A satellite remote sensing small target detection method based on context affine perception

Abstract

The application provides a satellite remote sensing small target detection method based on context affine perception, and comprises the following steps: image input, downsampling and initial feature extraction; an upsampling stage is used for recovering part of the spatial resolution; features after upsampling are subjected to feature affine, re-modeling and further upsampling processing, and a context perception dynamic affine module (CDAM) is introduced; high-resolution features obtained are input into a fine-grained small target detector (FSDetector); features are subjected to feature affine, re-modeling and re-downsampling processing; joint prediction of targets of different scales is realized through a multi-scale parallel detection mode; output results of the fine-grained small target detector are fused with output results of three different resolution detectors, so that final remote sensing image target detection results are obtained. The application realizes a remote sensing image target detection method which can inhibit semantic attenuation in feature transmission, give consideration to multi-scale perception and strengthen small target detection capability.

Description

Technical Field

[0001] This invention relates to the field of satellite remote sensing detection algorithms, and in particular to a satellite remote sensing small target detection method based on contextual affine perception. Background Technology

[0002] With the rapid development of satellite remote sensing technology, remote sensing target detection is increasingly widely used in urban and rural planning, disaster early warning, and environmental monitoring. In recent years, with breakthroughs in deep learning technology and improvements in computing resources, research on remote sensing target detection has made significant progress. However, limited by the inherent contradiction between spatial and spectral resolution during imaging and the modulation transfer function effect of the optical system, remote sensing images are prone to high-frequency information attenuation during imaging, leading to problems such as spectral shift, structural distortion, and blurred edge textures, ultimately affecting the detection accuracy of small targets.

[0003] The quality of remote sensing images is constrained by the inherent contradiction between spatial and spectral resolution, as well as the limitations of the modulation transfer function of optical sensors, leading to problems such as high-frequency information attenuation, blurred boundaries, weakened texture features, and low-contrast areas. Consequently, the semantic information extracted from degraded images often exhibits gradual attenuation as it propagates along the network hierarchy, making it difficult for detectors to distinguish small or visually insignificant targets in complex backgrounds. To mitigate the loss of discriminative features, early research focused primarily on image-level enhancement techniques—such as contrast stretching, histogram equalization, or illumination correction—attempting to recover lost texture and boundary details before feature extraction. However, these methods only operate in the pixel domain and struggle to achieve adaptive adjustment of feature semantics within the deep model. With technological advancements, super-resolution and image restoration methods based on generative adversarial networks and convolutional neural networks have been introduced at the feature level to enhance high-frequency details and structural clarity. While these methods are effective in certain scenarios, they incur significant computational costs and may introduce artifacts that interfere with subsequent detection processes.

[0004] With the development of deep neural networks, attention-based feature enhancement mechanisms have been widely explored to improve semantic representation and context awareness. Typical examples include SENet, which uses global average pooling to reweight channels; CBAM and ECA modules, which introduce spatial and channel adaptive calibration; and global attention models based on the Transformer architecture. In remote sensing, these techniques have been extended to highlight salient regions and suppress background noise, thereby improving target discrimination capabilities in complex environments. However, existing methods often focus on global context modeling or channel-level weight adjustment, paying insufficient attention to the semantic degradation caused by shallow, weak textures and blurred edges.

[0005] To address these challenges, an increasing number of studies are designing dedicated detection frameworks for remote sensing images, aiming to improve the feature representation capabilities of small targets under complex imaging conditions. Although these methods differ in their architectural design, they all strive to enhance the representational power of fine-grained features and improve the robustness of small target perception. However, most existing methods overlook a crucial issue: the semantic decay effect caused by imaging degradation and continuous downsampling. Degraded remote sensing images weaken high-frequency details and weak texture information in shallow features, thereby reducing the discriminative power of small target features in the deep semantic space; simultaneously, feature compression during downsampling can cause small target signals to be obscured by background information. This degradation in feature integrity and semantic alignment ultimately affects the network's accurate perception of real target regions during feature fusion and detection. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a Collaborative Context Affine Awareness Network (CCAPNet). This algorithm is implemented through four plug-and-play complementary modules: a Context Affine Awareness Dynamics (CDAM) module, a Dual-Domain Awareness Fusion (DPFM) module, a Collaborative Multi-Kernel Module (CMKB) module, and a fine-grained small object detection head (FSDetector). It systematically alleviates the feature degradation problem of small objects caused by weak texture, scale variations, and structural sparsity. Specifically, the four plug-and-play complementary modules are as follows:

[0007] The Context-aware Dynamic Affine Module (CDAM) enhances the semantic representation of deep features through multi-scale convolutional perception and adaptive context aggregation, and achieves the separation and reconstruction of high and low frequency features based on affine transformation, thereby preserving fine-grained spatial details under complex imaging conditions.

[0008] The Dual-range Perception Fusion Module (DPFM) constructs a dual-range perception path and dynamically balances the local and global receptive fields to ensure the preservation of discriminative features of targets at different scales.

[0009] The Collaborative Multi-Kernel Block (CMKB) alleviates the feature sparsity degradation problem in small target regions. This module effectively improves target saliency and stabilizes key regions by enhancing local-global feature contrast and suppressing background redundancy responses.

[0010] The Fine-grained Small-object Detector (FSDetector) module is built upon CMKB enhanced features. By compensating for the loss of fine-grained features due to continuous downsampling, it improves spatial positioning accuracy and maintains consistency between feature enhancement and detection. These components together constitute a complete detection framework, achieving a comprehensive improvement in robust perception, accurate positioning, and effective recognition of small targets in complex remote sensing scenarios.

[0011] The method of the present invention includes the following steps:

[0012] STEP1: Image initialization; including image input, downsampling, and feature extraction;

[0013] STEP2: Feature Filtering; CMKB; Feature filtering is achieved through a reweighting mechanism;

[0014] STEP 3: Enhance DPFM for the feature response selected in step 2;

[0015] STEP4: Upsampling stage; Restoring spatial resolution;

[0016] STEP 5: Feature Affine, Remodeling and Resampling, including CDAM;

[0017] STEP 6: Input the upsampled features from step 5 into the fine-grained small target detector to narrow the receptive field and lock onto the small target;

[0018] STEP 7: Feature affine transformation, remodeling and resampling to expand the receptive field and lock onto large targets;

[0019] STEP 8: Input the different scale features obtained in step 7 into three detectors with different resolutions;

[0020] STEP 9: Combine the output of the small target detector in step 6 with the output of the three detectors with different resolutions in step 8 to obtain the final satellite image detection result.

[0021] STEP1 includes the following:

[0022] STEP 1: Image Initialization. This includes image input, downsampling, and initial feature extraction. First, the remote sensing image to be detected is input into the network as the raw input data. Then, the input image is scaled through continuous convolutional downsampling operations, reducing spatial resolution while extracting basic visual features. This stage primarily extracts low-level features such as edges, textures, and local structures from the raw image, providing initial feature representations for subsequent semantic modeling. This step effectively reduces computational complexity while ensuring that key structural information is not completely lost, and simultaneously completes the mapping from the image domain to the feature domain.

[0023] STEP 2: Feature Selection. STEP 2 performs feature selection using the features obtained in STEP 1. This feature selection is implemented through a Collaborative Multi-Kernel Module (CMKB). In this step, a dynamic feature reweighting mechanism based on multiple receptive fields is introduced to adaptively adjust the feature responses at different spatial locations and channel dimensions, thereby highlighting effective features relevant to the target and suppressing redundant or unstable responses in the background region. This process effectively alleviates the problem of sparse pixels and unfocused feature responses for small targets in remote sensing images, providing cleaner and more focused feature inputs for subsequent feature enhancement.

[0024] STEP 3: Feature Response Enhancement. STEP 3 further enhances the feature responses selected in STEP 2, based on the Dual-Domain Perceptual Fusion (DPFM) module. By introducing multi-scale convolutional perception and local refinement operations, the structural features and detailed information within the target region are enhanced, while maintaining the correlation between the local and neighborhood contexts. This step further improves the saliency of the target region, making the feature responses more stable at both spatial and semantic levels. This enhancement process effectively reduces the risk of response degradation or breakage in deep features for small targets.

[0025] STEP 4: Upsampling Stage. STEP 4 is the upsampling stage, used to restore some spatial resolution. In this step, the enhanced features are upsampled, gradually restoring the feature map to a higher resolution in the spatial dimension, thus creating conditions for subsequent fine feature modeling and small object detection. This process helps alleviate the spatial information loss problem caused by multiple downsampling steps.

[0026] STEP 5: Feature Affine Mapping, Remodeling, and Re-upsampling. STEP 5 performs feature affine mapping, remodeling, and further upsampling on the upsampled features, introducing a Context-Aware Dynamic Affine Module (CDAM). In this step, affine mapping operations on the feature domain separate and reconstruct structure-related and detail-related information from the features, thereby suppressing semantic information decay during feature propagation. Through collaborative modeling of multi-scale contextual information, shallow detail features can be more effectively propagated to higher semantic layers. This step provides a feature representation with stronger semantic consistency and more complete detail preservation for subsequent small object detection.

[0027] STEP 6: Fine-grained small target detection. STEP 6 inputs the high-resolution features obtained in STEP 5 into the fine-grained small target detector (FSDetector). In this stage, by reducing the effective receptive field and enhancing the local response, the detector can more accurately focus on small target regions, thereby improving the localization accuracy and detection reliability of small targets. This step is mainly aimed at small-scale target scenes that are difficult for traditional detection heads to handle.

[0028] STEP 7: Feature Affine, Remodeling, and Resampling. STEP 7 performs feature affine transformation, remodeling, and resampling to expand the receptive field and enhance the perception of large-scale targets. By scaling and reconstructing the context of the features, the network can acquire a wider range of contextual information while maintaining semantic consistency, thereby improving its overall perception of medium- to large-scale targets.

[0029] STEP8: Multi-scale target detection. STEP8 inputs the different scale features obtained in STEP7 into three detectors with different resolutions. Through multi-scale parallel detection, it achieves joint prediction of targets at different scales, enabling the network to adapt to the detection needs of small, medium-scale, and large targets simultaneously, thereby improving the overall robustness and generalization ability of the detection.

[0030] STEP9: Detection Result Fusion. STEP9 fuses the output of the fine-grained small target detector in STEP6 with the output of the three detectors with different resolutions in STEP8 to obtain the final target detection result of the remote sensing image.

[0031] By comprehensively processing the multi-scale detection results, the advantages of detectors at each scale can be fully utilized to achieve unified detection of targets of different sizes in complex remote sensing scenes.

[0032] In YOLOv8, a detection head typically consists of three feature layers at different scales to achieve multi-scale object detection. These three detection heads are essentially structurally identical, but differ in input feature resolution, receptive field, and the scale of the targets they detect. Simply put, the final output is a concatenation of predictions from these three scales. These three detection heads are included by default in the system.

[0033] Through the above steps, this invention realizes a remote sensing image target detection method that suppresses semantic decay during feature transfer, balances multi-scale perception, and enhances small target detection capabilities. It significantly improves detection stability and accuracy under conditions of weak texture, low contrast, and complex backgrounds. The context-aware dynamic affine module proposed in this invention establishes a bidirectional cross-scale information pathway, maintaining the complementary relationship between shallow details and deep semantics during feature fusion. Furthermore, the introduction of a feature reweighting mechanism and contextual consistency constraints effectively alleviates feature decay during multi-stage fusion, thereby significantly improving the representation and discrimination capabilities of targets, especially small targets. Attached Figure Description

[0034] Figure 1 Diagram of the collaborative context affine perception network architecture.

[0035] Figure 2 A context-aware dynamic affine module diagram.

[0036] Figure 3 This is a diagram of a dual-range perception fusion module.

[0037] Figure 4 This is a diagram of a collaborative multi-core module. Detailed Implementation

[0038] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0039] The overall architecture of CCAPNet is as follows: Figure 1As shown, given an input remote sensing image, multi-scale feature maps are first extracted using the CSPDarkNet53 backbone network. These feature maps are then processed by a Context-Aware Dynamic Affine Module (CDAM), which mitigates semantic degradation caused by weak texture and low contrast. Through context-aware dynamic color transformation, CDAM adaptively models structural and detail components, recalibrating semantic information to enhance fine-grained feature representations in areas with blurred boundaries or missing textures, while maintaining global consistency. This process effectively improves the semantic separability and discriminative ability of small targets in complex remote sensing scenes. To enhance scale adaptability during feature fusion, a Dual-Domain Aware Fusion Module (DPFM) adaptively integrates multi-scale features. By dynamically adjusting the receptive field, DPFM can flexibly select feature responses according to changes in target scale, effectively handling the problem of dense small target distribution and scale imbalance. The optimized features are then processed by a Collaborative Multi-Kernel Module (CMKB), which combines local context awareness with cross-channel attention mechanisms to refine semantic responses and suppress background interference. By focusing on salient regions, CMKB enhances feature discrimination and stability, ensuring accurate small target perception in cluttered backgrounds.

[0040] Enhanced and context-interactive features are passed to the Feature-Sensitive Detector (FSDetector), which employs a multi-scale decoding structure to achieve accurate localization and classification. Overall, CCAPNet forms an efficient end-to-end detection framework that significantly improves the robustness and detection accuracy of small and complex targets in challenging remote sensing environments.

[0041] A. Context-Aware Dynamic Affine Module (CDAM)

[0042] like Figure 2 As shown, small targets in remote sensing images typically exhibit weak texture, low contrast, and blurred boundaries. This leads to feature confusion between different target categories when traditional convolutional neural networks extract features. To effectively separate structural information from detailed features and enhance context awareness, this invention proposes the use of a CDAM module. This module employs a frequency-inspired affine transformation to decompose the input feature map into structure-dominated low-frequency components and detail-dominated high-frequency components. Unlike traditional Fourier transform-based methods, this method operates entirely in the spatial domain, avoiding computational overhead and interpolation artifacts, while simultaneously achieving effective low-frequency and high-frequency separation.

[0043] The CDAM module adopts a hierarchical structure consisting of three sub-modules: Structure-Detail Optimization Module (SDRM), Context Adaptive Enhancement Block (CAEB), and Hierarchical Bottleneck Integrator (HBI). The overall architecture is as follows: Figure 3 As shown.

[0044] (1) Structure-Detail Optimization Module (SDRM)

[0045] SDRM is based on frequency-inspired affine transformation, which explicitly separates low-frequency structural information from high-frequency detail information. Given an input feature map X∈R^(B×C×H×W), the structure-dominant low-frequency components are first extracted by downsampling: Xs=Downk(X) (1)

[0046] Then upsample it back to the original resolution. Low-frequency structural features are obtained by element-wise modulation: XL=LN(ConvL(σ(X⊙Upsample(Xs)))) (2)

[0047] Where ⊙ represents element-wise multiplication, σ(·) is the GELU activation function, and LN represents layer normalization. The high-frequency detail components are calculated as follows:

[0048] XH=LN(ConvH(σ(X-α·Upsample(Xs)))) (3)

[0049] Where α∈[0,1] are learnable coefficients controlling low-frequency suppression. The low-frequency and high-frequency features are concatenated and projected to generate the SDRM output:

[0050] XSDRM=σ(Convproj([XL;XH])) (4)

[0051] To achieve dynamic frequency sensing enhancement, adaptive weights are calculated:

[0052] WL=Sigmoid(Convw(XL));WH=Sigmoid(Convw(XH)) (5)

[0053] The final SDRM features are:

[0054] XSDRM=σ(WL⊙XL+WH⊙XH) (6)

[0055] This operation enables the network to adaptively adjust the ratio of structural and detailed information according to different scenarios, enhancing its ability to distinguish small targets with weak textures.

[0056] (2) Context Adaptive Enhancement Block (CAEB)

[0057] CAEB further enriches features by integrating remote context and enhanced details. The input XSDRM is first processed by depthwise convolution and then added to the residual:

[0058] X0=XSDRM+LN(σ(DWConv(XSDRM))) (7)

[0059] This operation preserves local details while retaining the original information. Subsequently, channel compression is applied to the features, and broader spatial dependencies are captured through contextual convolution.

[0060] Xp=LN(σ(Proj1(X0)));Xc=LN(σ(CtxConv(Xp))) (8)

[0061] Enhancements can be made by reapplying SDRM:

[0062] Xe=SDRM(Xp) (9)

[0063] Finally, contextual and detailed features are fused using learnable coefficients:

[0064] XCAEB=σ(βc·Xc+βe·Xe);βc+βe=1 (10)

[0065] This design enhances target boundaries and textures while reducing false detections by leveraging context.

[0066] (3) Hierarchical Bottleneck Integrator (HBI)

[0067] The HBI module integrates multi-level features while preserving residual connections. Given input Xin, the hierarchical transformation is: XHBI=Xin+γ·CAEB(Conv2(Conv1(Xin))) (11)

[0068] Where γ∈[0,1] are learnable coefficients. This structure aggregates shallow and deep features, improves the representation of small targets, and ensures stable gradient propagation.

[0069] (4) CDAM Overall Architecture

[0070] CDAM constructs a hierarchical feature enhancement network by stacking HBI modules. Let Xin be the input and n be the number of stages: Y0=Convexpand(Xin);Yi=HBIi(Yi-1);i=1,...,n (12)

[0071] The final CDAM output is obtained by splicing together the outputs of all stages and projecting them:

[0072] XCDAM=Convproj([Y0,Y1,...,Yn]) (13)

[0073] This hierarchical integration, combining frequency-inspired affine separation and context enhancement, effectively preserves high-frequency details while capturing low-frequency structural and contextual information, thus improving the accuracy and robustness of small target detection in remote sensing images. Through context-aware dynamic affine transformation, CDAM adaptively models structural and detail feature components, performing semantic recalibration. This process maintains global consistency while enhancing fine-grained feature representations in areas with blurred boundaries and degraded textures, effectively improving the semantic separability and discriminative ability of small targets in complex remote sensing environments.

[0074] B. Dual-Domain Sensing Fusion Module (DPFM)

[0075] Due to differences in imaging resolution, flight altitude, sensor viewpoint, and complex background scenes, targets in remote sensing images often exhibit significant scale variations. During feature extraction and downsampling, small target features are easily overwhelmed or lost. Although traditional feature pyramid structures (such as FPN and PAN) achieve multi-scale modeling through hierarchical fusion, small target features are easily dominated by high-level features with large receptive fields, leading to detail degradation, scale response misalignment, and ultimately performance degradation.

[0076] To address this issue, based on the structure-details decoupling strategy in CDAM, we propose a Dual-Domain Aware Fusion Module (DPFM), whose structure is as follows: Figure 4 As shown, DPFM introduces a learnable scale prediction mechanism and a multi-level cross-domain fusion strategy in the spatial domain to achieve a dynamic balance between local details and global semantic features. This enables the network to maintain cross-scale detail and semantic consistency, significantly improving the robustness and generalization ability of small object detection.

[0077] (1) Scale prediction mechanism

[0078] Given an input feature map X∈R^(B×C1×H×W), the channel dimension is first reduced by a 1×1 convolution: Xr=Conv1×1(X) (14)

[0079] This operation reduces channel redundancy, improves computational efficiency, and preserves the dominant structural features.

[0080] Next, the dimensionality-reduced feature map is used to predict the scale response map through lightweight convolution:

[0081] Slogit=Conv1×1(σ(BN(Conv3×3(Xr)))) (15)

[0082] Where σ(·) represents ReLU activation, and Slogit∈R^(B×S×H×W) represents S original responses at different scales.

[0083] To enhance learnability and avoid scale bias, DPFM applies temperature normalization to adaptively adjust the scale probability distribution:

[0084] αs=exp((Slogit(s)+bs) / τ) / Σexp((Slogit(j)+bj) / τ) (16)

[0085] Where τ=exp(ω) is the learnable temperature parameter, and bs is the scale bias term. The temperature parameter controls the sharpness of the distribution and balances the competition between different scales during training.

[0086] Introducing a regularization term to stabilize scale learning:

[0087] Lscale=Σ∥αs-α¯∥22;α¯=(1 / S)Σαs (17)

[0088] This constraint prevents excessive concentration of scale weights and encourages balanced multi-scale responses.

[0089] (2) Multi-level dual-domain fusion

[0090] Using the learned scale weights αs, DPFM extracts multi-scale detail features through pooling:

[0091] Fs=Pools(Xr);s=1,2,...,S (18)

[0092] These features are aggregated by scale-weighted summation to obtain the first-level dynamic scale fusion features:

[0093] F(1)=Σαs⊙Fs (19)

[0094] Achieve adaptive receptive field reconstruction to enhance the salience of small target regions.

[0095] To capture higher-order interactions across scales, the second-level scale convolutional layer performs cross-scale context reconstruction:

[0096] F(2)=Σβs⊙Pools(F(1));βs=Softmax(GAP(Fs)) (20)

[0097] βs is generated through global average pooling to capture global scale dependencies and establish implicit coupling between local details and the global semantic domain.

[0098] Furthermore, a serial multi-layer max-pooling path extracts structural features of the ultra-large receptive field:

[0099] F(3)=Poolkmax(Poolkmax(Poolkmax(Xr))) (21)

[0100] As a supplementary channel to the semantic domain, it provides stable global guidance in dense or complex contexts.

[0101] Finally, all domain features are concatenated and adaptively fused: Fcat=[Xr,F(1),F(2),F(3)] (22)

[0102] C. Collaborative Multicore Module (CMKB)

[0103] In remote sensing images, targets often exhibit significant scale variations and are embedded in highly complex backgrounds. The limited number of pixels in small targets leads to sparse and incomplete feature representations, making it difficult for traditional convolutional structures with fixed receptive fields to balance local detail preservation and global semantic extraction. This imbalance frequently results in feature degradation and unstable responses, ultimately reducing the discriminative power of the detector. To address these challenges, a Collaborative Multi-Kernel Module (CMKB) is proposed, which establishes a collaborative path mechanism consisting of dynamic multi-kernel selection, local optimization, and channel attention. These three paths work together to enhance structural adaptability, recover fine-grained details, and optimize semantic relevance within the spatial domain.

[0104] (1) Dynamic multi-core selection:

[0105] Given an input feature map X∈R^(B×C×H×W), a set of dilated depthwise convolutions with different receptive fields is used to capture multi-scale contextual information:

[0106] Fi = Convdi(X); i ∈ {1, 2, ..., K}

[0107] Where Convdi represents a depthwise convolution with a dilation rate of di. To adaptively fuse features from different receptive fields, CMKB generates a set of dynamic weights through adaptive average pooling, two 1×1 convolutions, and softmax activation.

[0108] w=Softmax(W2δ(W1AAP(X)))

[0109] Where AAP(·) is adaptive average pooling, δ(·) is ReLU activation, and W1 and W2 represent learnable parameters. Dynamically weighted feature aggregation is represented as:

[0110] Fdyn=ΣwiFi

[0111] This adaptive selection mechanism enables the network to maintain the optimal receptive field for each spatial location, effectively balancing semantic abstraction and spatial accuracy. By allowing different hollow kernels to respond to different structural contexts, CMKB ensures the stability of feature representations even under sparse pixel distributions.

[0112] (2) Local feature optimization:

[0113] While dynamic kernel fusion improves scale adaptability, it can lead to reduced edge continuity and blurred fine-grained textures. To address this issue, CMKB introduces a local optimization path, defined as:

[0114] Floc=Conv3×3(σ(BN(Conv3×3(X))))

[0115] Where σ(·) represents ReLU activation and BN represents batch normalization. This path performs a compact residual-like transformation, optimizing spatial continuity and enhancing texture fidelity. The addition of non-grouped convolution ensures local feature recombination, enabling CMKB to recover suppressed boundaries and strengthen discriminative cues, which is crucial for delineation of small objects in complex scenes.

[0116] (3) Channel attention:

[0117] Inspired by the compression incentive mechanism, CMKB introduces a lightweight channel attention branch, adaptively reweighting and fusing features based on semantic relevance. Specifically, the optimized feature Fref is obtained as follows:

[0118] Fref=Fdyn+Floc

[0119] Then, an adaptive channel weighting process is performed: Ac=σ(W4δ(W3AAP(Fref)))Fatt=Fref⊙Ac

[0120] Where ⊙ represents element-wise multiplication. The adaptive pooling operation captures global contextual cues while maintaining scale invariance, and the attention weight Ac selectively enhances target-relevant channels while suppressing redundant or background-dominated activations.

[0121] Finally, the output of CMKB is optimized using a 1×1 convolution to integrate the reweighted features: Y = Conv1×1(Fatt)

[0122] Through the synergistic integration of three mechanisms, CMKB achieves a fine balance between global perception and local fidelity. It effectively mitigates feature degradation caused by sparse representation, stabilizes cross-scale responses, and enhances the detector's discriminative ability in small target scenes.

[0123] D. Fine-grained small target detector (FSDetector)

[0124] In remote sensing images, small targets typically contain limited pixel information and suffer significant spatial degradation due to multiple downsampling operations. Consequently, fine-grained details are gradually lost during feature propagation, leading to incomplete semantic representation and unstable detection responses. Traditional multi-scale detection heads are insufficient in these scenarios because they struggle to preserve spatial granularity while maintaining semantic consistency, especially in regions with densely packed small targets.

[0125] To address these challenges, we introduce the Fine-Grained Small Object Detector (FSDetector), a dedicated detection branch added to the existing YOLOv8 architecture. This branch is specifically designed to recover fine-grained spatial information and enhance the discriminative representation of small objects. Unlike standard detection heads, FSDetector operates on high-resolution shallow features, enabling it to preserve detailed local textures while maintaining the semantic consistency required for accurate detection.

[0126] Formally, given a shallow high-resolution feature map Fs∈R^(B×Cs×Hs×Ws) extracted from the backbone network, the feature enhancement and detection process of FSDetector can be represented as:

[0127] Fenh = Φenh(Fs)

[0128] Here, Φenh(·) represents the fine-grained enhancement operation, which optimizes the local structure and strengthens boundary cues through spatially adaptive feature recalibration. This operation enables the network to suppress irrelevant background noise while emphasizing key spatial responses relevant to small targets.

[0129] Subsequently, the enhanced feature Fenh is passed through a convolutional projection layer to improve feature compactness and discriminativity: Fproj = Conv3 × 3(Fenh)

[0130] Conv3×3(·) represents a 3×3 convolution, which aggregates contextual information while maintaining positioning accuracy.

[0131] Finally, the detection prediction is performed as follows: Ŷs=Detect(Fproj)

[0132] Where Ŷs represents the prediction result of the small target branch. Unlike the traditional three-scale detection hierarchy {P3, P4, P5} in YOLOv8, FSDetector introduces a new fine-grained level P2, forming a four-scale structure {P2, P3, P4, P5}. This design ensures that small targets can be detected from feature maps with higher spatial resolution and more detailed semantic encoding.

[0133] Furthermore, FSDetector collaborates with upstream CDAM and DPFM to form a progressively optimized pipeline. Specifically, CDAM decouples and enhances structure-detail representations, DPFM integrates cross-domain semantic cues, and FSDetector further optimizes fine-grained spatial features and detection response. These modules together construct a hierarchical adaptive feature enhancement system that effectively mitigates the feature degradation of small targets caused by resolution loss, achieving more robust and accurate small target detection in complex remote sensing scenarios.

[0134] The symbols in the above formulas have the following meanings:

[0135] X: Input feature map; B: Batch size; C: Number of channels; H: Feature map height; W: Feature map width; σ(·): Activation function (in this context, SDRM refers to GELU, and the other part refers to ReLU, see the explanation in the formula); LN: Layer Normalization; ⊙: Element-wise Multiplication; [;]: Concatenation; Conv: Convolution; Sigmoid: Sigmoid activation function.

[0136] Based on several publicly available remote sensing target detection datasets, each with unique challenges, the evaluation metrics and experimental settings are then detailed. The effectiveness of the proposed method is validated through ablation experiments, demonstrating the independent contribution of each module. Detection accuracy, feature extraction performance, and comparisons with state-of-the-art methods will be discussed in detail later. The effectiveness is further validated by integrating CCAPNet into different network architectures (particularly the YOLOv8 framework) and datasets.

[0137] A. Introduction to the Experimental Dataset: NWPUVHR-10 Dataset

[0138] The NWPUVHR-10 dataset, released by Northwestern Polytechnical University in 2014, is a commonly used resource in the field of remote sensing applications. Its latest version contains 1,172 images covering ten target categories: aircraft, ships, oil tanks, baseball fields, tennis courts, basketball courts, athletic fields, ports, bridges, and vehicles. These categories are further divided into positive and negative sample subsets. The image spatial resolution ranges from 0.08 to 2 meters.

[0139] B. Evaluation Indicators

[0140] This study uses the mean precision (AP) of each category and the mean mean precision (mAP) of all categories as evaluation metrics, specifically defined as follows:

[0141] AP=∫。 1Pr(Re)dRe(34)

[0142] mAP=(1 / Ncls)ΣAPi(35)

[0143] Where Pr and Re represent precision and recall, respectively, and Ncls is the total number of classes. The formulas for calculating precision (Pr) and recall (Re) are as follows:

[0144] Pr = NTP / (NTP + NFP) (36)

[0145] Re = NTP / (NTP+NFN)(37)

[0146] Here, NTP, NFP, and NFN represent the number of true positives, false positives, and false negatives, respectively. Specifically, a true positive is an instance that is correctly identified and accurately classified; a false positive is an instance that is incorrectly detected or misclassified; and a false negative is an instance that is not detected. During inference, when the Intersection over Union (IoU) between the predicted bounding box and the ground truth bounding box exceeds 0.5, the prediction is considered a true positive; otherwise, it is classified as a false positive. The following are the results of a comparative experiment:

[0147]

[0148] This invention proposes CCAPNet, a unified detection framework designed to address long-standing technical challenges in small target detection in remote sensing images. This framework effectively alleviates problems such as semantic degradation, feature suppression, and insufficient structural representation by integrating four complementary modules: CDAM reconstructs high- and low-frequency components through affine guided decomposition and context-aware fusion, enhancing the discriminative representation of weakly textured regions; DPFM establishes a dual-domain perception mechanism, adaptively balancing local detail cues and global semantic dependencies to ensure scale-consistent response under aggressive downsampling conditions; CMKB improves structural adaptability and fine-grained feature integrity through collaborative multi-kernel perception and refined channel interaction; and FSDetector operates based on high-resolution shallow features, effectively recovering spatial details lost during downsampling and significantly improving the accuracy of small target localization. Extensive experiments on multiple remote sensing image benchmark datasets demonstrate that CCAPNet significantly improves detection performance, particularly in challenging scenarios such as low contrast, scale variations, and densely distributed small targets. Generalization studies further confirm that CDAM, DPFM, and CMKB exhibit strong adaptability across different detection frameworks, demonstrating their plug-and-play flexibility. Although integrating multiple adaptive modules introduces additional computational costs and increases optimization complexity, future work will focus on improving model efficiency and enhancing robustness in practical applications.

[0149] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A satellite remote sensing small target detection method based on contextual affine perception, characterized in that, Includes the following steps: STEP1: Image initialization, including image input, downsampling, and initial feature extraction; STEP2: Feature filtering, using the features obtained in STEP1 to perform feature filtering, which is implemented through the Collaborative Multi-Core Module (CMKB); STEP3: Feature Response Enhancement. The feature responses selected in STEP2 are further enhanced based on the Dual Domain Perception Fusion (DPFM) module. STEP4: Upsampling stage, used to restore part of the spatial resolution; STEP5: Perform feature affine transformation, remodeling, and further upsampling on the upsampled features, and introduce a context-aware dynamic affine transformation module (CDAM). STEP6: Input the high-resolution features obtained in STEP5 into the fine-grained small object detector (FSDetector). STEP7: Perform feature affine transformation, remodeling, and resampling on the features; STEP8: Input the features at different scales obtained in STEP7 into the detector respectively, and achieve joint prediction of targets at different scales through multi-scale parallel detection; STEP9: The output of the fine-grained small target detector in STEP6 is fused with the output of the detector in STEP8 to obtain the final target detection result of the remote sensing image.

2. The method according to claim 1, characterized in that, In STEP2, the collaborative multi-core module establishes a collaborative path mechanism consisting of dynamic multi-core selection, local optimization, and channel attention; in STEP3, the dual-domain perception fusion module introduces a learnable scale prediction mechanism and a multi-level cross-domain fusion strategy in the spatial domain to achieve a dynamic balance between local details and global semantic features.

3. The method according to claim 1, characterized in that, In STEP5, the context-aware dynamic affine module uses frequency-inspired affine transformation to decompose the input feature map into structure-dominated low-frequency components and detail-dominated high-frequency components. It adopts a hierarchical structure consisting of three sub-modules: Structure-Detail Optimization Module (SDRM), Context Adaptive Enhancement Block (CAEB), and Hierarchical Bottleneck Integrator (HBI).

4. The method according to claim 1, characterized in that, In STEP6, the small target detector works in collaboration with upstream CDAM and DPFM to form a progressive optimization process. CDAM decouples and enhances structure-detail representation, DPFM integrates cross-domain semantic cues, and FSDetector optimizes fine-grained spatial features and detection response.

5. The method according to claim 1, characterized in that, In STEP1, the input image is scaled by continuous convolutional downsampling operations. While reducing the spatial resolution, basic visual features are extracted to obtain low-level features such as edges, textures and local structures from the original image, and to provide initial feature representations for subsequent semantic modeling.

6. The method according to claim 1, characterized in that, In STEP2, a dynamic feature reweighting mechanism based on multiple receptive fields is introduced to adaptively adjust the feature responses at different spatial locations and channel dimensions, thereby highlighting effective features related to the target and suppressing redundant or unstable responses in the background region.

7. The method according to claim 1, characterized in that, In STEP3, multi-scale convolutional perception and local refinement operations are introduced to enhance the structural features and detailed information within the target region, while maintaining the correlation between the local and neighborhood contexts.

8. The method according to claim 1, characterized in that, In STEP4, the enhanced features are upsampled to gradually restore the feature map to a higher resolution in the spatial dimension, thus creating conditions for subsequent fine feature modeling and small target detection.

9. The method according to claim 1, characterized in that, In STEP5, the structure-related information and detail-related information of the features are separated and reconstructed through affine mapping operations of the feature domain, thereby suppressing the attenuation of semantic information during feature transmission; through collaborative modeling of multi-scale contextual information, shallow detail features can be more effectively transmitted to the higher semantic layer.

10. The method according to claim 1, characterized in that, In STEP6, by reducing the effective receptive field and enhancing the local response, the detector can focus more precisely on the small target area, thereby improving the positioning accuracy and detection reliability of small targets.

Images