A target detection system based on state space model and multi-scale attention
Patent Information
- Application Number
- CN202580005847.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-28
- Publication Date
- 2026-10-09
AI Technical Summary
Existing aerial target detection methods are ineffective in low-visibility environments, have high computational costs, and their model complexity is not compatible with the computational and storage limitations of UAVs.
A target detection system based on a state-space model and multi-scale attention is adopted, which combines the HAFE module, the MAG module and knowledge distillation technology. The HAFE module enhances feature extraction in hazy environments, the MAG module optimizes the detection of small targets and fuzzy features, and the knowledge distillation strategy reduces model complexity.
It improves detection accuracy in hazy environments, reduces model complexity, adapts to the computing and storage limitations of UAVs, has a low false detection rate, and is robust, making it suitable for fields such as environmental monitoring and national defense early warning.
Smart Images
Figure CN122893142A_ABST
Abstract
Description
A target detection system based on state-space model and multi-scale attention Technical Field
[0001] This invention relates to image recognition and processing technology, and in particular to a target detection system based on a state-space model and multi-scale attention. Background Technology
[0002] Aerial target detection is an important research direction in the field of remote sensing information processing, aiming to accurately identify and locate ground targets using aerial imagery. This technology has attracted widespread attention in various fields such as environmental monitoring, urban intelligence systems, and defense early warning frameworks. Existing methods mainly rely on carefully compiled high-resolution remote sensing datasets for training and evaluation, while ignoring image degradation caused by low visibility (such as haze). This limitation is particularly evident in scenarios requiring consistency across multiple time scales of imaging, as degradation can propagate along the feature processing flow, thereby weakening the model's generalization ability and robustness.
[0003] To address this issue, mainstream approaches employ a two-stage process: first, a dehazing algorithm is applied, and then the enhanced image is fed into a detector trained on the HR dataset. However, this staged approach leads to tight algorithm coupling and high computational costs.
[0004] Furthermore, deploying complex models on drone equipment presents challenges. Knowledge distillation, an efficient optimization technique, offers practical solutions and has been widely used in recent years to address target detection in complex environments. However, these methods primarily focus on improving detection accuracy and do not fully address the issue of whether model size is compatible with the computational and storage limitations of drones.
[0005] Therefore, how to improve the detection performance of aerial targets in low-visibility environments, reduce computational costs, and match the model size with the computational and storage limitations of UAVs are the technical problems that need to be solved. Summary of the Invention
[0006] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is to provide a target detection system based on a state space model and multi-scale attention, which is suitable for UAVs to perform tasks in scenarios with poor visual conditions; through knowledge distillation strategy, the model complexity can be reduced and the adverse effects of environmental occlusion on perception accuracy can be mitigated.
[0007] To achieve the above objectives, this invention provides a target detection system based on a state-space model and multi-scale attention, including a state-space model comprising:
[0008] The Mamba backbone network is used to extract multi-level features. The Mamba backbone network includes the HazySimple module and the MS-VSSBlock module. The HazySimple module integrates the HAFE module, which enhances feature extraction in hazy environments. The MS-VSSBlock module integrates the MAG module to optimize the detection of small targets and blurred features. The HAFE module provides high-quality feature input to the MAG module, which further optimizes the detection of small targets and blurred features based on this input.
[0009] The PAFPN neck structure receives multi-level features output from the backbone network, and finely fuses high-level semantic features with low-level spatial features through XSSBlock before outputting them to the detection head, thereby enhancing the richness of feature representation.
[0010] The state-space model extracts multi-level features from the Mamba backbone network and then fuses these features into PAFPN. XSSBlock is responsible for refining and fusing high-level semantic features with low-level spatial features, and finally outputs the features to the decoupled head to obtain the detection results.
[0011] As a further improvement of the present invention, the HAFE module for input features The HAFE module first generates low-frequency features. To suppress noise, while extracting multi-scale features through convolution operations. ; then and The data is concatenated, then enhanced by a convolution and SiLU activation function, and finally merged with the residuals to generate the output. .
[0012] As a further improvement of the present invention, the HAFE module has improved input features. The processing procedure satisfies:
[0013] Low-frequency characteristics:
[0014] ;
[0015] Multi-scale features:
[0016] , ;
[0017] Output characteristics:
[0018] .
[0019] As a further improvement of the present invention, the MAG module for input features The MAG module extracts main branch features through bottleneck structures. And generate the remaining branch features. The main branch is then merged with the remaining branches, and multi-scale convolution is used to extract different receptive field features. and Next, these features are concatenated and fused through convolutional operations to generate the output. .
[0020] As a further improvement of the present invention, the processing procedure of the MAG module satisfies:
[0021] Remaining branch characteristics: ;
[0022] Main branch characteristics:
[0023] , ,
[0024] ;
[0025] Fusion Output:
[0026] ;
[0027] ,
[0028] .
[0029] As a further improvement of the present invention, a knowledge distillation module is also included, which performs knowledge distillation through an adaptive fog perception distillation loss to transfer the professional knowledge of the complex teacher network to the lightweight student network; the knowledge distillation module combines fog perception weights and KL divergence loss, wherein the fog perception weights are implemented through dark channel prior knowledge.
[0030] As a further improvement of the present invention, the adaptive haze perception distillation loss function of the knowledge distillation module satisfies:
[0031] Weighted classification loss:
[0032] ;
[0033] Weighted regression loss:
[0034] ;
[0035] Total loss:
[0036] .
[0037] As a further improvement to the present invention, haze perception weight satisfy:
[0038] ;
[0039] ;
[0040] .
[0041] As a further improvement to the present invention, it also includes:
[0042] The server is used to install the Hazy-Mamba model and perform subsequent model training and distillation.
[0043] On the application side, it is used to acquire images and detect aerial targets based on the onboard student network, and feed the detection results back to the state space model for optimization training.
[0044] As a further improvement of the present invention, the operation method of the target detection system includes the following steps:
[0045] The S100 module trains a high-precision complex state-space model and network to enhance detection performance. The HAFE module employs low-pass filtering, multi-scale convolution, and contrast enhancement techniques to capture multi-scale context and sharp features from fog-affected images. The MAG module uses a bottleneck structure to compress and enhance feature representations, thereby reducing computational complexity while preserving key information. Multi-scale convolution captures features of targets of different sizes, enhancing the saliency of small targets, while residual connections improve feature stability.
[0046] S200 employs knowledge distillation technology, using a complex network as a teacher network to train a lightweight student network, thereby achieving a balance between accuracy and model complexity;
[0047] S300 deploys the student network to the application to quickly detect targets in images, feeds the detection results back to the Hazy-Mamba model for optimization training, and then distributes them to the application after distillation via S200. Beneficial effects
[0048] The Hazy-Mamba model of this invention is an innovative framework based on the MambaYOLO model. It integrates the Haze Adaptive Feature Enhancement (HAFE) module and the Multi-Scale Attention Gating (MAG) module. This invention solves the problems of low detection accuracy, high model complexity, and difficulty in adapting to UAV deployment in haze environments through knowledge distillation technology.
[0049] This invention boasts three core advantages: First, high detection accuracy. The HAFE and MAG modules work synergistically, achieving mAP of 58.9% and 39.2% on the HazyDet test set and RDDTS, respectively, significantly outperforming existing methods. Second, lightweight deployment. Knowledge distillation technology reduces model parameters to 6M, keeping GFLOPs at 95.0, adapting to the computational and storage limitations of UAVs. Third, strong environmental adaptability. Adaptive haze perception distillation loss dynamically adjusts the learning focus, resulting in a low false detection rate and excellent robustness in real haze scenarios. The combination of these three advantages effectively balances accuracy and efficiency, providing a highly efficient solution for aerial target detection in low-visibility environments. This invention is applicable to UAVs and can be widely used in environmental monitoring, national defense early warning, and other fields. Attached Figure Description
[0050] Figure 1 is a schematic block diagram of the system configuration of the present invention;
[0051] Figure 2 is a schematic diagram of the architecture of the Hazy-Mamba model;
[0052] Figure 3 is a schematic diagram of the HAFE module architecture; where Cin represents the number of input feature map channels, k represents the kernel size, and g represents the number of groups;
[0053] Figure 4(a) is a schematic diagram of the MS-VSSBlock architecture;
[0054] Figure 4(b) is a schematic diagram of the MAG module architecture;
[0055] Figure 5 is a schematic diagram of the knowledge distillation process in the knowledge distillation module;
[0056] Figure 6 is a comparison of photos taken using the MambaYOLO model and the Hazy-Mamba model; these images were taken by drones under simulated and real haze conditions, showing two sets of comparison results, one with confidence level and category label, and the other without.
[0057] Figure 7 is a comparison of the characteristic heatmaps of the Baseline Model and the Hazy-Mamba Model; the two images represent UAV images under simulated haze conditions and real haze conditions, respectively, and from left to right they represent characteristic heatmaps at different levels of the backbone network. Detailed Implementation
[0058] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.
[0059] Referring to Figure 1, the target detection system based on the state-space model and multi-scale attention in this embodiment includes a state-space model (Hazy-Mamba model). The Hazy-Mamba model is an innovative framework based on the MambaYOLO model. The Hazy-Mamba model includes:
[0060] The Mamba backbone network is used to extract multi-level features. It includes the HazySimple module and the MS-VSSBlock module. The HazySimple module integrates the HAFE module to enhance feature extraction in hazy environments. The MS-VSSBlock module integrates the MAG module to optimize the detection of small targets and blurred features. The HazySimple module is built on top of the SimpleStem module and integrates the innovatively designed HAFE module. Specifically, the HazySimple module is extended by integrating the specially designed HAFE module on top of the SimpleStem module.
[0061] The PAFPN neck structure receives multi-level features output from the backbone network, and finely fuses high-level semantic features with low-level spatial features through XSSBlock before outputting them to the detection head, thereby enhancing the richness of feature representation.
[0062] The Mamba backbone network first performs downsampling and feature extraction through the HazySimple module, and then further processes it through the MS-VSSBlock and VisionClueMerge modules.
[0063] Referring to Figure 2, the Hazy-Mamba model utilizes the HazySimple module and the MS-VSSBlock module to construct its backbone network. Multi-level features, such as {C3, C4, C5}, are extracted from the backbone network and then fused into the PAFPN. XSSBlock is responsible for refining and fusing high-level semantic features with low-level spatial features, ultimately outputting the features {P3, P4, P5} to the decoupled head to obtain the detection results.
[0064] To enable UAVs to detect blurred features and small targets in foggy environments, this embodiment introduces Haze Adaptive Feature Enhancement (HAFE) and Multi-Scale Attention Gating (MAG) modules. The HAFE module is integrated into the SimpleStem module, enhancing feature extraction and effectively eliminating haze interference through low-pass filtering, multi-scale convolution, and contrast enhancement. The MAG module is integrated into the MS-VSSBlock module, optimizing small target detection and blurred feature detection through bottleneck structures and residual connections, thereby improving the detection capability for small targets and blurred features.
[0065] The HAFE module provides high-quality feature input to the MAG module. Based on these features, the MAG module further optimizes the detection of small targets and blurred features, forming a progressive processing chain of "dehazing - fine detection". The HAFE module focuses on global dehazing and feature enhancement to solve the overall image degradation problem; the MAG module focuses on local small target / blurred feature optimization to make up for the shortcomings in detail recognition. The two work together to improve detection robustness.
[0066] Referring to Figure 3, the Haze Adaptive Feature Enhancement (HAFE) module and the Hazy Simple module are designed to extract multi-scale context and dehazing features from fog-affected images through low-pass filtering and multi-scale convolution.
[0067] For input features The HAFE module first generates low-frequency features. To suppress noise, while through and Convolution operations extract multi-scale features Where Cin represents the number of input feature map channels, k represents the kernel size, and g represents the number of groups; subsequently, and Piece them together, and then through a Convolution and SiLU activation functions are used to enhance key features, which are then combined with the residuals to generate the output. :
[0068] ,
[0069] , ,
[0070] ;
[0071] The HazySimple module is built on top of the SimpleStem module and integrates the innovatively designed HAFE module, which can improve the feature extraction capability of UAV images under hazy conditions.
[0072] Referring to Figures 4(a) and 4(b), the MAG module effectively solves the problems of blurred features and small target detection in UAV fog images through a carefully designed bottleneck structure and multi-scale feature extraction with residual connections.
[0073] The HAFE module employs low-pass filtering, multi-scale convolution, and contrast enhancement techniques to capture multi-scale context and sharp features from fog-affected images. The MAG module uses a bottleneck structure to compress and enhance feature representations, thereby reducing computational complexity while preserving key information. Multi-scale convolution can capture target features of different sizes and enhance the saliency of small targets, while residual connections help improve feature stability, thus effectively addressing the challenges of detecting blurred features and small targets in foggy scenes.
[0074] The scattering and absorption effects of atmospheric haze lead to three main degradation characteristics (i.e., contrast degradation, noise enhancement, and attenuation). Under low visibility conditions, high-frequency texture components are particularly noticeable in UAV remote sensing images, directly impacting target saliency perception and detail resolution capabilities.
[0075] The HAFE module works as follows: low-pass filtering can suppress high-frequency noise and thus enhance low-frequency structural information in the image, while multi-scale convolution can capture target features at different scales, thereby improving feature robustness in blurred scenes.
[0076] While the MS-VSSBlock module can effectively extract and enhance features, it cannot handle the challenges of detecting blurred features and small targets in hazy remote sensing images. To address this issue, the MAG module was proposed.
[0077] As shown in Figure 4(a), the MAG module is integrated into a new path branch in the MS-VSSBlock module and works with the MS-VSSBlock components to extract features. The architecture of the MAG module is shown in Figure 4(b). ① The pointing arrow indicates the remaining connections.
[0078] Specifically, for input features The MAG module extracts main branch features through bottleneck structures. And generate the remaining branch features. :
[0079] ;
[0080] , ;
[0081] ;
[0082] Subsequently, the main branch and the remaining branches are merged, and multi-scale convolution is used to extract different receptive field features. and Next, these features are concatenated and fused through convolution operations to generate the output. :
[0083] ,
[0084] ;
[0085] .
[0086] MS-VSSBlock integrates multiple modules, such as LSBlock, SSM, MLP, and MAG, enabling efficient feature extraction and enhancement. Furthermore, through residual connections and DropPath, the model's stability and generalization ability are improved, making it particularly suitable for target detection in foggy scenes from a drone's perspective.
[0087] Referring to Figure 5, to address the challenge of deploying complex networks on UAV platforms, this embodiment adds a knowledge distillation module. This module transfers knowledge from the teacher network to the student network using an adaptive haze perception distillation loss function. Specifically, the knowledge distillation module employs an adaptive haze perception distillation loss to transfer the expertise of the complex teacher network to the lightweight student network, achieving a balance between efficiency and accuracy. Through this knowledge distillation strategy, the framework reduces model complexity and mitigates the adverse effects of environmental occlusion on perception accuracy.
[0088] This method transfers knowledge from complex networks to lightweight student networks, achieving a balance between model complexity and high detection accuracy. The distillation process is shown in Figure 5. By combining haze perception weights and KL divergence loss, the adaptive haze perception distillation loss dynamically adjusts the learning focus, optimizing the student network's detection performance in complex scenes. Specifically, the haze perception weights are implemented using dark channel prior knowledge, effectively estimating the haze concentration in the image and further enhancing the model's adaptability to haze scenes. The loss can be expressed as:
[0089]
[0090] in, is the loss scaling factor, which controls the overall strength of the distillation loss. N represents the number of predictions made by the network. It is the loss of the weighted classification task, while This is the loss from the weighted regression task. and It can be defined as:
[0091] ;
[0092] ;
[0093] in, It is the divergence loss, and S is the softmax operation. It is the distillation temperature. , , , These are classification predictions and regression predictions for student networks and teacher networks, respectively. It's the weighting for smog perception. Simply put, The implementation concept is based on the dark channel value, which reflects the brightness of areas in the image that are not completely obscured by haze. The higher the haze concentration, the larger the dark channel value, because haze increases the overall brightness. The calculation uses the following formula:
[0094] ;
[0095] ;
[0096] ;
[0097] in, These are the pixel values of the input image in channel C. It is the sigmoid function, where d is the offset, used to adjust the sensitivity in the fog area. It is the normalization factor in weight calculation, used to adjust the weight range to a reasonable interval. These are parameters of the fog area weighting coefficient, used to control the fog area ( The influence of the weighting is denoted by R, G, and B, which are the R, G, and B values of the image. This approach combines the distillation loss with the student network's regular detection loss to jointly optimize the student network's parameters, enabling it to dynamically adjust its learning focus and achieve better performance in hazy scenes.
[0098] The target detection system based on state-space model and multi-scale attention also includes:
[0099] The server is used to install the Hazy-Mamba model and perform subsequent model training.
[0100] On the application side, images are acquired and aerial targets are detected using the onboard student network. The detection results are then fed back to the Hazy-Mamba model for optimization and training. In this embodiment, the application side is a drone, specifically an image acquisition device, such as a camera, mounted on the drone.
[0101] The target detection method based on state-space model and multi-scale attention includes the following steps:
[0102] The S100 module trains a high-precision, complex Hazy-Mamba model and network to enhance detection performance. The HAFE module employs low-pass filtering, multi-scale convolution, and contrast enhancement techniques to capture multi-scale context and sharp features from fog-affected images. The MAG module uses a bottleneck structure to compress and enhance feature representations, thereby reducing computational complexity while preserving key information. Multi-scale convolution can capture target features of different sizes, enhancing the saliency of small targets, while residual connections help improve feature stability, thus effectively addressing the challenges of detecting blurred features and small targets in foggy scenes.
[0103] S200 employs knowledge distillation technology, using a complex network as a teacher network to train a lightweight student network, thereby achieving a balance between accuracy and model complexity;
[0104] S300 deploys the student network to the application to quickly detect targets in images, feeds the detection results back to the Hazy-Mamba model for optimization training, and then distributes them to the application after distillation via S200. Control experiment
[0105] 3.1 Experimental Setup
[0106] Due to the lack of large-scale datasets for target detection from a drone's perspective under challenging weather conditions, experiments were conducted on the recently released synthetic haze aerial dataset HazyDet. This dataset contains approximately 383,000 instances and includes the following components:
[0107] 1) Simulated fuzzy data: Generated from drone images in clear weather using an atmospheric scattering model (ASM), covering various environments (such as urban, suburban and coastal areas), different shooting angles (such as oblique and direct views), and lighting conditions ranging from bright to dim.
[0108] 2) Realistic Blurred Data (RDDTS): Accurate bounding box labels are generated using a semi-automatic annotation process based on real-world drone images captured in foggy weather and collected from the internet.
[0109] HazyDet consists of a training set, a validation set, and a test set. The training set contains simulated data and some real data, while the test set includes simulated fuzzy data from RDDTS and real fuzzy data designed to evaluate the model's performance in virtual and real scenarios. All experiments were conducted using the PyTorch framework with the SGD optimizer for model training. The training process lasted 300 stages with an initial learning rate of 0.01, dynamically adjusted according to a predetermined strategy during training. All experiments were performed on a single NVIDIA RTX 5070 Ti GPU; comparison images are shown in Figure 6.
[0110] 3.2 Performance Verification
[0111] State of the art (SOTA) compared Hazy-Mamba with various state-of-the-art object detection and dehazing methods on the HazyDet dataset, covering single-stage, two-stage, and end-to-end approaches. All models were trained using synthetic data from the HazyDet training set and evaluated on the test set and RDDTS based on accuracy and efficiency metrics. Detailed results are shown in Table 1.
[0112] Table 1: Performance comparison of object detection methods on the HazyDet and RD-DTS datasets. The top two results are highlighted with a box and bold black text.
[0113]
[0114] As shown in Table 1, Hazy-Mamba achieved mAP50-95 values of 58.9% and 39.2% on the HazyDet test set and RDDTS, respectively, which indicates a significant improvement compared to the state-of-the-art (SOTA) method.
[0115] This advancement is attributed to its HAFE module, which effectively eliminates haze interference and improves feature representation through low-pass filtering and multi-scale convolution; and its MAG module, which optimizes small object detection through bottleneck structures and residual connections.
[0116] Furthermore, the adaptive haze perception distillation loss dynamically adjusts the learning focus using knowledge distillation, thereby mitigating the impact of haze on student network predictions and improving accuracy. The number of parameters was reduced to 6M while maintaining 95.0 GFLOPS, primarily due to the lightweight design of HAFE and MAG, which reduces redundant computations, and the knowledge distillation process, which compresses the complex neural network structure. Figure 6 demonstrates the advantages of Hazy-Mamba over MambaYOLo in detecting haze and small targets, with a significantly reduced false positive rate.
[0117] 3.3 Elimination Study
[0118] This section provides a comprehensive validation of each component of the Hazy-Mamba module through ablation studies, using MambaYOLO as a benchmark. Evaluation metrics include the detector's accuracy on synthetic and real datasets, as well as the number of model parameters and computational burden.
[0119] Table 2: Performance comparison of HAFE and defogging methods. The top two results are highlighted with a border and bold black text.
[0120]
[0121] 1) Effectiveness of HAFE: As shown in Table 2, AOD and GatedDIP showed relatively limited performance on the test set and RDDTS, while adding HAFE significantly improved detection accuracy, especially on RDDTS. This is because AOD and GatedDIP primarily focus on improving overall image quality, neglecting the necessity of preserving target semantic information and details. In contrast, HAFE focuses on feature extraction and contrast enhancement, which improves feature quality in hazy scenes, thereby enhancing model performance.
[0122] 2) Effectiveness of MAG: As shown in Table 3, the impact of different intermediate channel numbers (Cmid) in the MAG module on accuracy and model parameters was verified. Referring to Figure 7, the higher intermediate dimension in MAG enhances multi-scale feature extraction, thereby improving the detection capability for small targets and blurred features. However, it also increases the computational units and feature dimensions in the network layers, thus increasing the model's complexity.
[0123] Table 3: Performance comparison of different intermediate channel numbers (Cmid) in MAG.
[0124]
[0125] 3) Effectiveness of Knowledge Distillation Adjustment: As shown in Table 4, the proposed "fog-awareness distillation loss strategy" significantly outperforms MGD on RD-DTS. This is because MGD treats prediction errors across all regions equally and cannot dynamically adjust weights based on the environment. In contrast, the distillation strategy in this application combines "fog-awareness weights" and "KL divergence," enabling dynamic adjustment of learning focus to adapt to changes in fog. Compared to the baseline model, the student network exhibits minimal changes in model parameters but significantly improved accuracy, achieving a balance between low model complexity and high detection accuracy.
[0126] Table 4: Ablation Study of Knowledge Distillation Strategies
[0127]
[0128] 4) Visualization: Referring to Figure 7, Figure Hazy-Mamba exhibits more concentrated feature activation regions, especially in small targets and blurred areas. This is because HAFE mitigates blur interference through low-pass filtering and multi-scale convolution to optimize feature representation; MAG enhances feature extraction of small targets through bottleneck structures; and the knowledge distillation strategy dynamically adjusts the learning focus to reduce the impact of blur on the teacher network's predictions. These three components work together to improve feature quality. In contrast, the baseline model's scatter heatmap shows dispersed activations, especially in realistic blurred scenes, reflecting its limited adaptability to complex environments. This further confirms Hazy-Mamba's advantage in improving detection robustness.
[0129] 4. Conclusion
[0130] The Hazy-Mamba model is an end-to-end aerial target detection framework specifically designed for foggy scenarios. This framework integrates the HAFE and MAG modules, significantly improving feature extraction and detection capabilities under foggy conditions.
[0131] The HAFE module extracts dehazing features through low-pass filtering, multi-scale convolution, and contrast enhancement, thereby improving feature representation. The MAG module addresses the challenges of detecting blurred features and small objects through bottleneck structures and multi-scale feature extraction.
[0132] An adaptive haze perception distillation loss was developed using knowledge distillation techniques to transfer knowledge from a complex teacher network to a lightweight student network, thereby achieving a balance between superior accuracy and reduced complexity.
[0133] Experimental results show that the Hazy-Mamba model outperforms existing methods on the HazyDet dataset, achieving higher mAP with a significant reduction in parameters, thus making it more suitable for deployment on drones. This confirms the effectiveness of the HAFE module, MAG module, and adaptive haze perception distillation loss (knowledge distillation module).
Claims
1. A target detection system based on a state-space model and multi-scale attention, characterized in that, Including state-space models, which include: The Mamba backbone network is used to extract multi-level features. The Mamba backbone network includes the HazySimple module and the MS-VSSBlock module. The HazySimple module integrates the HAFE module, which enhances feature extraction in hazy environments. The MS-VSSBlock module integrates the MAG module to optimize the detection of small targets and blurred features. The HAFE module provides high-quality feature input to the MAG module, which further optimizes the detection of small targets and blurred features based on this input. The PAFPN neck structure receives multi-level features output from the backbone network, and finely fuses high-level semantic features with low-level spatial features through XSSBlock before outputting them to the detection head, thereby enhancing the richness of feature representation. The state-space model extracts multi-level features from the Mamba backbone network and then fuses these features into PAFPN. XSSBlock is responsible for refining and fusing high-level semantic features with low-level spatial features, and finally outputs the features to the decoupled head to obtain the detection results.
2. The target detection system according to claim 1, characterized in that, The HAFE module for input features The HAFE module first generates low-frequency features. To suppress noise, while extracting multi-scale features through convolution operations. ; then and The data is concatenated, then enhanced by a convolution and SiLU activation function, and finally merged with the residuals to generate the output. .
3. The target detection system according to claim 2, characterized in that, The HAFE module for input features The processing procedure satisfies: Low-frequency characteristics: ； Multi-scale features: , ； Output feature: 。 4. The target detection system according to claim 1, characterized in that, The MAG module for input features The MAG module extracts main branch features through bottleneck structures. And generate the remaining branch features. The main branch is then merged with the remaining branches, and multi-scale convolution is used to extract different receptive field features. and Next, these features are concatenated and fused through convolution operations to generate the output. .
5. The target detection system according to claim 4, characterized in that, The processing procedure of the MAG module satisfies: Remaining branch characteristics: ； Main branch characteristics: ， ； ； Fusion Output: ； ， 。 6. The target detection system according to any one of claims 1-5, characterized in that, It also includes a knowledge distillation module, which performs knowledge distillation through adaptive fog perception distillation loss, transferring the professional knowledge of complex teacher networks to lightweight student networks; the knowledge distillation module combines fog perception weights and KL divergence loss, where fog perception weights are implemented through dark channel prior knowledge.
7. The target detection system according to claim 6, characterized in that, The adaptive haze perception distillation loss function of the knowledge distillation module satisfies: Weighted classification loss: ； Weighted regression loss: ； Total loss: 。 8. The target detection system according to claim 6, characterized in that, Haze perception weight satisfy: ； ； 。 9. The target detection system according to any one of claims 1, 2, 3, 4, 5, 7, and 8, characterized in that, Also includes: The server is used to install the Hazy-Mamba model and perform subsequent model training and distillation. On the application side, it is used to acquire images and detect aerial targets based on the onboard student network, and feed the detection results back to the state space model for optimization training.
10. The target detection system according to claim 9, characterized in that, The operation method of the target detection system includes the following steps: S100 trains a high-precision complex state-space model and network to enhance detection performance; the HAFE module employs low-pass filtering, multi-scale convolution, and contrast enhancement techniques to capture multi-scale context and sharp features from fog-affected images. The MAG module employs a bottleneck structure to compress and enhance feature representations, thereby reducing computational complexity while preserving key information; multi-scale convolutions capture target features of different sizes, enhancing the saliency of small targets, while residual connections improve feature stability. S200 employs knowledge distillation technology, using a complex network as a teacher network to train a lightweight student network, thereby achieving a balance between accuracy and model complexity; S300 deploys the student network to the application to quickly detect targets in images, feeds the detection results back to the Hazy-Mamba model for optimization training, and then distributes them to the application after distillation via S200.