摘要
针对现有图像拼接技术因依赖特征匹配导致的伪影、失真等问题,提出一种基于无监督的图像拼接方法。该方法包含无监督图像配准和无监督图像融合两个阶段。在配准阶段,通过融合多尺度特征和高效局部注意力机制(efficient local attention,ELA)改进残差块,进行跨尺度特征融合和动态特征增强,并结合交叉注意力(cross-attention)构建交互感知增强模块,促进图像对之间特征信息的深度交互与融合。进一步提出多尺度渐进变换配准模块,通过分层优化策略逐步校准图像变换关系,显著提升对齐精度。在融合阶段,引入动态软掩膜预测机制,基于像素级连续权重学习,实现重叠区域的平滑过渡与细节保持。为支撑无监督训练,构建了涵盖复杂光照、多视差场景的真实图像拼接数据集。实验表明,相较于现有传统和深度学习拼接算法,该方法在PSNR、SSIM上分别达到27.31、0.84,且视觉上拼接效果更好,抗干扰能力更强。
To address the issues of artifacts and distortion in existing image stitching technologies caused by reliance on feature matching,this paper proposed an unsupervised image stitching method.The method consisted of two stages:unsupervised ima-ge registration and unsupervised image fusion.In the registration stage,the paper improved residual blocks by fusing multi-scale features and the efficient local attention mechanism(ELA)to perform cross-scale feature fusion and dynamic feature enhancement.Meanwhile,it constructed an interactive perception enhancement module by incorporating cross-attention to promote the deep interaction and fusion of feature information between image pairs.Additionally,the paper proposed a multi-scale progressive transformation registration module.This module adopted a hierarchical optimization strategy to gradually calibrate the image transformation relationship,which significantly improved the alignment accuracy.In the fusion stage,the paper introduced a dynamic soft mask prediction mechanism.Based on pixel-level continuous weight learning,this mechanism achieved smooth transitions and detail preservation in overlapping regions.To support unsupervised training,the paper constructed a real-world image stitching dataset covering complex lighting and multi-parallax scenes.Experiments show that compared with existing traditional and deep learning stitching algorithms,the method achieves PSNR and SSIM values of 27.31 and 0.84,respectively.Visually,it provides better stitching effects and stronger anti-interference capabilities.
作者
赵文龙
王列伟
王军华
杨吉祥
Zhao Wenlong;Wang Liewei;Wang Junhua;Yang Jixiang(School of Computer Science&Engineering,Anhui University of Science&Technology,Huainan Anhui 232001,China;Nanjing Paiguang Intelligent Perception Information Technology Co.,Ltd.,Nanjing 210032,China)
出处
《计算机应用研究》
北大核心
2025年第12期3785-3792,共8页
Application Research of Computers
基金
广东省重点领域研发计划资助项目(2019B111105001)
中国铁路上海局集团有限公司科研计划资助项目(2022031)。
关键词
图像拼接
深度学习
图像配准
图像融合
无监督
image stitching
deep learning
image registration
image fusion
unsupervised
