融合隐藏状态压缩与空间依赖感知的Mamba-Transformer雾天目标检测方法OA
Mamba-Transformer foggy object detection method integrating hidden state compression and spatial dependency perception
在复杂天气与动态场景下,小目标检测易受到雾霾干扰和特征弱化的影响,现有方法在实时性和鲁棒性方面仍存在不足.为此,本文旨在构建一种兼具高效性与强鲁棒性的检测模型,以提升雾天小目标检测性能.本文提出基于Mamba-Transformer的混合状态空间检测网络HS-MambaDet.首先设计融合隐藏状态压缩与状态空间建模的HSM-SSD骨干结构以降低计算复杂度;随后引入包含高频感知模块与空间依赖感知模块的多尺度频空融合器(MFSF),强化小目标的细节表达与空间上下文建模.整体框架采用CNN、Transformer与Mamba的混合架构,分别实现局部特征提取、全局上下文建模与线性递归结构的高效结合.在RTTS与CityScapes数据集上的大量实验表明,HS-MambaDet在精度与效率上均优于当前主流模型.完整模型在RTTS上的预测精度为87.3%、召回率为73.1%、mAP@0.5为81.2%、mAP@(0.5～0.95)为51.0%,分别较对比模型最高提升3.8%、3.9%、3.6%和3.3%;同时推理时间为0.26s,保持了良好的实时性.此外,在小目标场景中,本模型的mAP@0.5提升最多达4.4%,并在跨雾强度测试中表现出更强的泛化能力.HS-MambaDet通过引入高效的隐藏状态压缩机制与多尺度频空融合结构,有效增强了雾天条件下对小目标的细节感知和空间建模能力,在检测精度、鲁棒性与推理效率上均取得显著优势,为动态恶劣环境中的实时目标检测提供了一种可行且高效的解决方案.
Object detection in foggy and dynamic environments remains challenging,especially for small objects whose features are easily degraded by low visibility.Existing methods often suffer from insufficient robustness and limited real-time performance.This study aims to develop a high-efficiency and highly robust detection framework tailored for foggy small-object detection.We propose HS-MambaDet,a hybrid state-space detection network built upon the Mamba-Transformer architecture.A hidden-state compression-enhanced backbone(HSM-SSD)is designed to reduce computational overhead while preserving global dependency modeling.Additionally,a Multi-scale Frequency-Spatial Fusion(MFSF)module—consisting of a high-frequency perception branch and a spatial dependency modeling branch—is introduced to enhance fine-grained details and long-range spatial relations for small objects.The overall framework integrates CNN,Transformer,and Mamba components to jointly capture local textures,global context,and efficient sequential dependencies.Extensive experiments on the RTTS and CityScapes datasets demonstrate that HS-MambaDet outperforms mainstream SOTA models in both accuracy and efficiency.On RTTS,the complete model achieves 87.3%precision,73.1%recall,81.2%mAP@0.5,and 51.0%mAP@(0.5～0.95),exceeding the best comparison baselines by up to 3.8%,3.9%,3.6%,and 3.3%,respectively.The inference time remains low at 0.26 s,ensuring real-time capability.In small-object scenarios,HS-MambaDet improves mAP@0.5 by up to 4.4%,and cross-domain tests under varying fog intensities further verify its superior generalization and robustness.By combining hidden state compression with multi-scale frequency-spatial fusion,HS-MambaDet effectively enhances fine-detail perception and spatial dependency modeling under foggy conditions.The proposed framework achieves strong performance in accuracy,robustness,and inference speed,offering a practical and efficient solution for real-time object detection in adverse weather and dynamic environments.
陈悦;许锋;宋京昊
中国刑事警察学院 公安信息技术与情报学院,辽宁 沈阳 110854中国刑事警察学院 公安信息技术与情报学院,辽宁 沈阳 110854中国刑事警察学院 公安信息技术与情报学院,辽宁 沈阳 110854
信息技术与安全科学
目标检测Mamba高频感知融合隐藏状态压缩
object detectionMambahigh-frequency perceptionfusion hidden state compression
《液晶与显示》 2026 (2)
222-239,18
十三五国家重点研发技术项目(No.2017YFC0821004)2022辽宁省教育厅基本科研重大攻关项目(No.LJKZZ20220007)中央高校基本科研业务重大培育项目(No.3242022004)Supported by National Key R&D Program of China(13th Five-Year Plan)(No.2017YFC0821004)2022 Major Fundamental Research Project of the Liaoning Provincial Department of Education(No.LJKZZ20220007)Major Cultivation Project for Fundamental Research in Central Universities(No.3242022004)
评论