Rotating target detection method and system based on entropy-guided angular super-resolution

Abstract

A rotating target detection method and system based on entropy-guided angular super-resolution includes the following steps: constructing multi-layer pyramid features of the input image, and determining the deep features to be compensated based on the decrease in local directional expressive power caused by downsampling; calculating spectral angular entropy and frequency band energy saliency based on the local frequency domain directional energy distribution of the deep features, and generating a continuous angle compensation weight map; aligning high-frequency angular details in the shallow features to the deep features, and injecting them in the form of residuals under the guidance of the compensation weights; performing high-frequency context calibration on the enhanced multi-layer features, and outputting the rotating target detection result. This invention adaptively locates regions with insufficient directional evidence through spectral angular entropy, recovers deep angular details using shallow structural cues, improves the stability of the rotating bounding box localization, and has good plug-in deployment value.

Description

Technical Field

[0001] This invention relates to a rotating target detection method and system based on entropy-guided angle super-resolution, belonging to the fields of computer vision, intelligent interpretation of remote sensing images, and rotating target detection technology. Background Technology

[0002] Remote sensing images are typically acquired from a high-altitude, top-down perspective, which differs significantly from the perspective projection commonly found in natural images. From this perspective, the planar geometry of a target is more clearly revealed. Constrained by physical limitations such as aerodynamics (e.g., aircraft) or hydrodynamics (e.g., ships), these typical targets often exhibit highly regular axial extensions and symmetrical structures. For such targets, traditional horizontal bounding boxes can only provide coarse localization and size estimations, containing not only a large amount of redundant background noise but also failing to accurately represent the target's true spatial orientation and attitude. Therefore, rotating target detection in this field is not simply about regressing an angle but also about capturing geometric evidence that defines the target's true orientation.

[0003] Existing rotating target detection paradigms typically rely on Feature Pyramid Networks (FPNs) to extract multi-scale semantic features, followed by angle regression or bounding box prediction in a dedicated detection head. This approach effectively addresses the challenge of drastic target scale variations in remote sensing scenes. However, constructing the feature pyramid requires repeated downsampling operations. While this process expands the receptive field and highly abstracts semantic information, it irreversibly and progressively weakens orientation-sensitive high-frequency geometric structures, such as oblique boundaries, corner transitions, and principal axis symmetry responses within the target.

[0004] From the mathematical essence of discrete sampling of digital signals, the root cause of the above problems lies in "angular resolution decay." As the feature hierarchy deepens, the size of the local sampling grid continuously shrinks, which mathematically limits the number of independent directions that can be unambiguously represented. When the deep features themselves have lost the structural evidence to support high-precision direction judgment, the model is prone to problems such as direction ambiguity, positioning jitter, and aliasing of dense target features. Currently, many studies are dedicated to improving angle representation, rotation-aware feature extraction mechanisms, or detector architecture design. However, these methods generally have a blind spot: they mainly focus on optimizing how to predict direction from the final output features, largely presupposing that the feature pyramid has naturally preserved sufficient angular information. This approach of only patching the prediction terminal fails to fundamentally solve the fact that directional details are progressively lost at the feature level.

[0005] Therefore, breaking through this bottleneck requires tracing back to the source and constructing a technical solution to directly compensate for angular resolution attenuation within the feature pyramid. This solution should enable deep semantic features to maintain their global receptive field and semantic discriminative power while using low-level pyramid features as a direct source of high-frequency geometric information to reacquire key angular details. Furthermore, considering the spatial non-uniformity of feature information attenuation, a set of quantitative indicators must be established to assess the directional uncertainty of local regions of deep features. This should then serve as a guide to adaptively control the compensation intensity, thereby accurately recovering angular evidence without introducing background noise interference. Summary of the Invention

[0006] The purpose of this invention is to address the problem in existing rotating target detection methods that easily lose high-frequency directional cues in deep features of the feature pyramid, leading to unstable angle localization. This invention proposes a rotating target detection method and system based on entropy-guided angle super-resolution. This invention measures the local directional uncertainty of deep features through spectral angular entropy and aligns high-frequency angular details from shallow features into the deep locations requiring compensation, thereby improving the representation of direction-aware features in rotating target detection.

[0007] The present invention is achieved using the following technical solution.

[0008] A rotating target detection method and system based on entropy-guided angular super-resolution includes the following steps:

[0009] Step 1: Construct multi-layer features and determine the angular resolution attenuation relationship.

[0010] The input remote sensing image is fed into the backbone network and feature pyramid of the rotating target detection network to obtain multi-layer features. Shallow features Preserving more boundaries, corners, and slender structures, deep features It has stronger semantic expression but is more prone to losing directional details. For a local sampling radius of... The set of axial directions that can be represented by the grid is:

[0011]

[0012] When feature map downsampling leads to a reduction in the effective sampling radius, it can be represented by a shrinking of the direction set, resulting in angular resolution attenuation. This step identifies shallow features as the source of high-frequency angular details and deep features as the target layer to be compensated.

[0013] Step 2: Generate continuous angle compensation weights based on spectral angle entropy.

[0014] For each deep feature to be compensated First through Convolution yields a single-channel response map, and at each spatial location... Extract the local window. Apply a two-dimensional Hann window to the local window, calculate the centered Fourier power spectrum, and then divide the ring frequency band (after removing the DC component and extreme high-frequency corner responses) according to the axial angle. Each directional sector is a soft directional sector. The directional spectral energy is obtained by aggregating the sectors in each direction. and normalized to a probability distribution. The local spectral angular entropy is defined as:

[0015]

[0016] in, A larger value indicates more dispersed local energy and less clear directional evidence of deep features at that location. To avoid flat areas being misjudged as locations requiring compensation, the frequency band energy significance ratio is further calculated. And generate continuous compensation weights:

[0017]

[0018] in, and For learnable scalar parameters, all spatial locations Composition weight graph .

[0019] Step 3: Align shallow high-frequency angular details with deep features.

[0020] shallow features Perform stepwise downsampling to make its spatial size match the target's deep features. Align, and get Then, deformable convolution is used to extract transferable angular details from the aligned shallow features:

[0021]

[0022] Then, perform residual injection based on the continuous weight graph obtained in step 2:

[0023]

[0024] in, For learnable hierarchical scaling parameters, This indicates element-wise multiplication. This step allows for more shallow directional details at locations with uncertain angles, while locations with sufficient directional evidence remain less perturbed.

[0025] Step 4: Perform high-frequency context calibration and output rotation detection results.

[0026] Features after injection and shallow features Perform high-frequency context calibration. For input features First, construct parameterless high-frequency priors:

[0027]

[0028] Then, spatial calibration weights are generated respectively. and channel calibration weights And obtain the calibration output:

[0029]

[0030] The calibrated multi-layer features are fed into the rotating target detection head, which outputs the target category, center position, size, and angle parameters to form the rotating target detection result.

[0031] Based on the above method, the present invention further proposes a rotating target detection system based on entropy-guided angular super-resolution, including: a multi-layer feature generation subsystem 10, a spectral angular entropy-guided subsystem 20, an angular detail injection subsystem 30, and a high-frequency calibration and detection subsystem 40.

[0032] The connections between the above components are as follows: the multi-layer feature generation subsystem 10 receives remote sensing images and outputs multi-layer pyramid features; the spectral angle entropy guidance subsystem 20 receives deep target features and generates a continuous compensation weight map; the angle detail injection subsystem 30 receives shallow features, deep features, and compensation weight map, and outputs deep features enhanced with angle details; the high-frequency calibration and detection subsystem 40 performs spatial and channel calibration on the enhanced features and outputs the rotating target detection results.

[0033] Beneficial effects

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] 1. This invention overcomes the limitations of relying solely on the detection head to predict angles. Starting from the mathematical essence of discrete sampling, this invention traces back to the source and directly compensates for direction-sensitive high-frequency geometric evidence across layers within the feature pyramid, eliminating directional ambiguity at the feature level and significantly improving the stability and accuracy of rotating target localization.

[0036] 2. This invention constructs a quantitative evaluation mechanism based on information theory. By calculating the ratio of local frequency domain spectral angular entropy to energy saliency, a continuous weighted map is generated to quantify the directional uncertainty of deep features. This gating mechanism enables on-demand compensation, effectively avoiding damage to the original sharp regions and amplification of background noise.

[0037] 3. This invention utilizes deformable alignment and cross-layer injection to explicitly reconstruct the oblique boundary and principal axis structures of deep defects, and combines a high-frequency context calibration mechanism to eliminate spatial response inhomogeneity and channel redundancy. This scheme employs a lightweight plug-in design to achieve consistent performance gains with minimal computational overhead. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the spectral entropy-guided angle detail injection structure of the present invention.

[0039] Figure 2 This is a schematic diagram of the system composition of the present invention. Detailed Implementation

[0040] To better illustrate the purpose and advantages of the present invention, the method of the invention will be further described below in conjunction with the accompanying drawings and examples.

[0041] Example

[0042] like Figure 1 As shown, a rotating target detection method and system based on entropy-guided angular super-resolution includes the following steps.

[0043] Step 1: The input remote sensing image is fed into the backbone network and feature pyramid to obtain multi-layer features from shallow to deep. Shallow features retain more local boundaries, corners, and target axial structures, while deep features retain stronger semantic information. Since downsampling reduces the number of distinguishable directions in the local sampling grid, deep features are identified as the objects requiring angular detail compensation.

[0044] Step 2: Calculate the spectral angular entropy for each deep feature to be compensated. Specifically, the deep feature is compressed into a single-channel response, local windows are extracted at each spatial location, the windows are frequency domain transformed, and the spectral energy distribution in different axial directions is statistically analyzed in the ring frequency band. When the directional spectral energy distribution is relatively dispersed and the region has obvious frequency band energy, it indicates that there is directional uncertainty at that location, and more angular detail compensation is needed.

[0045] Step 3: Using shallow features as the source of angular details, downsample them to the spatial size of the target deep features, and then extract transferable high-frequency angular cues through deformable convolution. Subsequently, using the continuous weight map obtained in Step 2, the angular details are injected into the deep features in the form of residuals, making the compensation process positionally adaptive.

[0046] Step 4: Perform high-frequency context calibration on the injected multi-layer features. This calibration first obtains high-frequency priors by utilizing the difference between the local average response and the original response, then calculates spatial weights and channel weights separately, and finally enhances the local structure and effective channels through residuals. The calibrated features are then fed into the rotating target detection head, which outputs the target category, location, scale, and angle.

[0047] Among them, the rotating target detection method and system based on entropy-guided angle super-resolution in this application solves the problem of unstable angle positioning caused by insufficient deep feature direction evidence through spectral angle entropy guidance, shallow angle detail injection, high frequency context calibration and rotation detection output, which helps to improve the quality of rotating target detection in complex remote sensing scenes.

[0048] Furthermore, in one example of this application, the spectral angle entropy guiding subsystem 20 calculates the directional spectral energy distribution for each local window of the deep feature and obtains the directional uncertainty by normalizing the Shannon entropy; the angle detail injection subsystem 30 uses this uncertainty as a continuous gating weight, so that the shallow high-frequency structure only has a strong influence on the deep feature at the required location.

[0049] Figure 2 A schematic diagram of a rotating target detection system based on entropy-guided angular super-resolution provided for this application example includes a multi-layer feature generation subsystem 10, a spectral angular entropy-guided subsystem 20, an angular detail injection subsystem 30, and a high-frequency calibration and detection subsystem 40.

[0050] The multi-layer feature generation subsystem 10 is used to generate multi-layer pyramid features from shallow to deep layers from the input image.

[0051] The spectral angle entropy guiding subsystem 20 is used to generate continuous angle compensation weights based on the local directional spectral energy distribution of deep features.

[0052] The angular detail injection subsystem 30 is used to align shallow high-frequency angular details and residually inject them into deep target features.

[0053] The high-frequency calibration and detection subsystem 40 is used to perform spatial and channel calibration on the enhanced multi-layer features and output the detection results of the rotating target.

[0054] The connection relationship between the above-mentioned components is as follows: the input remote sensing image first enters the multi-layer feature generation subsystem 10; the deep features in the multi-layer features enter the spectral angular entropy guidance subsystem 20 to generate compensation weights; the shallow features, deep features and compensation weights enter the angular detail injection subsystem 30 together to obtain enhanced features; the enhanced features enter the high-frequency calibration and detection subsystem 40 and output the rotation detection results.

[0055] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is merely a specific example of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A rotating target detection method based on entropy-guided angular super-resolution, characterized in that, Includes the following steps: Step 1: Feed the input image into the backbone network and feature pyramid to obtain multi-layer features. The shallow features are used as the source of high-frequency angular details, and the deep features are used as the target features to be compensated. Step 2: Calculate the ratio of local spectral angular entropy and frequency band energy significance for the deep features to be compensated, and generate a continuous angle compensation weight map; Step 3: Align the shallow features to the spatial dimensions of the deep target features, extract high-frequency angular details, and perform residual injection based on the continuous angle compensation weight map; Step 4: Perform high-frequency context calibration on the injected multi-layer features, and output the rotating target detection results based on the calibrated features.

2. The method as described in claim 1, characterized in that, The angular resolution attenuation mentioned in step 1 is determined by the set of representable directions of the local sampling grid, wherein the set of representable directions is: When feature map downsampling reduces the local sampling radius, the set of directions that can be represented decreases accordingly, and deep features are identified as the target layer that requires angular detail compensation.

3. The method as described in claim 1, characterized in that, Step 2 specifically includes: deep features pass Convolutional mapping is a single-channel response map; In spatial location Extract a local window; Apply a two-dimensional window function to the local window and calculate the centered Fourier power spectrum; Within the ring frequency band, soft-directional sectors are divided according to the axial direction; The spectral energy of each sector is aggregated and normalized into a directional probability distribution.

4. The method as described in claim 3, characterized in that, The local spectral angular entropy is: And combined with the frequency band energy significance ratio Generate continuous angle compensation weights: in, This represents the number of sectors in the soft direction. and For learnable scalar parameters, It is a stable term.

5. The method as described in claim 1, characterized in that, The high-frequency angular details mentioned in step 3 are obtained and injected in the following way: in, To align shallow features to the first Downsampling operator for layer depth feature size, For deformable convolution, This is the scaling parameter for the learnable hierarchy.

6. The method as described in claim 1, characterized in that, The high-frequency context calibration described in step 4 includes: pass Construct high-frequency priors; Spatial calibration weights are generated based on high-frequency priors; Channel calibration weights are generated based on the original features; and through The calibration output is obtained.

7. A rotating target detection system based on entropy-guided angular super-resolution, characterized in that, include: Multi-layer feature generation subsystem 10: used to generate multi-layer pyramid features from shallow to deep layers from the input image; Spectral angle entropy guiding subsystem 20: used to calculate directional uncertainty based on the local directional spectral energy distribution of deep features and generate a continuous angle compensation weight map; Angle detail injection subsystem 30: used to align shallow high-frequency angle details to deep target features and perform residual injection based on a continuous angle compensation weight map; High-frequency calibration and detection subsystem 40: used to perform spatial and channel calibration on the enhanced multi-layer features and output the detection results of rotating targets.

Images