航空兵器2026，Vol.33Issue(1)：44-53,10.DOI:10.12132/ISSN.1673-5048.2025.0137
基于Mamba的跨模态无人机热红外图像超分辨率算法
Cross-Modality UAV Thermal Images Super-Resolution Algorithm Based on Mamba
摘要
Abstract
Objective Unmanned aerial vehicle(UAV)platform equipped with sensors has become an essen-tial technical support in fields such as object detection,trajectory tracking,and military technology.However,constrained by sensor hardware performance and complex environmental imaging condi-tions,thermal images inherently suffer from low spatial resolution,strong noise interference,and weak edge contrast.Single image super resolution(SISR)technology is an effective way to solve this problem,but thermal image super resolution still faces many practical challenges due to its character-istics of high background uniformity and sparse high frequency details.Existing CNN methods are lim-ited by local receptive fields and cannot capture the long range spatial correlation of sparse targets in UAV aerial perspective images.While Transformer methods excel at capturing global context,their self-attention mechanism incurs quadratic computational complexity,making them inefficient for pro-cessing the high-resolution images typical in UAV applications.Visible images guided super resolu-tion methods fail to systematically solve the problems of modal difference alignment,inefficient long range dependency modeling and insufficient decoupling of structure-texture collaborative reconstruc-tion,resulting in unsatisfactory fine texture reconstruction effects.Therefore,it is of great practical significance to propose a super resolution algorithm that is adapted to the characteristics of UAV ther-mal images and can efficiently integrate cross modal information to improve reconstruction quality. Methods To produce better reconstruction results,this paper proposes a cross modality UAV ther-mal images super resolution algorithm based on Mamba(CM-Mamba),which constructs a dual branch model framework integrating three core modules:residual state space block(RSSB),cross modality feature integration framework(CMFIF)and dual channel collaborative reconstruction mecha-nism(DCCRM).Firstly,RSSB is designed to model the long range spatial dependency of thermal images efficiently.This module standardizes the input features through layer normalization,captures the long range dependency across spatial and frequency domains via spatial focal-state space model(SF-SSM),and dynamically adjusts channel features combined with channel attention mechanism.A local convolution layer is introduced after SF-SSM to make up for the loss of local information,and residual operation with scaling factor is used to ensure the stable transmission of features,which real-izes the efficient modeling of global dependency with linear computational complexity.Secondly,aim-ing at the feature misalignment caused by multi-scale feature redundancy and resolution mismatch in cross modal super resolution tasks,a three-stage CMFIF is constructed including feature interaction block(FIB),feature refinement block(FRB)and feature enhancement block(FEB).FIB extracts the differences between thermal and visible features through self-attention and cross attention mecha-nisms,and realizes the joint representation of dual modalities by channel concatenation to avoid the dilution of thermal radiation characteristics by visible texture;FRB completes cross-modal and cross-scale channel alignment,and refines the features in spatial and channel dimensions to integrate multi-scale information of dual modalities;FEB generates spatial weighting coefficients through average pool-ing and maximum pooling operations,suppresses multi-scale feature redundancy,and provides high quality feature input for subsequent reconstruction.Finally,DCCRM decouples the super-resolution reconstruction process into structure branch and appearance branch to solve the trade-off problem of structural distortion and texture blurring.The structure branch adopts edge-preserving smoothing mechanism,combined with reconstruction loss and adversarial loss,to restore the global structural contour of the image and eliminate high-frequency noise;the appearance branch introduces Gaussian sampling to expand the receptive field,establishes long range regional relationships through appear-ance flow mechanism,and synthesizes realistic high frequency texture details with the prior knowledge of the structure branch,and uses sampling correction loss,reconstruction loss and adversarial loss to ensure the semantic alignment and visual realism of the generated texture. Results and Discussions The experimental research is carried out on the VGTSR benchmark data-set containing 1 025 pairs of visible-thermal UAV aerial images,with 800 pairs for model training and 225 pairs for testing,covering typical scenes such as campus and urban streets under complex envi-ronmental conditions such as cloudy days and nights.Peak signal to noise ratio(PSNR)and struc-tural similarity(SSIM)are used as image quality evaluation indicators.Ablation experiments(Table 1)verify the effectiveness of each core module of the proposed algorithm:the introduction of RSSB improves the PSNR by 0.66 dB compared with the traditional Transformer component;the combina-tion of FIB,FRB and FEB realizes efficient cross-modal feature fusion;DCCRM further optimizes the collaborative reconstruction effect of global structure and local texture,and the comprehensive use of all modules makes the model reach the optimal performance.Comparative experiments with main-stream algorithms such as EDSR,Restormer,UGSR,MGNet and LTSR(Table 2)show that CM-Mamba achieves excellent performance in×4,×8 and×16 super-resolution tasks,with PSNR reach-ing 32.02 dB,26.74 dB and 22.59 dB respectively,and SSIM corresponding to 0.910 0,0.765 9 and 0.647 5,which are all higher than the existing state-of-the-art methods.Qualitative visual com-parison results(Fig.3,Fig.4,Fig.5)show that the thermal images reconstructed by CM-Mamba have sharp building contour edges,realistic road surface material texture details,and significantly improved detail in thermal sensitive areas such as road signs,which effectively solves the problems of blurring,artifact and misjudgment in the reconstruction results of existing methods.In addition,the model parameters of CM-Mamba are only 10.6 M,which has a good balance between reconstruction performance and computational complexity. Conclusions CM-Mamba effectively responds to the challenge of existing super resolution methods for UAV thermal images,and the three core innovative designs realize the efficient modeling of long range dependency,accurate alignment and fusion of cross-modal features,and decoupled reconstruc-tion of structure and texture respectively,which significantly improves the reconstruction quality of thermal images and provides high-quality data support for subsequent downstream visual tasks such as target detection and trajectory tracking.关键词
跨模态信息整合/热红外图像/超分辨率/无人机/状态空间模型Key words
cross-modality integration/thermal images/super-resolution/unmanned aerial vehicle/state-space module分类
军事科技引用本文复制引用
高颂,周冰婵,贾淼,马琪惠,金鑫,王欢..基于Mamba的跨模态无人机热红外图像超分辨率算法[J].航空兵器,2026,33(1):44-53,10.基金项目
航空科学基金项目(20240001042002) （20240001042002）
河南省科技攻关项目(252102211023) （252102211023）
河南省高等学校重点科研项目(25A120005) （25A120005）
