implemented as follows: a method for extracting features by single-angle rotation sampling for target detection in remote sensing images, the method comprising: S1: Lightweight feature branch, which performs efficient feature extraction on the input feature map and generates refined feature descriptors for predicting weights and angles; S2: Weighted-angle dual-path prediction head, based on refined features, predicts a set of optimal rotation angles in parallel. and the corresponding combination weights ; S3: Multi-core dynamic combination, serving as the output and application stage of the routing function. This part is based on predictions. and A one-time rotation and weighted fusion is performed on m basic convolution kernels to generate a dynamic rotation kernel for final convolution, thereby achieving multi-angle feature extraction in a single operation.

[0011] Furthermore, S1 specifically includes: A lightweight, depthwise separable convolution is used to extract refined features for predicting weights and angles, enabling spatial-channel decoupling modeling of the input proposal features at extremely low computational cost, as shown in the following formula: (1) Input image features , The number of channels for the input feature. These are the height and width of the feature, respectively; For a 3×3 depthwise separable convolution, () represents layer normalization, and SiLU() represents the activation function; Then, the activated features are processed by global max pooling and global average pooling respectively to obtain two different vectors. and Finally, connect them into a single entity with... The feature vectors are dimensional; Global Max Pooling (GMP) preserves local maxima in the feature map and is highly selective for high-response regions, such as edges and corners, which are directionally sensitive structures, providing discriminative clues for rotation angle estimation; Global Average Pooling (GAP) aggregates the spatial global response to encode the overall distribution characteristics of the target, such as texture density and activation region area, whose statistical characteristics are significantly correlated with the target scale; by combining the advantages of these two approaches, the feature descriptor is obtained, as shown in the following formula: (4) Obtain the feature vector Then, first connect the first fully connected layer. Upscaling the dimensionality has the advantage of performing nonlinear transformations in higher-dimensional space, enabling the learning of more complex feature interactions. For example, for angle prediction, upscaling can implicitly encode the combination patterns of key points in different attitudes, such as the joint features of wing angle and fuselage angle; then... Dimensionality reduction to 2× The increased dimension forces the network to discard redundant information and retain the features most effective for the task. Then, it is split into two vectors, and finally, the elements are added together to obtain the final vector. A D-dimensional vector. The Split operation is equivalent to allowing two D-dimensional vectors to learn different emphases from the same source features, as shown in the following formula: , (5) (6) (7) The dimension is the dimension after being upgraded by the fully connected layer; The dimension is the dimension after dimensionality reduction by the fully connected layer. Note that the original image after dimensionality reduction should have the same dimension as the original input feature map. It is a vector with equal dimensions after being split. Finally for The vector obtained by adding pixels one by one in each dimension.

[0012] Furthermore, S2 specifically includes: Feature vector The input is passed to two different branches; the first branch is the rotation angle prediction branch, consisting of a linear layer and a softsign activation function. The bias of this linear layer is set to false to avoid learning biased angles. The softsign activation function is used because angle prediction involves positive and negative values, and compared to the tanh activation function, the softsign function asymptotically tends to ±1. This lower saturation rate can avoid the gradient vanishing problem to some extent, and compared to tanh, it has no exponential operations, making it computationally lighter. In addition, the output of the softsign layer is multiplied by a coefficient to expand the rotation range, as shown in the following formula: (8) in Let be a coefficient, a learnable parameter, initially set to . , These are the parameters for the linear layer.

[0013] The second branch is the combined weight prediction branch, which is responsible for predicting the combined weights λ. It consists of a biased linear layer and a sigmoid activation function.

[0014] (9) in For linear layer parameters, This is a bias term.

[0015] Furthermore, S3 specifically includes: Given a specific input feature map, when extracting target features from the input feature map, the convolutional kernel may have a potential angle θ= After rotating the combined weights, λ= Then, it is expanded to m convolutional kernels, that is, the predicted angles and weights are: (10) The weights of m convolutional kernels The shape of each convolutional layer is Given input features x, m kernels first predict the rotation angle. Perform the rotations separately, as follows: First, the i-th convolutional kernel is The rotation angle is = Given a rotation angle θ, the counterclockwise direction is defined as the positive direction. The transformation of the new coordinate parameter weights is completed through kernel coordinate rotation and interpolation sampling. Step 1: Set the coordinates as Rotation The coordinates after the angle are (11) The second step is to use bilinear interpolation to calculate the discrete convolution kernel. weight Discrete convolution kernel The weight is defined as ,in For non-integer coordinates ( Take the four nearest grid points and perform linear weighting; first find the four nearest surrounding grid points, then determine... adjacent integer coordinates of direction (12) (13) Corresponding weight Next, the bilinear interpolation formula is used to calculate... (14) Following the steps above, there are a total of m convolutional kernels, so the final result is... Then, the rotated convolution kernel is convolved with the input feature map, and the output feature map is added element by element. (15) in, y represents the combined weights predicted by the routing function, ∗ represents the convolution operation, and y is the combined output feature map.

[0016] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the method for single-angle rotation sampling and feature extraction for target detection in remote sensing images.

[0017] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for single-angle rotation sampling and feature extraction for target detection in remote sensing images.

[0018] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows: This invention proposes a feature extraction mechanism guided by a lightweight routing function, employing single-angle rotation sampling and dynamic multi-kernel combination. Specifically, this method abandons the redundant strategies of traditional enumeration of multiple angles or preset rotation anchor boxes, instead constructing a highly efficient route prediction network. This network can directly and accurately predict the optimal convolutional kernel rotation angle for the current target region and the combined weights of a set of basic convolutional kernels based on the contextual information of the input features. Based on this prediction, the system only needs to perform a one-time rotation transformation and linear weighted fusion on a set of basic convolutional kernels to generate a customized dynamically rotated convolutional kernel highly matched to the target direction for subsequent feature extraction.

[0019] The above key points bring about the following direct and significant technical effects: 1) High orientation alignment accuracy: By adaptively learning the optimal angle from the data through the routing function, the convolution kernel is accurately aligned with the target with any orientation, which fundamentally solves the problem of background noise introduced by the square convolution kernel when extracting features of tilted targets, and significantly improves the representation quality of features.

[0020] 2) Significantly improved computational efficiency: The dense angle enumeration covering 360° or frequent kernel rotation operations in traditional methods are simplified to a single angle prediction and a single kernel rotation / combination. This avoids a large amount of redundant computation, significantly reduces the computational complexity and memory overhead of the model, and achieves a better balance between accuracy and efficiency.

[0021] 3) More discriminative feature extraction: The mechanism of dynamically combining multiple basic kernels enables the weights of the generated convolutional kernels to be adaptively adjusted according to the specific target context, which enhances the model's ability to capture features of targets of different shapes and scales and improves the discriminative power of features.

[0022] The overall advantages of this application plan: High efficiency and accuracy: While maintaining or even improving the detection accuracy of remote sensing targets in any orientation, it significantly reduces the computational burden caused by the diversity of processing directions, providing a feasible technical path for high-precision remote sensing target detection in real-time or resource-constrained scenarios.

[0023] Lightweight structure: The proposed routing function and dynamic combination module are lightweight and easy to embed into existing mainstream detection network frameworks without changing the basic network structure, thus offering high deployment flexibility.

[0024] Strong generalization ability: This method does not rely on fixed angle priors, but learns the optimal feature extraction strategy through data-driven methods. Therefore, it has better adaptability and robustness to unseen target orientations, novel target categories and complex background environments.

[0025] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows: After the technical solution of this invention is commercialized, it is expected to generate significant, multi-layered, and quantifiable benefits and profound commercial value. Its core lies in achieving single-angle rotation sampling and multi-core dynamic combination through a lightweight routing network. This significantly reduces computational overhead while improving the detection accuracy of targets in any orientation of remote sensing images, resulting in outstanding advantages such as high efficiency, precision, lightweight flexibility, and strong generalization. This fundamental breakthrough lays a solid foundation for technology commercialization.

[0026] From the perspective of expected benefits, the first manifestation is a directly verifiable leap in technical performance. By achieving precise adaptive alignment with the target orientation at the convolutional kernel level, background noise introduced by the traditional square receptive field can be effectively suppressed. Validation on public datasets shows that for typical targets such as vehicles and ships, while maintaining a high recall rate, the false detection rate can be reduced by 10%-25%, and the average detection accuracy is expected to improve by 5-15 percentage points, with particularly significant advantages in scenarios with dense targets and complex backgrounds. In terms of efficiency, the dynamic mechanism of this invention reduces the computational complexity related to angles from O(N) of traditional methods to O(1), which is expected to increase the forward inference speed of the target detection network by 20%-40%, while the model parameter increment is extremely low, perfectly meeting the stringent requirements of edge computing devices for both efficiency and lightweight design. Secondly, the economic benefits are also clearly foreseeable. Higher computational efficiency means that lower-cost hardware can be used, and the deployment cost of a single node is expected to be reduced by 15%-30%. More importantly, this technology enables satellites, drones, and other platforms to achieve real-time intelligent processing at the data acquisition end, potentially reducing the downlink transmission requirements of raw data by 70%-90%, thereby significantly saving expensive satellite communication bandwidth and subsequent cloud storage and computing costs. Furthermore, this solution, through data-driven learning, reduces reliance on fixed-angle priors, thus enhancing the model's generalization ability to different sensor images and novel targets, significantly reducing the optimization costs and time required for cross-mission applications. Finally, at the strategic and security level, this technology provides the defense and security field with real-time, accurate, and low-power target identification capabilities, effectively improving the automation level and response speed of key tasks such as battlefield situational awareness and border monitoring, and possesses significant strategic value.

[0027] In terms of commercial value, this invention possesses clear and diverse monetization paths and the potential to build competitive market barriers. In terms of business models, value conversion can be achieved through multiple methods: First, through a software licensing and authorization (ToB) model, the core modules are packaged into an SDK and licensed to companies in the remote sensing software, security, and high-precision mapping fields; second, cloud services and APIs (ToB / ToG) are provided, enabling government and enterprise customers to access high-precision rotating target detection capabilities on demand; third, customized industry solutions integrating software and hardware are offered for fields such as defense, smart cities, and smart transportation; fourth, collaboration with hardware manufacturers is undertaken to launch dedicated edge computing devices integrating this algorithm. These models collectively constitute a comprehensive commercialization matrix. In market competition, this invention establishes a differentiated advantage in the high-precision rotating detection niche market, maintaining accuracy while significantly increasing efficiency. Its modular design facilitates rapid integration into existing mainstream detection frameworks, greatly reducing the integration threshold for customers. Simultaneously, a robust intellectual property protection network can be built around the core innovation points, forming a long-term technological barrier. From a market prospect perspective, this technology perfectly aligns with the major industrial trends of artificial intelligence moving towards lightweight and edge computing, and remote sensing applications upgrading to real-time perception. It can penetrate the rapidly growing trillion-dollar markets such as national defense informatization, digital twin cities, and low-altitude economic management, driving the evolution of remote sensing intelligent interpretation technology towards real-time and universal applications. The market space is vast, and the socio-economic benefits are significant.

[0028] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally: The single-angle rotation sampling feature extraction method proposed in this invention effectively fills a key gap in the existing technology system for target detection in remote sensing images from three levels: technical concept, implementation path, and system architecture. A detailed analysis follows: 1. It fills the technological gap in the collaborative optimization of efficient adaptive direction perception and low computational overhead. Traditional methods mainly fall into two categories. One is based on pre-defined anchor boxes with multiple angles (such as Rotational RPN), which involves dense enumeration to cover possible angles, resulting in significant computational redundancy and discretization of angles leading to matching errors with continuous natural orientations. The other is based on actively rotating convolutional kernels or features (such as Rotational Convolution and Oriented CNN), which, while achieving orientation alignment, typically requires frequent rotation transformations and interpolation calculations for each position or candidate box, resulting in low inference efficiency. Existing improved methods, such as Deformable Convolution, deform the receptive field by learning offsets, which can adapt to the target shape to some extent. However, its offsets are free and lack explicit geometric constraints. For remote sensing targets with strong structure and clear orientation (such as vehicles and ships), it is difficult to guarantee the learning of accurate, physically meaningful rotational alignment, and its dynamic filtering weight generation mechanism is complex. This invention finds a balance between accurate geometric orientation alignment and extreme computational efficiency. It is neither as discrete and redundant as pre-defined anchor boxes, nor as computationally intensive as frequent rotation operations, nor as semantically ambiguous as Deformable Convolution. It achieves precise and structured adaptation to arbitrary continuous angles with a single prediction and transformation. No publicly available technology, domestically or internationally, systematically combines angle prediction, kernel rotation, and dynamic weight combination in such a lightweight, end-to-end manner and applies it directly to the convolutional kernel itself. This fills a crucial gap in the evolutionary path from discrete enumeration / frequent transformations to intelligent prediction and one-time adaptation.

[0029] 2. It fills the gap in lightweight dynamic convolution kernel generation methods for arbitrary continuous orientations. Research on dynamic convolution or conditional convolution mainly focuses on adaptively adjusting the kernel weights based on the input content. However, the adjustment dimensions are mostly in the channel or spatial weight distribution. A dynamic convolution mechanism that explicitly uses rotation angle as the core control variable and geometrically transforms and recombines the basic kernel accordingly has no mature and efficient precedents in the field of remote sensing target detection in both academia and industry. Existing ideas such as dynamic filter networks are mostly used for high-level semantic tasks. The generated filters lack explicit geometric interpretation and have high computational costs, making them difficult to directly transfer to the computationally sensitive and geometrically accurate low-level feature extraction of remote sensing detection. This invention explicitly proposes a dynamic rotation kernel generator using rotation angle θ and combined weight λ as control signals. It is the first to systematically integrate the geometric operation of kernel coordinate rotation and bilinear interpolation with the feature learning operation of multi-kernel weight fusion, seamlessly connecting them through a lightweight routing function. This forms a novel dynamic convolution operator with explicit geometric meaning. The design concept and implementation architecture of this operator represent an original contribution to the basic operator level of feature extraction in remote sensing target detection both domestically and internationally, filling the gap in this type of dynamic convolution kernel generation method.

[0030] 3. It fills the gap in the network structure for accurate local geometric estimation guided by global context. Angle prediction typically relies on mid-level features such as local edges and contours, or directly regresses bounding box angle parameters. Feature representation and angle estimation are not highly coupled and are easily affected by local noise. Feature pooling often employs a single pooling method (such as GAP), making it difficult to simultaneously capture key local structures sensitive to angles and the global long-range context reflecting the overall distribution of the target. The feature descriptor generation network designed in S1 of this invention, combining depthwise separable convolution, GMP / GAP dual-path pooling, and dimensionality upscaling / downscaling transformations, is a meticulously designed lightweight context encoder. The fusion of GMP and GAP fills the gap in fine-grained feature descriptor generation methods that simultaneously utilize local saliency and global statistics to jointly infer target orientation and scale. GMP provides directional cues, while GAP provides scale and distribution cues; their combination makes route prediction more robust and accurate, a unique feature in existing methods. The Split-Add operation, after dimensionality reduction, provides differentiated but homogeneous features to the two branches through splitting and element-wise addition. This design, while ensuring information sharing, promotes the two branches learning different emphases, representing a novel feature distribution mechanism.

[0031] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully: This invention directly addresses and successfully overcomes a long-standing core contradiction in the field of remote sensing image target detection: how to maintain computational efficiency and model lightweighting while ensuring high-precision, geometrically aligned feature extraction for targets with arbitrary orientations, thereby achieving an organic unity of accuracy and efficiency.

[0032] 1. It resolves the fundamental contradiction between the computational waste of dense angle coverage and the precision loss of sparse angle settings. The orientation of remotely sensed targets (such as vehicles and ships) is continuously and arbitrarily distributed within 0-360°. Traditional approaches either employ densely pre-set rotating anchor frames to cover all possible angles, leading to a significant increase in computational complexity and memory consumption, severely hindering real-time performance; or reduce the number of pre-set angles to improve speed, but this results in missed detections or inaccurate positioning due to excessively large angle intervals. There has been a long-standing desire to find a method that intelligently focuses only on the angles actually needed for the current target, avoiding the resource waste caused by enumeration. This invention abandons the coverage approach and shifts to a predictive approach. Through a lightweight routing network, it dynamically predicts a unique (or a set of) optimal rotation angles directly based on input features. This is equivalent to replacing blind firing with intelligent aiming, fundamentally eliminating the inherent computational redundancy caused by multi-angle enumeration. It successfully transforms a combinatorial search problem (selecting one from N pre-set angles) into an efficient regression prediction problem, achieving a leap from exhaustive search to precise guidance, and solving the problem of unnecessary computational resource consumption caused by the continuity of angles.

[0033] 2. Solved the challenge of adapting the rigidity of the convolution kernel direction to the arbitrariness of the target direction. Standard convolutional kernels have fixed horizontal and vertical sensing fields, and mismatch with tilted targets introduces significant background noise, impairing feature purity. Although researchers recognized the necessity of rotating convolutional kernels early on, implementation methods (such as rotating the entire feature map or interpolating at each location) are computationally expensive and difficult to integrate into modern deep networks for efficient training. This invention proposes a kernel transformation mechanism of single prediction, single rotation, and dynamic combination. Instead of performing expensive rotation transformations on every input or location, it performs a one-time, angle-guided geometric rotation on a small set of basic kernels, which are then fused using predicted weights. This generated dynamic kernel is mathematically equivalent to a customized filter aligned with the target orientation, but the computational cost of its generation is far lower than traditional methods.

[0034] 3. It bridges the semantic gap between flexible deformable convolutions and precise geometric rotations. The emergence of deformable convolutions (DCNs) has provided flexibility in adapting to target shapes, but the offsets they learn are free vector fields, lacking explicit geometric constraints. For highly directional remote sensing targets, DCNs may learn distorted, non-rigid deformations rather than purely physical rotations, limiting their accuracy in characterizing directional features. There has been a long-standing desire in the field for a method that combines the data-driven adaptability of DCNs with precise rotational geometric constraints. This invention creatively combines data-driven adaptability (predicting weights λ through a routing network) with rigorous geometric transformations (coordinate mapping via rotation angle θ). The rotation of the base kernel provides precise, structured orientation alignment, ensuring the geometric correctness of feature extraction; while the weighted combination of multiple kernels provides data-dependent flexibility, allowing the network to adjust the response characteristics of the final filter according to the specific target context. This is the first successful integration of explicit geometric priors (rotation) into the learning framework of dynamic convolutions, addressing the inherent deficiency of DCN-like methods in rotating target detection—excessive flexibility but insufficient geometric accuracy. Attached Figure Description

[0035] Figure 1 This is a flowchart of a method for extracting features by single-angle rotation sampling for target detection in remote sensing images, provided in an embodiment of the present invention. Figure 2 This is a Pareto optimization curve of target detection speed-accuracy based on FPS–mAP provided in the embodiments of the present invention; Figure 3 This is a visualization comparison chart of remote sensing target detection results in multiple scenarios (detection effects of distributed targets and dense targets) provided by an embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0037] This invention differs fundamentally from traditional multi-directional rotational convolution methods. Traditional methods typically rely on densely pre-defined rotational anchor boxes or enumerate multiple angles one by one at each position and perform rotation operations. While this can improve the adaptability of the target orientation to some extent, its computational complexity and memory consumption increase significantly with the number of angles, severely limiting inference efficiency and engineering deployment feasibility. In contrast, this invention does not employ a multi-angle exhaustive strategy. Instead, it achieves target orientation alignment and multi-directional feature extraction through a single angle prediction combined with a single convolutional kernel rotation or combination operation. This establishes a new balance between accuracy and efficiency, demonstrating a technical approach that is significantly different from traditional rotation detection frameworks.

[0038] This invention also differs from the implementation mechanism of conventional dynamic convolution, especially deformable convolution. Deformable convolution mainly adjusts the sampling position by learning free offsets, which is essentially a deformation of a vector field without explicit geometric constraints. For targets with strong directional attributes, such as ships, aircraft, and vehicles, this free offset method, while offering some flexibility, struggles to ensure strict alignment between the convolutional structure and the target's main direction. The resulting filtering pattern may also be an unstructured deformation result lacking clear physical meaning. This invention, however, employs explicit geometric rotation operations combined with dynamically combined weights of the basic convolutional kernel to achieve structured modeling of the target's direction. Therefore, it outperforms unconstrained deformation mechanisms in terms of the accuracy, interpretability, and stability of directional representation.

[0039] The S1 stage of this invention employs a processing path of "depthiably separable convolution + dual-path pooling of GMP and GAP + dimensionality increase-dimensionality reduction-split-add," which is significantly different from traditional schemes that rely solely on a single path to compress features. Using only GAP often weakens orientation-sensitive information such as edges and corners; using only GMP easily ignores the overall distribution, scale, and texture trends of the target. This invention, through dual-path parallel encoding, collaboratively fuses local salient structural information with global distribution information, enabling the final feature descriptor to possess both orientation discrimination and scale adaptability. This provides a more stable and targeted input for subsequent angle prediction and combined weight generation. This is not a common feature compression method, but an encoding strategy specifically designed for orientation modeling tasks.

[0040] This scheme is particularly well-suited to the practical needs of target detection in remote sensing images. Target orientations in remote sensing images are typically continuously distributed within the range of 0° to 360°, lacking stable priors, and exhibiting drastic scale variations and significant aspect ratio differences. Targets such as ships, vehicles, and aircraft may appear in any orientation, and similar targets may present drastically different sizes and morphological characteristics under different shooting conditions. Faced with such complex scenarios, traditional methods relying on dense angular anchor frames or position-by-position multi-angle rotation not only introduce a large amount of redundant computation but also struggle to balance lightweight design and real-time performance. This invention, however, uses a single-angle rotation sampling mechanism to predict only the optimal orientation, combined with dynamic weighted fusion of multiple basic convolutional kernels, to achieve an effective representation of multi-directional structural features with a single transformation. Therefore, it better meets the application requirements of high efficiency and high robustness in remote sensing scenarios.

[0041] Remote sensing targets often possess prominent local edges and geometric contours, such as ship hull outlines, vehicle edges, and aircraft wings and tails. This information serves as both a basis for identification and crucial clues for orientation determination. Traditional methods are prone to orientation estimation biases when local features are unclear, target density is low, or noise levels are high, further impacting subsequent detection accuracy. This invention extracts orientation-sensitive structures such as edges and corners using Geometric Modeling (GMP), and then utilizes Geometric Approach (GAP) to encode the overall distribution and scale trends. This allows local and global information to participate in angle prediction and convolution kernel generation, thereby reducing the interference of local errors on global orientation determination and improving stability in complex backgrounds. Particularly when the target's aspect ratio varies significantly, this mechanism maintains strong geometric adaptability without requiring independent modeling of multiple orientations.

[0042] The "single-angle rotation sampling and multi-kernel dynamic combination" mechanism proposed in this invention is not a simple replacement of existing rotational convolution or dynamic convolution, but a systematic optimization scheme formed under the specific technical scenario of remote sensing image target detection. Its core lies in: capturing the most representative principal direction through a single direction prediction, achieving collaborative expression of multiple basic kernels through weight combination, and then completing direction alignment with explicit geometric rotation, thereby obtaining high-quality direction-sensitive features with lower computational cost. This scheme not only significantly reduces the computational and memory burden caused by multi-angle enumeration, but also improves the model's accuracy, stability, interpretability, and engineering deployment value, demonstrating substantial technological progress that is difficult to foresee compared to traditional methods.

[0043] like Figure 1 As shown, this embodiment of the invention provides a method for extracting features by single-angle rotation sampling for target detection in remote sensing images. The method includes: S1: Lightweight feature branch, which performs efficient feature extraction on the input feature map and generates refined feature descriptors for predicting weights and angles; S2: Weighted-angle dual-path prediction head, based on refined features, predicts a set of optimal rotation angles in parallel. and the corresponding combination weights ; S3: Multi-core dynamic combination, serving as the output and application stage of the routing function. This part is based on predictions. and A one-time rotation and weighted fusion is performed on m basic convolution kernels to generate a dynamic rotation kernel for final convolution, thereby achieving multi-angle feature extraction in a single operation.

[0044] S1 specifically includes: A lightweight, depthwise separable convolution is used to extract refined features for predicting weights and angles, enabling spatial-channel decoupling modeling of the input proposal features at extremely low computational cost, as shown in the following formula: (1) Input image features , The number of channels for the input feature. These are the height and width of the feature, respectively; For a 3×3 depthwise separable convolution, () represents layer normalization, and SiLU() represents the activation function; Then, the activated features are processed by global max pooling and global average pooling respectively to obtain two different vectors. and Finally, connect them into a single entity with... The feature vectors are dimensional; Global Max Pooling (GMP) preserves local maxima in the feature map and is highly selective for high-response regions, such as edges and corners, providing discriminative cues for rotation angle estimation; Global Average Pooling (GAP) aggregates spatial global responses to encode the overall distribution characteristics of the target, such as texture density and activation region area, whose statistical properties are significantly correlated with the target scale; by combining the advantages of both, concat is used to connect the vectors. and To obtain the feature descriptor, the formula is as follows: (4) Obtain the feature vector Then, first connect the first fully connected layer. Upscaling the dimensionality has the advantage of performing nonlinear transformations in higher-dimensional space, enabling the learning of more complex feature interactions. For example, for angle prediction, upscaling can implicitly encode the combination patterns of key points in different attitudes, such as the joint features of wing angle and fuselage angle; then... Dimensionality reduction to 2× The increased dimension forces the network to discard redundant information and retain the features most effective for the task. Then, it is split into two vectors, and finally, the elements are added together to obtain the final vector. A D-dimensional vector. The Split operation is equivalent to allowing two D-dimensional vectors to learn different emphases from the same source features, as shown in the following formula: , (5) (6) (7) The dimension is the dimension after being upgraded by the fully connected layer; The dimension is the dimension after dimensionality reduction by the fully connected layer. Note that the original image after dimensionality reduction should have the same dimension as the original input feature map. It is a vector with equal dimensions after being split. Finally for The vector obtained by adding pixels one by one in each dimension.

[0045] S2 specifically includes: Feature vector The input is passed to two different branches; the first branch is the rotation angle prediction branch, consisting of a linear layer and a softsign activation function. The bias of this linear layer is set to false to avoid learning biased angles. The softsign activation function is used because angle prediction involves positive and negative values, and compared to the tanh activation function, the softsign function asymptotically tends to ±1. This lower saturation rate can avoid the gradient vanishing problem to some extent, and compared to tanh, it has no exponential operations, making it computationally lighter. In addition, the output of the softsign layer is multiplied by a coefficient to expand the rotation range, as shown in the following formula: (8) in Let be a coefficient, a learnable parameter, initially set to . , These are the parameters for the linear layer.

[0046] The second branch is the combined weight prediction branch, which is responsible for predicting the combined weights λ. It consists of a biased linear layer and a sigmoid activation function.

[0047] (9) in For linear layer parameters, This is a bias term.

[0048] S3 specifically includes: Given a specific input feature map, when extracting target features from the input feature map, the convolutional kernel may have a potential angle θ= After rotating the combined weights, λ= Then, it is expanded to m convolutional kernels, that is, the predicted angles and weights are: (10) The weights of m convolutional kernels The shape of each convolutional layer is Given input features x, m kernels first predict the rotation angle. Perform the rotations separately, as follows: First, the i-th convolutional kernel is The rotation angle is = Given a rotation angle θ, the counterclockwise direction is defined as the positive direction. The transformation of the new coordinate parameter weights is completed through kernel coordinate rotation and interpolation sampling. Step 1: Set the coordinates as Rotation The coordinates after the angle are (11) The second step is to use bilinear interpolation to calculate the discrete convolution kernel. weight Discrete convolution kernel The weight is defined as ,in For non-integer coordinates ( Take the four nearest grid points and perform linear weighting; first find the four nearest surrounding grid points, then determine... adjacent integer coordinates of direction (12) (13) Corresponding weight Next, the bilinear interpolation formula is used to calculate... (14) Following the steps above, there are a total of m convolutional kernels, so the final result is... Then, the rotated convolution kernel is convolved with the input feature map, and the output feature map is added element by element. (15) in, y represents the combined weights predicted by the routing function, ∗ represents the convolution operation, and y is the combined output feature map.

[0049] This invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method for single-angle rotation sampling and feature extraction for target detection in remote sensing images.

[0050] This invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for single-angle rotation sampling and feature extraction for target detection in remote sensing images.

[0051] To objectively and quantitatively verify the effectiveness of the technical solution of this invention, a series of rigorous comparative experiments were designed and implemented. The experiments aimed to comprehensively evaluate the performance advantages of this invention compared to existing mainstream rotating target detection methods from four core dimensions: detection accuracy, computational efficiency, model complexity, and generalization ability. The experiments were conducted on the widely recognized remote sensing target detection benchmark dataset DOTA-v1.0, which contains a large number of aerial image targets of different scales, orientations, and backgrounds.

[0052] 1. Experimental setup and comparison methods Dataset and Evaluation Metrics: The DOTA-v1.0 dataset was used. The model was trained on the training set according to the official partitioning, and the results were reported on the validation set. The primary evaluation metric was mean average precision (mAP), with an IoU threshold of 0.5 during computation. Additionally, the model's parameter count (params), floating-point operations (FLOPs, calculated with an input image size of 1024×1024), and inference speed (FPS, frames per second) measured on actual hardware (NVIDIA Tesla V100) were reported to comprehensively measure efficiency.

[0053] Three representative types of rotating target detectors were selected as baselines for comparison: The method based on dense rotating anchor boxes is Rotated Faster R-CNN (RoI Transformer). This method aligns features on pre-defined dense angular anchor points and is a classic method with high accuracy.

[0054] Angle regression-based method: S²A-Net. This method optimizes angle prediction through feature decoupling and adaptive anchor point mechanism, achieving a good balance between accuracy and efficiency.

[0055] Anchor-Free Method: CSL (Circular Smooth Label) + FCOS. This method treats angle prediction as a classification problem, avoiding anchor box design and resulting in a relatively simple structure.

[0056] In order to ensure a fair comparison, the core Dynamic Rotation Convolution (DRC) module of this invention was used as an enhancement plugin for feature extraction networks (such as ResNet-50) and embedded into detection frameworks based on RetinaNet and FCOS respectively, thus constructing two models: DRC-RetinaNet and DRC-FCOS.

[0057] 2. Main Experimental Results and Analysis (1) Comparison of detection accuracy (mAP) The table below shows the mean accuracy (AP) and overall mAP of the present invention and the comparison method on the DOTA-v1.0 validation set for each category.

[0058] Table 1: Comparison of detection accuracy on the DOTA-v1.0 validation set (mAP@0.5) As shown in Table 1, both implementations of this invention demonstrate superior performance. DRC-RetinaNet achieved an overall mAP of 81.53%, and DRC-FCOS achieved an overall mAP of 81.15%, both significantly outperforming all baseline methods, including RotatedR-CNN (76.31%) and S²A-Net (77.49%) based on dense anchor boxes, as well as the current state-of-the-art anchor-free method CSL-FCOS (79.60%). This fully demonstrates that the core module (DRC) of this invention can bring significant performance gains in different detection frameworks (RetinaNet and FCOS). The advantages of this invention are particularly evident in categories with drastic changes in target shape and orientation, such as bridges (BR), small vehicles (SV), heliports (HC), and large vehicles (LV). This directly proves that the dynamically rotated convolutional kernel can better fit the geometric features of targets with arbitrary orientations, extracting purer and more discriminative features, thereby bringing a substantial improvement in accuracy. In particular, DRC-RetinaNet achieved the highest AP across multiple categories (such as PL, BR, GTF, SH, SBF, HA, etc.), demonstrating its strong feature adaptation capabilities.

[0059] (2) Comparison of computational efficiency and model complexity To verify the core advantages of this invention in terms of high efficiency and accuracy, the model complexity and inference speed of various methods were compared.

[0060] Table 2: Comparison of Model Efficiency and Complexity (Backbone Network: ResNet-50) Combination Figure 2 and Figure 3 It can be seen that the present invention exhibits significant comprehensive performance advantages in complex remote sensing scenarios. Figure 2 This indicates that the method of this invention lies in the upper right region of the Pareto front in the FPS–mAP tradeoff relationship. While maintaining an inference speed of over 17 FPS, it achieves a detection accuracy of over 81%, demonstrating excellent synergy between computational efficiency and accuracy. Compared to traditional methods, which often involve a significant increase in computational load when improving accuracy, this invention effectively reduces redundant computation and achieves efficient inference through a lightweight routing mechanism and a single-angle sampling strategy. Figure 3The model's detection performance in various typical remote sensing scenarios was further verified, including densely parked vehicles, port ships, airport aircraft, and targets in residential areas. It was able to achieve high-precision positioning and stable detection, and maintained good robustness, especially under conditions of dense distribution of small targets and complex background interference. This demonstrates that the invention has strong generalization ability and engineering application value.

[0061] DRC-RetinaNet achieved the highest detection accuracy (81.53% mAP) while having a computational cost (FLOPs) of only 138.2G, the lowest among all compared methods. DRC-FCOS maintained one of the highest accuracy levels (81.15% mAP) while having the fewest parameters (31.9M). This directly confirms that the lightweight routing network and single-angle sampling mechanism effectively avoid the huge computational overhead of traditional multi-angle enumeration, achieving extreme optimization of computational resources. DRC-FCOS, with 31.9M parameters, is the model with the fewest parameters, even slightly lower than the structurally simple CSL-FCOS baseline (32.1M). This demonstrates the lightweight design of the dynamic convolutional kernel generation module, introducing only a very small number of additional parameters.

[0062] On the speed-accuracy tradeoff curve (FPS-mAP), this invention (two green stars) is located at the upper right of the Pareto front. This indicates that, at similar inference speeds (17.9-18.2 FPS), this invention achieves the highest detection accuracy; or, to achieve similar high accuracy (e.g., >81% mAP), this invention is the only method that can be implemented while maintaining high efficiency (>17 FPS). This perfectly solves the long-standing dilemma of "precision and efficiency being difficult to achieve simultaneously." Figure 2 Speed ​​(FPS) vs. accuracy (mAP) tradeoff curve. The horizontal axis represents FPS, and the vertical axis represents mAP. This invention (green star) is located in the upper right corner of the curve, indicating optimal performance. It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0063] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for extracting features by single-angle rotation sampling for target detection in remote sensing images, characterized in that, Includes the following steps: S1, the lightweight feature branch, performs depthwise separable convolution, layer normalization, and activation processing on the input feature map to obtain intermediate features; performs global max pooling and global average pooling on the intermediate features respectively, and concatenates the pooling results to obtain a feature vector; the feature vector is then subjected to dimensionality increase, dimensionality reduction, segmentation, and element-wise addition processing in sequence to obtain a feature vector for subsequent prediction. S2, weighted angle dual-path prediction head, inputs the feature vector into the rotation angle prediction branch and the combined weight prediction branch respectively, and outputs a set of rotation angles and a set of combined weights corresponding to the rotation angles; S3, Multi-kernel dynamic combination: Rotate multiple basic convolutional kernels according to the rotation angle, and then perform weighted fusion of the rotated basic convolutional kernels according to the corresponding combination weights to obtain a dynamic rotation kernel. The dynamic rotation kernel is then used to convolve the input feature map to output the target feature.

2. The method for extracting features by single-angle rotation sampling for target detection in remote sensing images according to claim 1, characterized in that, In step S1, after performing depthwise separable convolution on the input feature map, layer normalization and SiLU activation are performed sequentially. The activated features are subjected to global max pooling and global average pooling respectively to obtain two vectors representing high response region information and overall distribution information, respectively. The two vectors are concatenated and then fed into the first fully connected layer for dimensionality increase, and then fed into the second fully connected layer for dimensionality reduction. The dimensionality reduction result is divided into two equal-dimensional vectors, and the two equal-dimensional vectors are added element by element to obtain the feature vector used for subsequent prediction.

3. The method for extracting features by single-angle rotation sampling for target detection in remote sensing images according to claim 1, characterized in that, In step S2, the rotation angle prediction branch consists of a linear layer and softsign activation. The linear layer does not have a bias term, and the output after softsign activation is multiplied by the learnable coefficients to obtain the rotation angle. The combined weight prediction branch consists of a linear layer with a bias term and a Sigmoid activation to obtain combined weights corresponding to the rotation angle.

4. The method for extracting features by single-angle rotation sampling for target detection in remote sensing images according to claim 1, characterized in that, In step S3, there are m basic convolutional kernels, and each basic convolutional kernel corresponds to a set of predicted rotation angles and a set of predicted combination weights. When rotating each basic convolution kernel, the kernel coordinates are first rotated and transformed, and then bilinear interpolation is used to calculate the weight values ​​of the discrete convolution kernels to obtain the rotated convolution kernels.

5. The method for extracting features by single-angle rotation sampling for target detection in remote sensing images according to claim 4, characterized in that, In step S3, the convolution kernels obtained by rotating each basic convolution kernel are convolved with the input feature map, and the convolution results are summed element-wise according to the corresponding combination weights to obtain the output feature map.

6. A method for extracting features by single-angle rotation sampling for target detection in remote sensing images, characterized in that, The method includes: S1: The lightweight feature branch performs efficient feature extraction on the input feature map, generating refined feature descriptors for predicting weights and angles; S2: Weighted-angle dual-path prediction head, based on refined features, predicts a set of optimal rotation angles in parallel. and the corresponding combination weights ; S3: Multi-core dynamic combination, serving as the output and application stage of the routing function. This part is based on predictions. and A one-time rotation and weighted fusion is performed on m basic convolution kernels to generate a dynamic rotation kernel for final convolution, thereby achieving multi-angle feature extraction in a single operation.

7. The method for extracting features by single-angle rotation sampling for target detection in remote sensing images according to claim 6, characterized in that, S1 specifically includes: A lightweight, depthwise separable convolution is used to extract refined features for predicting weights and angles, enabling spatial-channel decoupling modeling of the input proposal features at extremely low computational cost, as shown in the following formula: (1) Input image features , The number of channels for the input feature. These are the height and width of the feature, respectively; For a 3×3 depthwise separable convolution, () represents layer normalization, and SiLU() represents the activation function; Then, the activated features are processed by global max pooling and global average pooling respectively to obtain two different vectors. and Finally, connect them into a single entity with... The feature vectors are dimensional; Global Max Pooling (GMP) preserves local maxima in the feature map and is highly selective for high-response regions, such as edges and corners, which are directionally sensitive structures, providing discriminative clues for rotation angle estimation; Global Average Pooling (GAP) aggregates the spatial global response to encode the overall distribution characteristics of the target, such as texture density and activation region area, whose statistical characteristics are significantly correlated with the target scale; by combining the advantages of these two approaches, the feature descriptor is obtained, as shown in the following formula: (4) Obtain the feature vector Then, first connect the first fully connected layer. Upscaling the dimensionality has the advantage of performing nonlinear transformations in higher-dimensional space, enabling the learning of more complex feature interactions. For example, for angle prediction, upscaling can implicitly encode the combination patterns of key points in different attitudes, such as the joint features of wing angle and fuselage angle; then... Dimensionality reduction to 2× The dimensionality forces the network to discard redundant information and retain the features most effective for the task; then it is split into two vectors, and finally the elements are added together to obtain the result. A D-dimensional vector; the Split operation is equivalent to allowing two D-dimensional vectors to learn different emphases from the same source features, as shown in the following formula: , (5) (6) (7) The dimension is the dimension after being upgraded by the fully connected layer; The dimension is the dimension after dimensionality reduction by the fully connected layer. Note that the original image after dimensionality reduction should have the same dimension as the original input feature map. It is a vector with equal dimensions after being split; finally for The vector obtained by adding pixels one by one in each dimension.

8. The method for extracting features by single-angle rotation sampling for target detection in remote sensing images as described in claim 6, characterized in that, S2 specifically includes: Feature vector The input is passed to two different branches; the first branch is the rotation angle prediction branch, consisting of a linear layer and a softsign activation function. The bias of this linear layer is set to false to avoid learning biased angles. The softsign activation function is used because angle prediction involves positive and negative values, and compared to the tanh activation function, the softsign function asymptotically tends to ±1. This lower saturation rate can avoid the gradient vanishing problem to some extent, and compared to tanh, it has no exponential operations, making it computationally lighter. In addition, the output of the softsign layer is multiplied by a coefficient to expand the rotation range, as shown in the following formula: (8) in Let be a coefficient, a learnable parameter, initially set to . , For linear layer parameters; The second branch is the combined weight prediction branch, which is responsible for predicting the combined weight λ; it consists of a biased linear layer and a sigmoid activation function. (9) in For linear layer parameters, This is a bias term.

9. The method for extracting features by single-angle rotation sampling for target detection in remote sensing images as described in claim 6, characterized in that, S3 specifically includes: Given a specific input feature map, when extracting target features from the input feature map, the convolutional kernel may have a potential angle θ= After rotating the combined weights, λ= Then, it is expanded to m convolutional kernels, that is, the predicted angles and weights are: (10) The weights of m convolutional kernels The shape of each convolutional layer is Given input features x, m kernels first predict the rotation angle. Perform the rotations separately, as follows: First, the i-th convolutional kernel is The rotation angle is = Given a rotation angle θ, the counterclockwise direction is defined as the positive direction; the transformation of the new coordinate parameter weights is completed through the steps of kernel coordinate rotation and interpolation sampling. Step 1: Set the coordinates as Rotation The coordinates after the angle are （11） The second step is to use bilinear interpolation to calculate the discrete convolution kernel. weight Discrete convolution kernel The weight is defined as ,in For non-integer coordinates ( Take the four nearest grid points and perform linear weighting; first find the four nearest surrounding grid points, then determine... adjacent integer coordinates of direction (12) (13) Corresponding weight Secondly, the bilinear interpolation formula is used to calculate... (14) Following the steps above, there are a total of m convolutional kernels, so the final result is... Then, the rotated convolution kernel is convolved with the input feature map, and the output feature map is added element by element. (15) in, y represents the combined weights predicted by the routing function, ∗ represents the convolution operation, and y is the combined output feature map.

10. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for single-angle rotation sampling and feature extraction for target detection in remote sensing images as described in any one of claims 1-5.

Images