Remote sensing target detection methods

Abstract

This invention provides a remote sensing target detection method. The method includes segmenting a remote sensing image into patches and generating location codes to obtain a patch sequence; inputting the patch sequence into a Transformer encoder to extract global semantic features to obtain a feature map; inputting the feature map into a synaptic fractal pruning module to remove redundant channels and retain key features to obtain a compressed feature map; inputting the compressed feature map into a multi-resolution manifold router to generate a feature pyramid containing multi-scale geometric information; and inputting the feature pyramid into a detection head to predict the target category and location to obtain the final detected target. This invention, combining synaptic fractal pruning and a multi-resolution manifold router, can significantly restore and enhance multi-scale representation capabilities without incurring additional computational overhead. While meeting the real-time requirements of edge deployment, it provides significant performance improvements for the detection of small remote sensing targets and complex structures.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology, and in particular relates to a remote sensing target detection method. Background Technology

[0002] Remote sensing target detection refers to the automatic or semi-automatic identification, localization, and extraction of targets of interest from data acquired from satellites, UAVs, airborne platforms, or ground sensors. Such images are characterized by significant scale variations, complex backgrounds, and diverse target morphologies. Small objects typically exhibit high background contrast, sharp and regular edges, obvious clustering, and sensitivity to occlusion; for example, the stark color difference between a swimming pool and a rooftop, or between rooftops, and the often neatly arranged, well-defined outlines of vehicles obscured by tree shadows. In contrast, large objects occupy a wider spatial range, contain richer textures and prominent shapes, such as clearly defined gates and symmetrical layouts in airport facilities, which facilitate multi-level semantic and contextual reasoning. The significant differences in texture, structure, scale, and spatial distribution between large and small targets place stringent demands on detection models, particularly in terms of resolution, contextual modeling, and robustness to occlusion and noise.

[0003] To address these challenges, Remote Sensing Foundation Models (RSFMs) have become a pillar of intelligent remote sensing through large-scale pre-training and cross-task adaptation, providing robust feature representations and portability. However, the improvement in accuracy often comes at the cost of heavy computational overhead; during training and inference, deeper architectures, large parameters, and complex multi-scale fusions lead to significant memory consumption, computational costs, and deployment burdens.

[0004] The current RSFM still has several key limitations:

[0005] (1) Pressure to capture features of interest: The model needs to maintain high-resolution perception and global context modeling in ultra-high-resolution feature maps to accurately capture regions of interest. This forces the base model to maintain a high configuration in terms of feature pyramid depth, channel bandwidth and self-focus range, which significantly increases parameter size, memory footprint and inference algorithm.

[0006] (2) The return on low precision for efficiency: The performance gains from multi-scale modeling, global attention and multimodal fusion come from a large amount of redundant computation, resulting in a moderate overall precision-efficiency ratio.

[0007] Faced with a heavy computational burden, the industry is reconsidering lightweight solutions: pruning, knowledge distillation, low-bit quantization, and efficient network design. Channel pruning has matured into an adaptive multi-resolution method that can achieve multiple parameter reductions; distillation consistently shifts power from teachers with too many parameters to students with fewer; quantization compresses weights to 8 bits or less, significantly reducing storage space while maintaining accuracy; combined with minimalist backbones and shared parameter schemes, these methods are guiding remote sensing foundation models toward higher operational efficiency.

[0008] Compression gains are typically achieved by simplifying the architecture and reducing parameters; however, these steps weaken the network's ability to perceive fine-grained or local cues, making it easy to miss small, boundary-oriented, and partially hidden targets. Because such cues have weak activations and complex textures, their loss directly reduces recognition accuracy. Furthermore, the significant scale variability in remote sensing images necessitates robust multi-scale modeling; methods like eliminating feature pyramids, inter-layer fusion, or global concerns disrupt this mechanism and distort performance at different scales.

[0009] Traditional remote sensing basic models rely on a large number of parameters and high inference budgets, but false positives and false negatives still exist in complex contexts.

[0010] Existing remote sensing baseline models are difficult to deploy on edge devices due to their high computational and storage requirements. While traditional lightweight techniques help reduce model size, they often compromise the ability to capture fine-grained features, leading to missed small and boundary targets. Furthermore, the degradation of multi-scale modeling further limits their robustness to scale variations. Summary of the Invention

[0011] The purpose of this invention is to address the shortcomings of existing technologies by providing a remote sensing target detection method that reduces model size, computation, and memory requirements while retaining key features, thereby improving the accuracy and precision of target detection.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0013] A remote sensing target detection method includes the following steps:

[0014] S1. Acquire remote sensing images;

[0015] S2. Divide the remote sensing image into patches and generate location codes to obtain a patch sequence;

[0016] S3. Input the patch sequence into the Transformer encoder to extract global semantic features and obtain a feature map;

[0017] S4. Input the feature map into the synaptic fractal pruning module to remove redundant channels and retain key features to obtain a compressed feature map;

[0018] S5. Input the compressed feature map into a multi-resolution manifold router to generate a feature pyramid containing multi-scale geometric information;

[0019] S6. Input the feature pyramid into the detection head to predict the target category and location, and obtain the final detected target;

[0020] The synaptic fractal pruning module comprises a fusion layer, an interleaved computation layer, a differential computation layer, a first normalization layer, a first gating layer, a first linear layer, a GELU activation layer, a second linear layer, and a residual injection layer connected in sequence. The left branch nerve layer and the right branch nerve layer are respectively connected to the fusion layer, and the interleaved computation layer is connected to the first linear layer. The Transformer encoder is connected to the left branch nerve layer, the right branch nerve layer, and the residual injection layer.

[0021] The multi-resolution manifold router includes a multi-resolution feature downsampling layer, a second gate layer, a first convolutional layer, a second convolutional layer, a probability-weighted fusion layer, and a cross-scale feature fusion layer connected in sequence. The multi-resolution feature downsampling layer is connected to a Laplacian geometric embedding layer. The Laplacian geometric embedding layer and the second gate layer are respectively connected to a query generation layer. The query generation layer and the Laplacian geometric embedding layer are respectively connected to a softmax activation layer. The softmax activation layer is connected to a second normalization layer. The second normalization layer is connected to a probability-weighted fusion layer.

[0022] The residual injection layer is connected to the multi-resolution feature downsampling layer.

[0023] This invention combines synaptic fractal pruning with a multi-resolution manifold router, integrating fractal geometry and manifold curvature principles into the feature learning process. Synaptic fractal pruning identifies redundant connections through differences between left and right neural branches and dynamically compresses the model by applying recursive pruning with differentiable gating, significantly reducing FLOPs and memory usage in inefficient channels while preserving key target detection. The multi-resolution manifold router constructs a multi-scale pyramid based on the compressed features from synaptic fractal pruning, establishes geometry awareness through Laplacian embedding, and guides labels to lightweight experts based on resolution and geometry adaptability. Curvature-aware sparse expert routes are deployed on the Laplacian embedding manifold to recover the multi-scale representations lost during compression, and multi-scale outputs are fused to enhance contextual representation and inference efficiency, achieving efficient and robust multi-scale feature modeling.

[0024] This invention combines synaptic fractal pruning with a multi-resolution manifold router to significantly recover and enhance multi-scale representation capabilities without incurring additional computational overhead. While meeting the real-time requirements of edge deployment, it provides significant performance improvements for remote sensing detection of small targets and complex structures. The collaboration between synaptic fractal pruning and the multi-resolution manifold router allows the model to maintain a complete semantic pyramid even after compression, while generating highly concentrated feature responses in regions with weak texture or small targets.

[0025] Furthermore, the specific process of S4 includes:

[0026] The feature maps are input from the left and right branch nerve layers. After copying the feature maps to the left and right nerve branches, the left and right nerve branches are projected onto the feature maps, and the responses of the left and right nerve branches are connected row by row to generate a dual-path stacked feature matrix.

[0027] The interleaved features are obtained by multiplying the interleaved sequence matrix from bottom to top by the dual-path stacked feature matrix;

[0028] By performing row-by-row differencing on the alternating features, we obtain the difference operator;

[0029] The difference operator is normalized to generate normalized difference features for each connection or channel;

[0030] The normalized difference features of each connection or channel are input into the first gating layer to obtain the retention probability of each connection or channel;

[0031] Based on the retention probability of each connection or channel, determine whether each connection or channel is retained or pruned. If the retention probability is greater than a threshold, the connection or channel is retained; if the retention probability is less than or equal to the threshold, the connection or channel is pruned.

[0032] The features of the retained connections or channels are fused into a diagonal mask matrix;

[0033] The diagonal mask matrix is ​​processed by the first linear layer, the GELU activation layer, and the second linear layer to obtain the updated feature vector;

[0034] The updated feature vector is input into the residual injection layer and fused with the feature map to obtain a compressed feature map.

[0035] Synaptic fractal pruning precisely removes redundant connections using fractal patterns through left-right branch differences and differentiable gated cyclic hard pruning. While preserving fine-grained information of the synaptic layer, it significantly reduces FLOPs and memory usage of inefficient channels.

[0036] Furthermore, the expression for the dual-path stacked feature matrix H is as follows:

[0037] ;

[0038] Among them, W L W represents the weight matrix of the left branch projection. R Let X represent the weight matrix of the right branch projection, and X be the feature map.

[0039] Furthermore, the interleaved feature X int The expression is as follows:

[0040] ;

[0041] ;

[0042] Where J represents the interleaved sequence matrix, I d Represents the identity matrix.

[0043] The interleaved feature simulates the motion of the left and right sensory axes, and the cross-processed signals can be directly compared in the same coordinate system.

[0044] Furthermore, the expression for the retention probability of each connection or channel is as follows:

[0045] ;

[0046] Among them, W g and b g The weight vector and bias, h, are used to calculate the gating score, respectively. i Let g represent the normalized difference feature of the i-th connection or channel, σ(·) represent the Sigmoid activation function, and g i G represents the retention probability of the i-th connection or channel. i ∈(0,1).

[0047] Furthermore, the expression for the diagonal mask matrix M is as follows:

[0048] ;

[0049] ;

[0050] Where, m i This indicates that the i-th connection or channel is reserved, 1 {·} The function is an indicator function; it outputs 1 if the condition is met and 0 if the condition is not met. τ is the threshold. diag() means converting the vector inside the brackets into a diagonal mask matrix M.

[0051] A diagonal mask matrix dynamically controls which connections are retained in the computation and which are deleted, thus achieving a pruning effect. This strategy sparsifies the computation graph during each forward pass, effectively reducing the consumption of computational resources.

[0052] Furthermore, compress the feature map The expression is as follows:

[0053] ;

[0054] ;

[0055] Where φ(·) represents the GELU nonlinear activation function, W h and W o These are the weight matrices of the first and second linear layers, respectively; Z is the updated feature vector; ⊙ denotes element-wise multiplication; α represents the residual scaling factor; M is the diagonal mask matrix; X... int It is an interlaced feature.

[0056] Furthermore, the specific process of S5 includes:

[0057] The compressed feature map is input into the multi-resolution feature downsampling layer, and downsampling is performed at each resolution or scale to generate a feature tensor for each resolution or scale.

[0058] The feature tensor of each resolution or scale is passed through the second gating layer to obtain the gated feature tensor of each resolution or scale;

[0059] The gated feature tensors at each resolution or scale are subjected to nonlinear transformation and regularization through the first and second convolutional layers in sequence to obtain the output features of each expert at each resolution or scale.

[0060] The feature tensor at each resolution or scale extracts low-dimensional coordinate information through a Laplacian geometric embedding layer to obtain the manifold coordinate tensor at each resolution or scale.

[0061] Input the manifold coordinate tensor and the gated feature tensor of each resolution or scale into the query generation layer to generate a query vector for each resolution or scale that integrates geometric and texture information.

[0062] The query vector at each resolution or scale is processed by softmax activation to obtain the route for each expert at each resolution or scale;

[0063] Normalize the routes of each expert at each resolution or scale to obtain the selection weights of each expert at each resolution or scale.

[0064] Based on the selection weights of each expert at each resolution or scale, the output features of all experts at each resolution or scale are weighted and fused to obtain the mixed features at each resolution or scale.

[0065] By fusing mixed features from all resolutions or scales, a feature pyramid containing multi-scale geometric information is generated.

[0066] The multi-scale manifold router constructs a multi-scale pyramid based on synaptic fractal pruning compression features, extracts curvature through Laplacian embedding, and guides labels to lightweight experts based on resolution and geometric adaptability. It also employs two-layer regularization of scale and expert to prevent collapse and fuses multi-scale outputs to enhance contextual representation and inference efficiency. The multi-scale manifold router uses curvature-driven sparse routing on the Laplacian manifold to dynamically allocate residual labels to the most suitable scale and geometric subspace, recovering the multi-scale semantic information lost after deep compression.

[0067] Furthermore, the expression for the gated feature tensor at each resolution or scale is as follows:

[0068] ;

[0069] ;

[0070] ;

[0071] Where σ(·) represents the Sigmoid activation function, v represents the gating vector, r represents the channel retention factor, and ⊙ represents element-wise multiplication. I represents the gated feature tensor at the s-th resolution or scale; I represents the compressed feature map, s represents the resolution or scale, and r represents the gated feature tensor at the s-th resolution or scale. s Pool is the downsampling ratio for the s-th resolution or scale. rs (·) denotes a convolution or mixed pooling operator, U (s) Let be the feature tensor of the s-th resolution or scale.

[0072] Energy efficiency can be achieved by simulating neuronal inhibition mechanisms by activating only the channels that are valuable for the task.

[0073] Furthermore, the expression for the query vector at each resolution or scale is as follows:

[0074] ;

[0075] ;

[0076] Where ψ(·) represents an alternating linear function, Q (s) Let C represent the query vector at the s-th resolution or scale, and GeomEmbed(·) represent the Laplacian eigenmap or diffusion map operator. (s) Represents the manifold coordinate tensor at the s-th resolution or scale.

[0077] Furthermore, the expression for the selection weights of each expert at each resolution or scale is as follows:

[0078] ;

[0079] ;

[0080] Among them, W k Let the gating weight vector of the k-th expert be represented. λ represents the Softmax temperature, λ represents the curvature bias coefficient, and k represents the Softmax temperature. (s) This represents the average manifold curvature at the s-th resolution or scale. This represents the route taken by the k-th expert at the s-th resolution or scale. This represents the route taken by the j-th expert at the s-th resolution or scale, where K is the total number of experts, and a k (s) ∈(0,1) represents the selection weight of the k-th expert at the s-th resolution or scale, Q (s) This represents the query vector at the s-th resolution or scale.

[0081] Furthermore, the expression for the blending features at each resolution or scale is as follows:

[0082] ;

[0083] ;

[0084] ;

[0085] Where φ(·) is the GELU activation function, W k (1) b k (1) H represents the weight matrix and bias vector of the first convolutional layer of the k-th expert, respectively. k (s) E represents the implicit representation of the k-th expert at the s-th resolution or scale; k (s) W represents the output feature of the k-th expert at the s-th resolution or scale. k (2) b k (2) Let be the weight matrix and bias vector of the second convolutional layer for the k-th expert, respectively. Drop(·) represents the regularization operation, and ⊙ represents element-wise multiplication. This represents the mixed feature at the s-th resolution or scale.

[0086] Furthermore, the expression for the feature pyramid is as follows:

[0087] ;

[0088] ;

[0089] Where O represents the feature pyramid, Up(·) represents the spatial size of the nearest neighbor interpolation or transpose convolution upsampling, and ρ s β is the weighting coefficient for the s-th resolution or scale. s Let ρ be the softmax normalized weight for the s-th resolution or scale, where S represents the total number of resolution layers or scales, and ρ is the weight for the s-th resolution or scale. t The exponent denominator is used for Softmax normalization, and t represents the resolution or scale index.

[0090] By integrating across scales, we can achieve integrated reasoning from coarse to fine and from local to global, in order to recover the semantic and geometric details of the target.

[0091] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0092] This invention employs synaptic fractal pruning, identifying redundant connections through gradient difference matrices and dynamically compressing the model by applying recursive pruning with differentiable gating. This method significantly reduces FLOPs and memory usage while preserving key object detection, achieving efficient edge deployment.

[0093] To alleviate the weak multi-scale representation problem caused by large-scale pruning in deep compression models, and the expert mismatch and collapse problems caused by the lack of geometric awareness and single-scale operation in traditional gated MoE (hybrid expert model), this invention adopts a multi-resolution stream-form router, which is analogous to an agent deployed in a manifold space of different resolutions to cooperate. It uses Laplacian embedding to establish geometric awareness, and combines sparse routing and expert selection to achieve efficient and robust multi-scale feature modeling.

[0094] This invention combines synaptic fractal pruning with a multi-resolution manifold router to significantly restore and enhance multi-scale representation capabilities without incurring additional computational overhead. While meeting the real-time requirements of edge deployment, it provides significant performance improvements for remote sensing detection of small targets and complex structures. Attached Figure Description

[0095] Figure 1 This is a comparison chart of the detection performance of different remote sensing target detection methods in this invention on the DOTA dataset;

[0096] Figure 2 This is a schematic diagram of the LiteFM-RSDet model structure according to an embodiment of the present invention;

[0097] Figure 3 This is a comparison graph of LiteFM-RSDet and LSKNet-S on the DOTAv1.0 dataset under multiple scales according to an embodiment of the present invention.

[0098] Figure 4This is a diagram showing the detection results of the method in this embodiment of the invention on the DOTA-v1.0 dataset;

[0099] Figure 5 The detection results and heatmap of a simple DOTA scene according to an embodiment of the present invention;

[0100] Figure 6 The above are the detection results and heatmaps of complex DOTA scenes according to an embodiment of the present invention. Detailed Implementation

[0101] The present invention will be described in detail below with reference to embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0102] Example 1

[0103] 1. Remote Sensing Basic Model

[0104] With the advancement of remote sensing technology, Remote Sensing Foundation Models (RSFMs) have significantly enhanced tasks such as scene classification and object detection through their powerful feature representation capabilities. These models have driven progress in practical applications such as architecture design, training strategies, and disaster monitoring, becoming key tools in the field. Based on their methodology and application focus, RSFMs can be categorized into transformer-based, target analysis-based, multimodal, self-supervised, and spatiotemporal models, each with its own advantages and challenges.

[0105] Transformer-based models, especially ViT and Swin Transformer, are widely used in remote sensing image classification and target detection due to the advantages of their self-attention mechanism in capturing long-range dependencies. ViT is suitable for large-scale analysis by dividing the image into blocks, while Swin Transformer reduces computational cost and improves efficiency for high-resolution tasks by using a local window mechanism; however, both require significant computational resources.

[0106] Models such as Fast R-CNN, Mask R-CNN, and RetinaNet perform excellently in remote sensing target detection. RPN and CNN effectively support multi-target recognition, with RetinaNet mitigating class imbalance through focus loss. However, they suffer from scalability and real-time bottlenecks in large-scale image or long sequence processing. To improve semantic understanding, multimodal models such as SkySense and RemoteCLIP fuse multi-source data, improving accuracy in complex scenes, but still face challenges related to modality alignment and data scarcity.

[0107] Beyond architectural innovation and multimodal fusion, self-supervised learning models (such as RingMoand SpectralGPT) leverage large amounts of unlabeled data to improve performance in object detection and change detection. While these models reduce reliance on labeled data, they face challenges in ensuring the robustness and representativeness of learned features, especially in complex remote sensing scenarios.

[0108] In challenging remote sensing tasks, these existing methods have limitations in terms of scalability, non-operational aspects, and feature representation. In contrast, LiteFM-RSDet in this embodiment integrates synaptic fractal pruning (SFP) and multi-resolution manifold router (MMR), significantly improving model efficiency and feature representation by leveraging cross-scale pruning and geometric awareness, effectively addressing these shortcomings.

[0109] 2. Lightweight detection

[0110] Current lightweighting strategies primarily focus on structural compression design, token pruning mechanisms, dynamic computation strategies, YOLO series optimizations, and large model compression. The core objective is to minimize computational and storage costs while maintaining sufficient model accuracy, thereby improving the deployment efficiency of models on edge devices and resource-constrained environments. Although these technical approaches emphasize different aspects, they exhibit a general trend of complementarity and synergistic development.

[0111] In terms of structural compression, researchers focus on reducing computational and storage requirements by optimizing network structure, replacing convolutional operators, and reconstructing model parameters. Models such as EfficientViT, XFeat, and Lite-HRNet optimize computational and storage requirements by changing network structure and replacing traditional convolutional operators. These methods improve inference speed and accuracy while reducing computational cost across various resolutions. MobileNetV3 and GhostNet further reduce parameters and improve computational efficiency, especially for embedded or mobile environments.

[0112] These structural compression methods are mainly applied to general-purpose networks, emphasizing module simplification and hardware compatibility. However, in task-specific detection scenarios such as object detection, lightweight strategies must be further customized to adapt to the nature of the task, often requiring targeted structural redesign and feature reconfiguration.

[0113] Building upon this, dynamic computation mechanisms aim to optimize inference efficiency by adjusting computational resources and paths during the inference process based on input features. SpAtten's cascaded token pruning and attention avatar pruning employ a progressive quantization strategy to dynamically prune tokens and avatars based on importance. This approach is well-suited for resource-constrained devices, enhancing the adaptability and responsiveness of inference across different hardware platforms.

[0114] In large-scale model compression, collaborative optimization strategies have been proposed to mitigate inference latency, memory overhead, and deployment challenges. For example, CHEX employs periodic channel pruning and regeneration to create high-performance subnets without pre-training, which is beneficial for classification and detection tasks. Dual compression reduces memory usage by 40% in large models like OPT with minimal accuracy loss.

[0115] The teacher-student paradigm shifts semantic knowledge from large models to compact models, maintaining performance while reducing model size. These methods support the transition from massive models to efficient, deployable solutions, facilitating real-time and edge deployment of large language and multimodal models.

[0116] However, these methods often struggle to balance real-time performance on edge devices (drones, autonomous vehicles, edge IoT cameras, industrial edge computing gateways such as Advantech and HPE Edgeline). Transformer-based models, while effective, are computationally expensive. Target analysis models excel at small target detection but face scalability issues in complex scenarios. Multimodal and self-supervised models improve data utilization but lack robustness to fine-grained targets. Traditional pruning and quantization methods also degrade model performance, especially in high-resolution remote sensing tasks. To address these challenges, this embodiment introduces Synaptic Fractal Pruning (SFP), which reduces computation and memory usage while preserving key features.

[0117] 3. LiteFM-RSDet

[0118] 1) Baseline: MTP

[0119] Multi-task Pretraining (MTP), employing a Vision Transformer backbone and utilizing joint task supervision, significantly improves the representation fidelity and downstream generalization ability of remote sensing images, serving as a fundamental architecture in this field. Structurally, MTP retains the standard Transformer layout, including an input embedding module, a backbone, and a set of task-specific headers. Unlike standard convolutional channels, MTP uses a Vision Transformer patch embedding technique to slice high-resolution remote sensing images into uniform patches, submitting the generated sequences to a Transformer encoder. Each patch has a positional encoding to preserve spatial information. This processing method significantly expands the global receptive field of the self-attention mechanism, effectively addressing extreme scale variations present in remote sensing scenes.

[0120] The Transformer encoder consists of a multi-head self-attention layer, a residual layer, and a layer normalization layer connected in sequence.

[0121] Within the backbone network, stacked Transformer encoders employ multi-head self-attention techniques to capture global semantics in remote sensing images. Residual connections and layer normalization alleviate gradient degradation in deep networks, thereby improving stability and convergence speed under multi-task supervision.

[0122] Real-world remote sensing tasks often face challenges such as extremely high image resolution, significant target scale variations, and complex background interference. Furthermore, remote sensing images not only contain numerous invalid regions but also often exhibit complex structures with dense targets, small scale, and weak texture. This necessitates models that maintain accuracy while possessing excellent fine-grained perception and multi-scale modeling capabilities. MTP, employing a Transformer architecture, features a global self-attention mechanism and task-coordinated pre-training capabilities, enabling it to more fully preserve the long-range dependencies and semantic consistency of remote sensing images. During compression, MTP's modular design facilitates the introduction of fine-grained structural optimizations based on tokens, channels, and paths, contributing to lightweight and high-fidelity remote sensing image understanding. It also provides a natural advantage and strong support for remote sensing model compression.

[0123] Therefore, this embodiment selects MTP as the structural baseline of the remote sensing base model, considering both its versatility in multi-task modeling and its ability to extract long-range correlation features from high-resolution remote sensing images. MTP has good generalization ability in remote sensing tasks, especially suitable for multi-task collaborative scenarios such as classification, detection, and change detection. However, MTP is based on a large-scale transformer architecture, which has high computational and memory overhead, and suffers from token redundancy and insufficient inter-layer information flow, making it prone to semantic loss during compression. If the structure can be optimized by combining the redundancy and multi-scale features of remote sensing images, the compression efficiency and performance will be significantly improved.

[0124] Based on this, this embodiment proposes a joint optimization lightweight scheme LiteFM-RSDet based on MTP architecture: a novel architecture that combines synaptic fractal pruning (SFP) with multi-scale manifold routers (MMR), such as... Figure 2 As shown.

[0125] 2) Synaptic Fractal Pruning (SFP)

[0126] As the resolution of remote sensing images increases, the amount of texture and structural detail per frame is also exploding. Traditional remote sensing base models respond by stacking deeper convolutional Transformer stages and introducing hundreds of millions of parameters, but this significantly increases inference costs and memory footprint, making real-time deployment on edge devices impractical. While channel pruning and low bit quantization can suppress FLOPs (floating-point operations per second), they often weaken the network's ability to perceive critical targets such as vehicles. Therefore, achieving significant computational savings while maintaining or even enhancing attention to these salient features has become a core challenge for next-generation remote sensing models.

[0127] Synaptic Fractal Pruning (SFP) aims to prune and compress the underlying remote sensing model while minimizing accuracy loss. Inspired by biological synaptic fractal pruning, this module prioritizes high gradient differences and applies recursive cross-scale pruning. SFP first uses a differential matrix to accurately identify redundant connections, then performs multi-level hard pruning using differentiable gating probabilities. The core objective is to significantly reduce FLOPs and memory usage during inference while preserving key feature representations to the greatest extent possible.

[0128] The synaptic fractal pruning module includes a fusion layer, an interleaved computation layer, a differential computation layer, a first normalization layer, a first gating layer (micro-gating layer), a first linear layer, a GELU activation layer, a second linear layer, and a residual injection layer connected in sequence. The interleaved computation layer is connected to the first linear layer. The left branch nerve layer and the right branch nerve layer are both connected to the fusion layer. The Transformer encoder is connected to the left branch nerve layer, the right branch nerve layer, and the residual injection layer, respectively.

[0129] The feature map output by the Transformer encoder is input into the left and right branch nerve layers. After copying the feature map into the left and right nerve branches, the left and right nerve branches are projected onto the feature map, and the responses of the left and right nerve branches are connected row by row to generate a dual-path stacked feature matrix.

[0130] The interleaved features are obtained by multiplying the interleaved sequence matrix from bottom to top by the dual-path stacked feature matrix;

[0131] By performing row-by-row differencing on the alternating features, we obtain the difference operator;

[0132] The difference operator is normalized to generate normalized difference features for each connection or channel;

[0133] The normalized difference features of each connection or channel are input into the first gating layer to obtain the retention probability of each connection or channel;

[0134] Based on the retention probability of each connection or channel, determine whether each connection or channel is retained or pruned. If the retention probability is greater than a threshold, the connection or channel is retained; if the retention probability is less than or equal to the threshold, the connection or channel is pruned.

[0135] The features of the retained connections or channels are fused into a diagonal mask matrix;

[0136] The diagonal mask matrix is ​​processed by the first linear layer, the GELU activation layer, and the second linear layer to obtain the updated feature vector;

[0137] The updated feature vector is input into the residual injection layer and fused with the feature map output by the Transformer encoder to obtain a compressed feature map.

[0138] The specific process of synaptic fractal pruning (SFP) is as follows:

[0139] Left and right branch neural layers: Vertebrates rely on bilaterally symmetrical sensors, such as two eyes, two ears, and paired antennae, to perceive environmental differences. Inspired by this principle, the feature map X obtained after processing by the Transformer encoder is copied into the left and right neural branches; then, different left and right projections are applied to the feature map X, and their responses are connected row by row to generate a dual-path stacked feature matrix H as a fractal representation:

[0140]

[0141] Among them, W L W represents the weight matrix of the left branch projection. R The weight matrix represents the projection of the right branch, and the dual-path stacked feature matrix H represents the concatenation result of the left and right branch features. This concatenation serves as the initial feature carrier for the entire branching process.

[0142] Interleaved computation layer: Using the interleaved sequence matrix J, multiply it from bottom to top by the dual-path stacked feature matrix H to generate interleaved features X. int This simulates the motion of the left and right sensory axes, and the signals processed by their crossover can be directly compared within the same coordinate system.

[0143]

[0144]

[0145] Among them, I d Represents the identity matrix.

[0146] Differential computation layer: Perform row-by-row differencing on the interleaved features to obtain the difference operator ∆:

[0147]

[0148] Here, ∆ describes the difference in contribution of each branch to the left and right signals, and its magnitude will be used later to measure the importance of the connection. N represents the total number of connections or channels. Indicates from the interlaced feature X int Take all columns from row 1 to row N. Indicates from the interlaced feature X int Take all columns from row N+1 to row 2N.

[0149] First normalization layer: To avoid numerical instability caused by extreme differences, layer normalization is introduced:

[0150]

[0151] Where LayerNorm() represents the layer normalization function.

[0152] This produces a feature h with zero mean and zero unit variance, which facilitates gradient propagation.

[0153] First gating layer (micro-gating layer): Feature h is inserted into a micro-gating layer and scored using the Sigmoid activation function.

[0154]

[0155]

[0156] Among them, W g and b g Let h be the weight vector and bias used to calculate the gated score, respectively, and be a learnable parameter. i G represents the normalized difference feature of the i-th connection or channel. i G represents the retention probability (gating score) of the i-th connection or channel. i ∈(0,1), N represents the total number of connections or channels, and σ(·) represents the Sigmoid activation function.

[0157] Mimicking biological action potential thresholds, neurons either generate impulses (1) or remain silent (0) for hard threshold pruning, and the probabilities are discretized during inference.

[0158]

[0159]

[0160] Where, m i This indicates that the i-th connection or channel is reserved (active): 1 indicates reserved, 0 indicates pruned. {·}τ is an indicator function that outputs 1 if the condition is met and 0 if the condition is not met; τ is the threshold; and diag() indicates that the vector inside the brackets is converted into a diagonal mask matrix M.

[0161] By m i The resulting diagonal mask matrix M dynamically controls which connections are retained in the computation and which are deleted, thus achieving a pruning effect. This strategy sparsifies the computation graph during each forward pass, effectively reducing the consumption of computational resources.

[0162] Preserve the feedforward path and perform a two-layer MLP only on the active features (the preserved features). The two-layer MLP consists of a first linear layer, a GELU activation layer, and a second linear layer.

[0163]

[0164] Where φ(·) represents the GELU nonlinear activation function, W h and W o These are the weight matrices for the first and second linear layers, respectively, and Z is the updated feature vector.

[0165] Residual Injection Layer: The updated feature vector Z is injected into the main branch using a mask that controls the residuals, stabilizing accuracy and maintaining smooth gradients.

[0166]

[0167] Here, X is the feature map input to the synaptic fractal pruning module, ⊙ represents element-wise multiplication, and α is a residual scaling factor that adjusts the amplitude of the injected signal (feature vector Z) to prevent excessive perturbation. Feature Fusion (Compressed feature maps) carry the newly injected information to subsequent network layers.

[0168] The SFP module algorithm is as follows:

[0169] Input: X is the input feature map; W L W is the weight matrix of the left branch projection; R W is the weight matrix of the right branch projection; g is the weight vector used to calculate the gated score; bg is the bias used to calculate the gated score; W h W is the weight matrix of the first linear layer in an MLP. o is the weight matrix of the second linear layer in the MLP; J is the predefined interleaved sequence matrix used for feature interleaving; τ is the threshold used for activation pruning mask; α is the residual scaling factor for residual feature fusion.

[0170] Output: It is a compressed feature map after pruning and residual enhancement.

[0171] 1: Using W L and W R The feature map X is divided into two branches, and then they are vertically connected to form a dual-path stacked feature matrix H; 2: An interleaved sequence matrix J is applied to interweave the two branches to obtain interleaved features X. int 3: For the interlaced feature X int Perform row-wise differencing to obtain the difference operator ∆; 4: Normalize the difference operator ∆ to obtain the feature h with zero mean and zero unit variance; 5: For each normalized difference feature h... i 6: Calculate the gated fraction g using the sigmoid function over a linear transformation. i 7: End; 8: For each gated score g i ;9: If g i If the threshold τ is exceeded, then m is set. i =1, otherwise set m i =0; 10: End; 11: From m i 12: Form a diagonal mask matrix M; 13: Input the diagonal mask matrix M into the two-layer MLP to obtain the updated feature vector Z; 14: The updated feature vector Z is fused with the feature map X into a compressed feature map by using a mask injection of the control residual into the main branch. .

[0172] 3) Multi-resolution manifold router (MMR)

[0173] In deep compression models, aggressive pruning of structure and parameters significantly weakens the model's ability to represent multiple scales, especially affecting the recall of small targets and regions. While traditional gated MoE (hybrid expert model) has a dynamic expert selection mechanism, it operates only at a single scale, lacks geometric awareness, and cannot effectively compensate for the scale information loss caused by pruning. This often leads to the misassignment of high-curvature regions to unsuitable experts, thus wasting computational resources and accelerating expert failure.

[0174] To this end, the expert system is analogous to a multi-agent system (MAS), where each expert is considered an autonomous agent deployed on a manifold space with different resolutions to collaboratively perform tasks representing multi-scale and multi-curvature regions in remote sensing images. Inspired by MAS in robot control and reinforcement learning, this embodiment designs a novel multi-resolution manifold router (MMR) mechanism. This mechanism utilizes Laplacian embedding to establish geometric perception and combines sparse routing and expert selection to achieve efficient and robust multi-scale feature modeling.

[0175] The multi-resolution manifold router comprises, in sequence, a multi-resolution feature downsampling layer, a second gating layer (channel micro-gating layer), a first convolutional layer (expert hidden layer), a second convolutional layer (expert output layer), a probability-weighted fusion layer, and a cross-scale feature fusion layer. The multi-resolution feature downsampling layer is connected to a Laplacian geometric embedding layer. The Laplacian geometric embedding layer and the second gating layer are connected to a query generation layer. The query generation layer and the Laplacian geometric embedding layer are connected to a softmax activation layer. The softmax activation layer is connected to a second normalization layer. The second normalization layer is connected to the probability-weighted fusion layer. The residual injection layer is connected to the multi-resolution feature downsampling layer.

[0176] The compressed feature map obtained by the synaptic fractal pruning module is input into the multi-resolution feature downsampling layer, and downsampling is performed at each resolution or scale to generate a feature tensor for each resolution or scale.

[0177] The feature tensor at each resolution or scale is passed through a second gating layer to filter out channels that are valuable to the task, resulting in a gated feature tensor at each resolution or scale.

[0178] The gated feature tensors at each resolution or scale are subjected to nonlinear transformation and regularization through the first and second convolutional layers in sequence to obtain the expert features of each expert at each resolution or scale.

[0179] The feature tensor at each resolution or scale extracts low-dimensional coordinate information through a Laplacian geometric embedding layer to obtain the manifold coordinate tensor at each resolution or scale.

[0180] Input the manifold coordinate tensor and the gated feature tensor of each resolution or scale into the query generation layer to generate a query vector for each resolution or scale that integrates geometric and texture information.

[0181] The query vector at each resolution or scale is processed by softmax activation to obtain the route for each expert at each resolution or scale;

[0182] Normalize the routes of each expert at each resolution or scale to obtain the selection weights of each expert at each resolution or scale.

[0183] Based on the selection weights of each expert at each resolution or scale, the expert features of all experts at each resolution or scale are weighted and fused to obtain the mixed features at each resolution or scale.

[0184] The mixed features of all resolutions or scales are fused to obtain the final output features (feature pyramid containing multi-scale geometric information) after expert routing and multi-scale fusion.

[0185] The specific process of a multi-resolution manifold router (MMR) is as follows:

[0186] Multi-resolution feature downsampling layer: Each resolution level s corresponds to a sub-agent group, and its observed features are represented as follows:

[0187]

[0188] Where I represents the 4-D feature tensor (compressed feature map) generated by the backbone network. ), s represents resolution or scale, r s Pool is the downsampling ratio for the s-th resolution or scale. rs (·) is a learnable convolution or mixed pooling operator, U (s) Let be the feature tensor of the s-th resolution or scale.

[0189] Laplacian geometry embedding layer: By embedding geometric information through Laplacian or diffusion fusion mapping, low-dimensional fluid coordinate extraction can be achieved, endowing intelligent out-of-scale texture curvature and neighborhood perception capabilities.

[0190]

[0191] Where GeomEmbed(·) represents the Laplacian eigenmap or diffusion map operator, C (s) Represents the manifold coordinate tensor at the s-th resolution or scale.

[0192] Second gating layer (channel micro-gating layer): This achieves energy efficiency by mimicking neuronal inhibition mechanisms by activating only channels valuable to the task.

[0193]

[0194]

[0195] Where σ(·) represents the Sigmoid activation function, v is a learnable gating vector, r represents the channel retention factor, and ⊙ represents element-wise multiplication. It is the gated feature tensor of the s-th resolution or scale. This process is equivalent to an agent filtering external stimuli and retaining only high-value inputs.

[0196] Query generation layer: fuses texture features with geometric coordinates to generate query vectors that distinguish scale and curvature.

[0197]

[0198] Where ψ(·) is an alternating linear function, Q (s) It is a query vector with hidden dimension dq at the s-th resolution or scale.

[0199] Softmax activation layer: Each agent (expert) uses a query vector Q that fuses geometry and texture. (s) To assess your suitability for the task:

[0200]

[0201] Among them, W k Let the gating weight vector of the k-th expert be represented. λ is the Softmax temperature, λ is the curvature bias coefficient, and k is the Softmax temperature. (s) =‖∆manifoldC (s) ‖2 is the average manifold curvature at the s-th resolution or scale, and ∆manifold denotes the Laplace-Beltrami operator (differential operator) over the manifold. This represents the route taken by the k-th expert at the s-th resolution or scale. This process is similar to a multi-intelligence bidding mechanism, where each expert (agent) scores the task based on their abilities, and the system selects the most suitable executor.

[0202] Second normalization layer: Normalizes routes into probabilities, forming sparse expert selection weights:

[0203]

[0204] Where K is the total number of experts, a k (s) ∈(0,1) represents the selection weight of the k-th expert at the s-th resolution or scale. This represents the route taken by the j-th expert at the s-th resolution or scale.

[0205] Expert Hidden Layer: Performs non-linear transformations on gated features to prepare for capturing expert patterns.

[0206]

[0207] Where φ(·) is the GELU activation function, W k (1) b k (1) Let H be the weight matrix and bias vector of the first convolutional layer (expert hidden layer) of the k-th expert, respectively. k (s) Let be the implicit representation of the k-th expert at the s-th resolution or scale.

[0208] Expert Output Layer: Generates expert output features and uses Dropout to suppress overfitting.

[0209]

[0210] Among them, E k (s) W represents the output feature of the k-th expert at the s-th resolution or scale. k (2) b k (2) These are the weight matrix and bias vector of the second convolutional layer (expert output layer) of the k-th expert, respectively. Drop(·) represents the regularization dropout operation. Dropout is a regularization method used to randomly block some neurons or channels to prevent overfitting.

[0211] Probability-weighted fusion layer: Selected experts generate output in a probability-weighted manner.

[0212]

[0213] Here, ⊙ represents element-wise multiplication. It is a mixed feature at the s-th resolution or scale.

[0214] Cross-scale feature fusion layer: Through cross-scale fusion, it achieves integrated reasoning from coarse to fine and from local to global to recover the semantic and geometric details of the target.

[0215]

[0216]

[0217] Where O represents the final output feature of MMR (a feature pyramid containing multi-scale geometric information), Up(·) represents the spatial size (H', W') of the nearest neighbor interpolation or transposed convolution upsampling, and ρ s Let β be the learnable weight coefficient at the s-th resolution or scale. s Let ρ be the softmax normalized weight for the s-th resolution or scale, where S represents the total number of resolution layers or scales, and ρ is the weight for the s-th resolution or scale. t It is the exponent denominator used in Softmax normalization, representing the sum of weights for all resolutions or scales, and t represents the resolution or scale index.

[0218] The algorithm for the MMR module is as follows:

[0219] Input: I is the feature tensor (compressed feature map) of the backbone network input; S is the total number of resolution layers or scales; K is the total number of experts used in the routing.

[0220] Output: O is the final output feature after expert routing and multi-scale fusion.

[0221] 1: Perform for each scale s = 1 to S; 2: Pool I with the sampling ratio r to obtain U (s) ;

[0222] 3: From U (s) Extracting geometric embeddings as C (s) ;4: Calculate the gated channel retention factor r using sigmoid activation;5: Set U (s) Multiply by r to obtain the gating feature 6: will and C (s) Connect them to form the query vector Q (s) ;7: From C (s) Estimating manifold curvature k (s) ;8: Perform for each expert k = 1 to K;9: Use Q (s) and k (s) Calculate routes ;10: End; 11: Execute for each expert k = 1 to K; 12: Pass The calculation expert selects the weight a k (s) ;13: Obtain hidden features H from the first convolutional layer (expert hidden layer) k (s) ;14: Obtain the expert output feature E from the second convolutional layer (expert output layer) k (s) 15: End; 16: Use weight a k (s) Transform expert output features into hybrid features 17: End; 18: Execute for each scale s = 1 to S; 19: Calculate the cross-scale weights β s ;20: End; 21: Through weighted upsampling and fusion Calculate the final output O.

[0223] Traditional remote sensing models rely on a large number of parameters and high inference budgets, yet they still suffer from false positives and false negatives in complex contexts. This embodiment of LiteFM-RSDet integrates fractal geometry and manifold curvature principles into the feature learning process. Synaptic fractal pruning (SFP) utilizes left-right branch differences and differentiable gated cyclic hard pruning to precisely remove redundant connections using fractal patterns, significantly reducing FLOPs and memory usage of inefficient channels while preserving fine-grained information at the synaptic layer. The multi-scale manifold router (MMR) constructs a multi-scale pyramid based on SFP compressed features, extracts curvature through Laplacian embedding, and guides labels to lightweight experts based on resolution and geometric adaptability. MMR also employs two-layer regularization of scale and expert to prevent collapse and fuses multi-scale outputs to enhance contextual representation and inference efficiency. This design enables LiteFM-RSDet to significantly recover and enhance multi-scale representation capabilities without incurring additional computational overhead, meeting the real-time requirements of edge deployments while providing significant performance improvements for the detection of small targets and complex structures in remote sensing. Figure 1 A comprehensive evaluation compares the detection performance of different remote sensing target detection methods on the DOTA dataset. LC is a conventional linear compression.

[0224] Example 2

[0225] 1. Dataset

[0226] DOTA-v1.0 is one of the largest rotating frame detection datasets currently available, containing 2,806 frames of remote sensing images with a resolution of up to 4000×4000, including 188,282 instances of 15 target types: aircraft (PL), baseball field (BD), bridge (BR), ground runway (GTF), small vehicle (SV), large vehicle (LV), ship (SH), tennis court (TC), basketball court (BC), storage tank (ST), football field (SBF), roundabout (RA), port (HA), swimming pool (SP), and helicopter (HC).

[0227] HRSC2016 focuses on ship detection, containing 1,061 high-resolution coastline images with a total of 2,976 labeled ship instances. HRSC2016 features extreme target aspect ratios and large-scale spans, making it a classic dataset for evaluating the performance of slender target detection.

[0228] DIOR-R is a rotating frame extension of the DIOR dataset, providing 23,463 multi-source remote sensing images covering 20 common categories (e.g., aircraft, vehicles, ports, tanks, etc.) with a total of approximately 190k instances of orientation labels.

[0229] 2. Experimental Setup

[0230] For the remote sensing object detection task, four RTX 4090 GPUs were used, with a batch size of 8, a learning rate of 0.00005, and training for 12 epochs. Results of the remote sensing object detection model on the DOTA-v1.0, HRSC2016, and FAIR1M-v1.0 datasets are reported. To ensure fairness, the same dataset preprocessing protocol as other mainstream methods was used. For the DOTA-v1.0 dataset, single-scale training and testing were employed. For the FAIR1M-v1.0 dataset, a multi-scale training and testing strategy was used. First, the image size was resized to three scales (0.5, 1.0, 1.5), and then each scaled image was cropped into 1024×1024 blocks with an overlap of 500 pixels. For the HRSC2016 dataset, the long side of the image was set to 800 pixels while preserving its aspect ratio, thus resizing the image.

[0231] Figure 4 The results generated by this method on the DOTA-v1.0 dataset show three different scenarios: dense targets, sparsely distributed targets, and targets submerged in the background.

[0232] 3. Evaluation Indicators

[0233] To evaluate the model's performance on remote sensing datasets, the performance metrics used are described. Mean precision (AP) and mean mean precision (mAP) are commonly used as evaluation metrics in object detection research involving both natural and remote sensing imagery. AP is calculated by measuring the area under the precision-recall curve, with recall ranging from 0% to 100%. Precision and recall are defined as follows:

[0234]

[0235]

[0236] Predicted bounding boxes with an IoU score higher than a predefined threshold are classified as true positives (TP) or false positives (FP). False negatives (FN) are ground-truth bounding boxes that were not detected. AP values ​​fluctuate with different IoU thresholds, generally decreasing as the IoU threshold increases. At a given IoU threshold, higher AP indicates better performance.

[0237]

[0238] Where Q represents the number of categories, |R(Q)| represents the number of images associated with category Q, j is the rank in the retrieved image sequence, n is the total number of retrieved images, P(k) represents the precision at the cutoff point j in the list, and R(q) represents an indicator function that assigns a value of 1 if the image with rank q is relevant, and a value of 0 otherwise. mAP 50 This refers to a scenario where the IoU threshold for calculating mAP is 0.5. IoU quantifies the overlap between the predicted bounding box and the ground truth bounding box. When IoU ≥ 0.5, the prediction is considered correct. Therefore, mAP... 50 A more lenient matching criterion is used, primarily evaluating whether an object has been approximately detected. In contrast, mAP... 75 A 0.75 IoU threshold is used, representing a stricter standard that requires greater consistency between the predicted and ground truth bounding boxes. mAP 75 The model's accuracy in object localization was rigorously evaluated. The AI-TOD dataset includes aerial objects of varying sizes. Objects are categorized by size as follows: very small objects range from [2, 8] pixels, small objects from [8, 16] pixels, relatively small objects from [16, 32] pixels, and medium-sized objects from [32, 64] pixels. There are no large objects in this dataset. The distributions for these categories are 13.3%, 72.3%, 12.3%, and 2.1%, respectively. The experimental section uses the AP metric to evaluate the size of these objects.

[0239] To measure the complexity of BNN, the total number of operations (items) is used as follows:

[0240]

[0241] BOPs quantify the computational load of binary operations, while the number of items accounts for the load of real-valued operations. Since this method does not introduce new binary operations, the computational methodology remains consistent with that used in ReactNet and BiT. Marginal increases in the number of items primarily stem from additional items.

[0242] 4. Main Results

[0243] Table 1 shows that our method maintains leading performance in both single-scale and multi-scale training tests on the DOTA-v1.0 dataset. It achieves significant mAP improvements on complex categories such as bridges, helicopters, ships, and roundabouts, outperforming existing state-of-the-art methods like AO2-DETR, EMO2-DETR, and LSKNet-S, demonstrating more balanced performance. It also surpasses similar models in multi-scale settings, showcasing its advantages in hierarchical feature extraction and spatial representation. Furthermore, it excels in handling small targets such as helicopters and small vehicles, validating its effectiveness in densely distributed and geometrically complex scenes.

[0244] Table 1. Comparison of Single-Scale and Multi-Scale Training and Testing for Various Remote Sensing Target Detection Methods

[0245]

[0246] The breakthrough is even more pronounced in the multi-scale model: with the same resource constraints (28.3M parameters / 195GFLOPs), LiteFM-RSDet achieves 81.80% mAP, surpassing all comparable models.

[0247] Tables 2 and 3 compare LiteFM-RSDet with existing remote sensing target detection methods under the same dataset, evaluation protocol, and complexity metrics. Figure 3 .

[0248] Representative methods from the HRSC2016 rotating ship inspection benchmark are presented, along with VOC-style mAP(07), mAP(12), model complexity parameters, and FLOPs (see Table 2). The results show that LiteFM-RSDet achieves the highest mAP(12) of 98.58% among all methods. Its mAP(07) is 90.60%, matching the current best results. Furthermore, LiteFM-RSDet requires only 28.35 M parameters / 181 G FLOPs, with a complexity approximately 90% of its strongest competitor, ReDet, and a quarter of MTP. In other words, while maintaining top-level accuracy, LiteFM-RSDet achieves the lowest computational and storage thresholds in its class.

[0249] Table 2 Comparison of various remote sensing target detection methods on the HRSC2016 dataset

[0250]

[0251] Table 3 Comparison of lightweight target detection methods for the DIOR-R dataset

[0252]

[0253] Figure 3 The comparison between LiteFM-RSDet and LSKNet-S on the DOTAv1.0 dataset at multiple scales demonstrates that the LiteFM-RSDet model has a significant advantage in detecting objects with complex structures or large-scale variations, with mAP gains typically exceeding 1%. In particular, improvements of 2-3% are achieved for SV and HC, highlighting the effectiveness of LiteFM-RSDet in small object modeling, rotation invariance, and background suppression.

[0254] 5. Ablation test

[0255] 1) Overall Ablation Experiment: Table 4 presents an ablation study evaluating the incremental contributions of Synaptic Fractal Pruning (SFP) and the Expert Hybrid Module (MoE) to the baseline linear compression framework. This benchmark achieved a mean accuracy (mAP) of 73.91% with a model size of 25.7M parameters and 165.2G FLOPs. The introduction of SFP improved the mAP gain by nearly 3% without incurring any additional computational or parameter costs. This demonstrates that SFP effectively improves characterization efficiency by structurally reorganizing the model capacity without increasing its footprint. Further integration...

[0256] After implementing the MoE module, the performance improved to 78.52% mAP, demonstrating that MoE possesses the ability to adaptively specialize computation and enriches the model's expressive power. It is noteworthy that this gain was achieved with only a modest increase in parameters and FLOPs, highlighting a favorable trade-off between accuracy and efficiency.

[0257] Table 4. Experimental data of ablation studies enhanced by synaptic fractal pruning (SFP) and MoE.

[0258]

[0259] SFP improves accuracy without increasing model size, while MoE further improves performance with acceptable overhead.

[0260] 2) Ablation Study of SFP on Network Layer Design: Table 5 shows a comprehensive ablation analysis of each component within the Synaptic Fractal Pruning (SFP) framework, aiming to dissect the functional significance of its key design elements. Empirical results show that the full SFP configuration achieves the highest mAP, reaching 76.87%, establishing a strong performance benchmark. Systematically removing each component, such as the Swap Matrix J, LayerNorm, HardThressholding, Gating (Random Mask), Residual Scaling α, or Dual-Branch Design, leads to a consistent performance degradation.

[0261] This indicates that each component is indispensable to model performance. When the gating layer is replaced by random masking, the accuracy drops to 74.58%, demonstrating that learnable attention is crucial for parameter utilization; removing the bi-branch design and staggered matrix also weakens spatial diversity, leading to performance degradation. Despite having the same number of parameters and computational cost, the difference in mAP between different variants proves that SFP's advantage stems from architectural choice rather than simple compression, and the collaborative design of a fully configured system is crucial to the overall pruning strategy.

[0262] Table 5. Experimental data on ablation studies of various neural network layers in Synaptic Fractal Pruning (SFP).

[0263]

[0264] 3) Ablation Study of SFP under Different Pruning Rates: Table 6 explores the impact of different pruning rates within the SFP framework. The results show a negative correlation between pruning rate and mAP: the full model has the highest mAP (78.60%) but requires a large computational load; 25% pruning, while maintaining high accuracy (76.87%), significantly reduces parameters and FLOPs to 25.7M and 165.2 GFLOPs, respectively, achieving the best balance between performance and efficiency. Beyond 25%, the accuracy decline intensifies, and the efficiency gains diminish.

[0265] Table 6 shows the experimental data of ablation studies on synaptic fractal pruning under different pruning ratios.

[0266]

[0267] 4) Effectiveness of Individual Components in MMR: Table 7 evaluates the contribution of each MMR component, validating its effectiveness in geometry-aware expert routing. Starting with a baseline model using only MoE (mAP 77.12%), performance was progressively improved by introducing Geometric Embedding, Multi-resolutioan Fusion, Curvature Bias, Sparse Routing, and Lapacian Sensing, ultimately achieving an mAP of 78.52% in the full configuration. Despite unchanged parameters and computational cost, removing any component resulted in a decrease in accuracy, demonstrating the crucial role of multi-scale geometric features and structured manifold signals within the MoE framework.

[0268] Table 7 shows the experimental data from the ablation study of multi-resolution manifold routers (MMR).

[0269]

[0270] 5) Performance under different numbers of experts in the MMR module: Table 8 analyzes the impact of the number of experts K in the MMR module. As K increases from 2 to 6, mAP steadily improves. With a moderate increase in parameters and FLOPs, performance is optimal at K=6 (78.52%). However, further increasing to 8 results in a slight decrease, indicating that too many experts introduce redundancy and interference. The results show that a balance should be struck between expert diversity and routing accuracy, with K=6 being the optimal compromise.

[0271] Table 8 Experimental data for different numbers of experts in multi-resolution manifold routers

[0272]

[0273] 8. Visualization

[0274] To qualitatively evaluate the feature focusing capability of LiteFM-RSDet, visual analysis was performed using eight representative remote sensing scenes from the DOTA dataset.

[0275] Two typical cases will be discussed in detail: the first case corresponds to a scene with regular structure and simple texture, such as Figure 5 In each set, the left image is the original image of the superimposed detection frame, and the right image is the attention heatmap corresponding to the feature map. The baseline heatmap is scattering noise with low target-background contrast. The second case involves a visually complex residential area, characterized by densely packed roofs, interwoven textures, and a complex road network, such as... Figure 6 .

[0276] For each scene, the original image covering the detection results is presented. In typical scenes, the attention response of the baseline model is largely homogeneous, with low contrast between foreground and background regions. This leads to increased detection box drift and false positives, indicating suboptimal localization and weak discriminative focus. In contrast, LiteFM-RSDet produces significantly high-response regions, precisely aligned with semantically meaningful structures such as vehicle outlines and gaps between train carriages, while maintaining low activation in irrelevant regions such as empty spaces and shadowed areas. The resulting detection boxes are closely aligned with the true object geometry, demonstrating the ability of our proposed model to accurately focus on small, regular targets.

[0277] In complex residential scenarios, the baseline model exhibits scattered and noisy activations in roof ridges, vegetation, and shadow areas, leading to significant missing detections, particularly in narrow roads and congested traffic. In contrast, LiteFM-RSDet generates spatially coherent activation patterns, with high-response regions corresponding to key semantic structures such as rooftops, road medians, and vehicle clusters, while suppressing texture clutter. The resulting densely distributed and precisely localized bounding boxes demonstrate the effectiveness of LiteFM-RSDet in preserving multi-scale semantic information, which typically degrades during deep compression in cluttered environments.

[0278] In summary, these visualizations highlight two key mechanisms underpinning the effectiveness of the LiteFM-RSDet model. First, SFP eliminates redundant channels, thereby promoting sparse and efficient feature representations. Second, MMR enhances the model's ability to reassign attention to geometrically salient regions in both regular and complex scenes. The synergy of these components enables LiteFM-RSDet to display high-contrast heatmaps superior to the baseline, emphasizing target-relevant regions while suppressing background noise, thus significantly reducing false positives and false negatives. The visualizations also demonstrate the corresponding attention heatmaps generated by the baseline model and LiteFM-RSDet, emphasizing target-relevant regions while suppressing background noise, further reducing false positives and false negatives.

[0279] This embodiment presents a lightweight remote sensing image multi-scale target detection framework, LiteFM-RSDet. Unlike existing lightweight methods, LiteFM-RSDet does not rely on additional differential convolutions or large kernel blocks. Instead, it focuses on synaptic fractal pruning (SFP) and multi-resolution manifold routers (MMR): SFP prunes channels and connections simultaneously through fractal recursion gates, significantly reducing parameter counting and computational redundancy; MMR employs curvature-driven sparse routing on the Laplacian manifold, dynamically assigning residual labels to the most suitable scale and geometric subspace, recovering the multi-scale semantic information lost after deep compression. This two-stage collaboration enables the model to maintain a complete semantic pyramid even after compression, while generating highly concentrated feature responses in regions with weak texture or small targets.

[0280] This embodiment systematically validates the proposed method on DOTA-v1.0, DIOR-R, and HRSC2016: LiteFM-RSDet, while maintaining end-to-end friendly 28M level parameters and a 190GFLOPs budget, comprehensively outperforms mainstream lightweight detection methods such as equivalent algorithms KFIoU, RTMDet R, and LSKNet-S in terms of mAP, and is equal to or exceeds the real-valued transformer backbone with several times the algorithm power. Visualization experiments further demonstrate that LiteFM-RSDet can generate heatmap patterns of bright spot concentration and background suppression in structured landfills and complex residential areas, significantly reducing false positives and false negatives compared to the baseline. It also visually demonstrates the enhanced saliency of the SFP-MMR framework. This proves the effectiveness of the SFP-MMR framework in enhancing saliency.

[0281] The LiteFM-RSDet of this embodiment has significant advantages: it achieves breakthroughs in accuracy in general target detection, rotating target recognition, and small target scenes with low parameter counting and computation. Meanwhile, heatmaps verify its ability to accurately focus on target regions and suppress background interference, effectively reducing the false detection rate in complex scenes.

[0282] Overall, LiteFM-RSDet achieves the best trade-off between computing power, storage, and detection accuracy through synaptic fractal pruning and multi-resolution manifold routers, without introducing additional large cores or ad hoc operators, providing an extremely practical idea for lightweight target detection in high-resolution remote sensing scenarios.

[0283] The above embodiments should be understood as being used only to illustrate the present invention more clearly, and not to limit the scope of the present invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art fall within the scope defined by the appended claims.

Claims

1. A remote sensing target detection method, characterized in that, Includes the following steps: S1. Acquire remote sensing images; S2. Divide the remote sensing image into patches and generate location codes to obtain a patch sequence; S3. Input the patch sequence into the Transformer encoder to extract global semantic features and obtain a feature map; S4. Input the feature map into the synaptic fractal pruning module to remove redundant channels and retain key features to obtain a compressed feature map; The specific process of S4 includes: The feature maps are input to the left and right branch nerve layers. After the feature maps are copied into the left and right nerve branches, the left and right nerve branches are projected onto the feature maps and the responses of the left and right nerve branches are connected row by row to generate a dual-path stacked feature matrix. The interleaved features are obtained by multiplying the interleaved sequence matrix from bottom to top by the dual-path stacked feature matrix; By performing row-by-row differencing on the alternating features, we obtain the difference operator; The difference operator is normalized to generate normalized difference features for each connection or channel; The normalized difference features of each connection or channel are input into the first gating layer to obtain the retention probability of each connection or channel; Based on the retention probability of each connection or channel, determine whether each connection or channel is retained or pruned. Connections or channels with a retention probability greater than a threshold are retained, and connections or channels with a retention probability less than or equal to the threshold are pruned. The features of the retained connections or channels are fused into a diagonal mask matrix; The diagonal mask matrix is ​​processed by the first linear layer, the GELU activation layer, and the second linear layer to obtain the updated feature vector; The updated feature vector is input into the residual injection layer and fused with the feature map to obtain a compressed feature map. S5. Input the compressed feature map into a multi-resolution manifold router to generate a feature pyramid containing multi-scale geometric information; The specific process of S5 includes: The compressed feature map is input into the multi-resolution feature downsampling layer, and downsampling is performed at each resolution or scale to generate a feature tensor for each resolution or scale. The feature tensor of each resolution or scale is passed through the second gating layer to obtain the gated feature tensor of each resolution or scale; The gated feature tensors at each resolution or scale are subjected to nonlinear transformation and regularization through the first and second convolutional layers in sequence to obtain the output features of each expert at each resolution or scale. The feature tensor at each resolution or scale extracts low-dimensional coordinate information through a Laplacian geometric embedding layer to obtain the manifold coordinate tensor at each resolution or scale. Input the manifold coordinate tensor and the gated feature tensor of each resolution or scale into the query generation layer to generate a query vector for each resolution or scale that integrates geometric and texture information. The query vector at each resolution or scale is processed by softmax activation to obtain the route for each expert at each resolution or scale; Normalize the routes of each expert at each resolution or scale to obtain the selection weights of each expert at each resolution or scale. Based on the selection weights of each expert at each resolution or scale, the output features of all experts at each resolution or scale are weighted and fused to obtain the mixed features at each resolution or scale. By fusing mixed features from all resolutions or scales, a feature pyramid containing multi-scale geometric information is generated. S6. Input the feature pyramid into the detection head to predict the target category and location, and obtain the final detected target; The synaptic fractal pruning module comprises a fusion layer, an interleaved computation layer, a differential computation layer, a first normalization layer, a first gating layer, a first linear layer, a GELU activation layer, a second linear layer, and a residual injection layer connected in sequence. The left branch neural layer and the right branch neural layer are respectively connected to the fusion layer, and the interleaved computation layer is connected to the first linear layer. The Transformer encoder is connected to the left branch neural layer, the right branch neural layer, and the residual injection layer. The multi-resolution manifold router includes a multi-resolution feature downsampling layer, a second gate layer, a first convolutional layer, a second convolutional layer, a probability-weighted fusion layer, and a cross-scale feature fusion layer connected in sequence. The multi-resolution feature downsampling layer is connected to a Laplacian geometric embedding layer. The Laplacian geometric embedding layer and the second gate layer are respectively connected to a query generation layer. The query generation layer and the Laplacian geometric embedding layer are respectively connected to a softmax activation layer. The softmax activation layer is connected to a second normalization layer. The second normalization layer is connected to a probability-weighted fusion layer. The residual injection layer is connected to the multi-resolution feature downsampling layer.

2. The remote sensing target detection method according to claim 1, characterized in that, Interlaced feature X int The expression is as follows: ； ； ； Among them, W L W represents the weight matrix of the left branch projection. R Let X represent the weight matrix of the right branch projection, and let I represent the feature map. d H represents the identity matrix, and H represents the dual-path stacked feature matrix.

3. The remote sensing target detection method according to claim 1, characterized in that, The expression for the diagonal mask matrix M is as follows: ； ； ； Where, m i This indicates that the i-th connection or channel is reserved, 1 {·} This is an indicator function; it outputs 1 if the condition is met, and 0 if not. τ is the threshold, and diag() indicates that the vector within the parentheses is converted into a diagonal mask matrix M; W g and b g The weight vector and bias, h, are used to calculate the gating score, respectively. i Let g represent the normalized difference feature of the i-th connection or channel, σ(·) represent the Sigmoid activation function, and g i This represents the retention probability of the i-th connection or channel.

4. The remote sensing target detection method according to claim 1, characterized in that, Compressed feature map The expression is as follows: ； ； Where ϕ(·) represents the GELU nonlinear activation function, W h and W o These are the weight matrices of the first and second linear layers, respectively; Z is the updated feature vector; ⊙ denotes element-wise multiplication; α represents the residual scaling factor; M is the diagonal mask matrix; X... int It is an interlaced feature.

5. The remote sensing target detection method according to claim 1, characterized in that, The expression for the gated feature tensor at each resolution or scale is as follows: ； ； ； Where σ(·) represents the Sigmoid activation function, v represents the gating vector, r represents the channel retention factor, and ⊙ represents element-wise multiplication. Represents the gated feature tensor at the s-th resolution or scale; I represents the compressed feature map, r s Pool is the downsampling ratio for the s-th resolution or scale. rs (·) denotes a convolution or mixed pooling operator, U (s) Let be the feature tensor of the s-th resolution or scale.

6. The remote sensing target detection method according to claim 1, characterized in that, The expression for the selection weights of each expert at each resolution or scale is as follows: ； ； ； ； Among them, W k This represents the gating vector of the k-th expert. λ represents the Softmax temperature, λ represents the curvature bias coefficient, and k represents the Softmax temperature. (s) This represents the average manifold curvature at the s-th resolution or scale. This represents the route taken by the k-th expert at the s-th resolution or scale. This represents the route taken by the j-th expert at the s-th resolution or scale, where K is the total number of experts, and a k (s) Q represents the selection weight of the k-th expert at the s-th resolution or scale. (s) Represents the query vector at the s-th resolution or scale; ψ(·) represents the interleaved linear function, GeomEmbed(·) represents the Laplacian eigenmap or diffusion map operator, and C (s) U represents the manifold coordinate tensor at the s-th resolution or scale. (s) Let be the feature tensor at the s-th resolution or scale. Represents the gated feature tensor at the s-th resolution or scale.

7. The remote sensing target detection method according to claim 1, characterized in that, The expression for the blending feature at each resolution or scale is as follows: ； ； ； in, a represents the blending feature at the s-th resolution or scale. k (s) Let φ(·) represent the selection weight of the k-th expert at the s-th resolution or scale, and W be the GELU activation function. k (1) b k (1) H represents the weight matrix and bias vector of the first convolutional layer of the k-th expert, respectively. k (s) E represents the implicit representation of the k-th expert at the s-th resolution or scale; k (s) W represents the output feature of the k-th expert at the s-th resolution or scale. k (2) b k (2) Let be the weight matrix and bias vector of the second convolutional layer for the k-th expert, respectively. Drop(·) represents the regularization operation, and ⊙ represents element-wise multiplication. Represents the gated feature tensor at the s-th resolution or scale.

8. The remote sensing target detection method according to claim 1, characterized in that, The expression for the feature pyramid is as follows: ； ； Where O represents the feature pyramid, Up(·) represents the spatial size of the nearest neighbor interpolation or transpose convolution upsampling, and ρ s β is the weighting coefficient for the s-th resolution or scale. s ρ represents the softmax normalized weights for the s-th resolution or scale, where S represents the total number of resolution layers or scales. t This represents the exponent denominator used for Softmax normalization, t represents the resolution or scale index, and Drop(·) represents the regularization operation.

Images