【作者】 刘勇；

【导师】 仲维；

【作者基本信息】 大连理工大学 ， 软件工程， 2025， 硕士

【摘要】 红外与可见光图像融合技术作为多源图像融合领域的重要研究方向,通过有效结合红外图像的热辐射特性和可见光图像丰富的纹理信息,已成为提升智能感知系统性能的关键手段。但是,由于不同传感器位置和成像原理存在较大差异,红外与可见光图像之间的几何错位、融合图像信息丢失、成像模糊等仍是亟待解决的问题。本文基于深度学习技术,对红外与可见光图像的配准和融合展开深入研究,主要工作概况如下:(1)针对红外与可见光图像因偏移、扭曲等形变导致的几何错位问题,本文提出一种基于模态转换与多尺度细化配准的图像配准方法。首先,通过引入跨模态风格迁移网络,将可见光图像转换为伪红外图像,从而将跨模态配准问题转化为近似的单模态配准任务,降低了因模态差异产生的图像配准误差。其次,本文采用多尺度细化配准策略,通过设计共享多尺度特征提取器和两个由粗到细的变形场估计模块,逐级细化得到准确的变形场。最后,通过变形场对待配准红外图像进行重采样,实现红外图像与可见光图像的精准对齐。实验结果表明,本文配准方法能够实现红外与可见光图像几何结构的精细化对齐,有效降低图像配准误差,相比其他代表性方法具有更高的配准精度和鲁棒性。(2)在红外与可见光图像融合方面,本文提出了一种基于多尺度特征表示与注意力引导的图像融合方法,旨在解决现有融合方法中因全局特征建模不足和跨域特征交互有限导致的融合图像信息丢失、成像模糊问题。首先,通过并行化空洞卷积实现不同尺度下的两种图像特征的提取,充分捕捉图像的全局和局部信息。其次,通过基于轻量化设计的自注意力和交叉注意力机制的特征融合模块,分别进行图像同一域内全局上下文建模和跨域间的特征交互,生成兼具两种模态信息的融合特征表示。这种轻量化设计的注意力机制,在降低计算复杂度的同时保持了模型性能。最后,通过卷积重建单元实现融合特征重建,生成最终的融合图像。实验结果表明,本文图像融合方法相比其他代表性方法,能够有效避免图像信息损失,提高融合图像的视觉效果,其丰富的纹理细节和清晰的目标轮廓,进一步提高了目标检测性能。更多还原

【Abstract】 As an important research direction of multi-source image fusion,infrared and visible image fusion technology has become a key means to enhance the performance of intelligent sensing systems by effectively combining the thermal radiation characteristics of infrared images and the rich texture information of visible light images.However,geometric misalignment between infrared and visible images,loss of information in fused images,and imaging blurring are still problems that need to be solved due to the large differences in different sensor positions and imaging principles.In this paper,based on deep learning technology,we carry out in-depth research on the alignment and fusion of infrared and visible images,and the main work is summarized as follows:(1)Aiming at the geometric misalignment of infrared and visible images due to displacement,distortion and other deformation,this paper proposes an image registration method based on modal transformation and multi-scale refinement registration.First,by introducing a cross-modal style transformation network,the visible image is converted into a pseudo-infrared image,thus transforming the cross-modal alignment problem into an approximate unimodal alignment task,and reducing the image registration error due to modal differences.Secondly,this paper adopts a multi-scale refinement registration strategy,and the accurate deformation field is obtained by designing a shared multi-scale feature extractor and two coarse-to-fine deformation field estimation modules for step-by-step refinement.Finally,the infrared image to be registered is resampled by the deformation field to realize the accurate alignment of the infrared image with the visible image.The experimental results show that the registration method in this paper can realize the fine alignment of the geometrical structure of infrared and visible images,effectively reduce the image registration error,and have higher registration accuracy and robustness than other representative methods.(2)In terms of infrared and visible image fusion,this paper proposes an image fusion method based on multi-scale feature representation and attention guidance,aiming at solving the problems of fused image information loss and imaging blurring caused by insufficient global feature modeling and limited cross-domain feature interaction in the existing fusion methods.First,the extraction of two image features at different scales is achieved by parallelized null convolution,which fully captures the global and local information of the image.Second,through the feature fusion module based on the lightweight design of self-attention and cross-attention mechanisms,the global context modeling within the same domain of the image and the cross-domain feature interaction are carried out respectively,to generate the fused feature representation with both modal information.This lightweight design of the attention mechanism reduces the computational complexity while maintaining the model performance.Finally,the fused feature reconstruction is realized by the convolutional reconstruction unit to generate the final fused image.The experimental results show that the image fusion method in this paper can effectively avoid the loss of image information and improve the visual effect of the fused image compared with other representative methods,and its rich texture details and clear target contours further improve the target detection performance.更多还原