【作者】 高洁；

【导师】 张欣欣；

【作者基本信息】 山东大学 ， 软件工程， 2025， 硕士

【摘要】 图像质量退化问题是图像处理领域的重要研究方向,其中反射光干扰是导致图像退化的主要原因之一。反射现象广泛存在于日常生活中,例如透过玻璃拍摄时,反射光会与目标图像混合,导致背景信息被遮挡、图像对比度下降、细节模糊等现象,严重影响后续视觉任务的性能。去除图像中的反射成分,恢复清晰的背景,具有重要的理论意义和应用价值。随着深度学习的快速发展,图像去反光取得了显著进展。本文基于现有的去反光方法,针对当前研究在全局信息建模和网络泛化能力方面的不足进行深入研究,主要贡献包括:(1)提出一种语义信息引导的双流自注意力反射去除网络,旨在解决单张图像去反光任务中存在的两个关键问题:一是现有反射建模在真实场景中的适应性差,二是现有方法感受野有限、长距离依赖建模能力不足,难以充分利用有效的上下文信息。本方法引入一个简单通用的残差预测模型,将反光图像建模为传输层和残差层的叠加,增强模型对合成和真实反光数据中反射成分的区分能力。在此基础上,本方法构建了一个两阶段双流网络(DRRformer):第一阶段通过语义特征融合网络,为反光去除提供多层次的全局语义上下文信息;第二阶段通过共享参数的双流注意力模块(DSA)执行进一步的特征编码与解码。DSA模块结合具有空间感知特性的注意力机制(MaSA),有效提升了网络的全局上下文建模能力和长距离感知能力,从而实现更精确的反光图像复原。此外,为了进一步提升模型的泛化能力,本研究设计了一种针对反光去除任务的数据增强策略。该策略对合成数据和真实数据分别施加运动模糊和通道不一致亮度增强,以模拟真实世界中复杂的图像退化和光照变化,增强模型对不同场景的适应能力,使其表现更加稳健。实验结果表明,所提出的方法在五个被广泛采用的测试集上具有优越性,平均 PSNR、SSIM、LMSE 指标分别为 25.929 dB、0.917、0.0036,均优于最优对比方法,在结构保持方面性能良好,同时保持了较低的网络参数量。(2)引入基于锐度感知与梯度匹配的泛化增强方法与基于图像配准的数据集扩充方案,以进一步提升模型在单张图像去反光任务中的鲁棒性与泛化能力。前者旨在解决反光数据多样性高且具有较大的分布差异造成的模型训练过程中的参数更新方向不一致问题,并降低参数对扰动的敏感度,提高模型的泛化性;后者旨在克服现有真实训练集规模有限而其收集过程中需要采取物理手段精准对齐的困难,简化反光图像数据的采集过程。本方法结合锐度感知和梯度匹配方法,构建混合损失以引导参数向更稳定且一致的方向更新;对于训练数据采集,本方法基于反光图像特性,引入SIFT-Flow方法进行不对齐数据集配准,为传统依赖物理手段的反光图像对采集提供新的解决方案,降低了数据采集的难度和成本。在两个基准模型上的实验结果表明,所提出的混合损失在提升网络泛化能力方面具有明显效果。与仅使用对齐不变损失的训练方法相比,使用配准后的图像训练的模型在多个测试集上的去反光效果更优,表明配准策略在真实场景中具有广泛的应用潜力,为去反光任务提供了一种新的技术路径。更多还原

【Abstract】 Image degradation is a fundamental research topic in the field of image processing,with reflection interference stands out as a significant factor.Reflections frequently occur in daily imaging scenarios—for example,when photographing through glass,the reflected light is superimposed onto the scene,resulting in occlusion of background content,reduced image contrast,and loss of fine details.Such degradation severely affects the performance of subsequent high-level vision tasks.Therefore,removing reflection components from images and recovering clean background scenes is of both theoretical importance and practical value.With the rapid development of deep learning,single-image reflection removal has witnessed remarkable progress.This thesis conducts an in-depth study of the limitations in existing SIRR methods,particularly in global information modeling and network generalization.The main contributions are summarized as follows:(1)A semantic information-guided dual-stream self-attention reflection removal network is proposed to address two key issues in single-image reflection removal tasks:first,the poor adaptability of existing reflection modeling in real-world scenarios,and second,the limited receptive field and insufficient long-range dependency modeling capability of existing methods,which makes it difficult to fully utilize effective contextual information.The proposed method introduces a simple and general residual prediction model,modeling the reflected image as the sum of a transmission layer and a residual layer,enhancing the model’s ability to distinguish reflection components in both synthetic and real-world reflected data.Based on this,a two-stage dual-stream network(DRRformer)is constructed:in the first stage,a semantic feature fusion network provides multi-level global semantic context information for reflection removal;in the second stage,a dual-stream attention module(DSA)with shared parameters performs further feature encoding and decoding.The DSA module combines an attention mechanism with spatial awareness(MaS A),effectively improving the network’s global context modeling ability and long-range perception,leading to more precise reflection image restoration.Additionally,to further enhance the model’s generalization ability,a data augmentation strategy specifically for reflection removal tasks is designed.This strategy applies motion blur and channel inconsistent brightness enhancement to synthetic and real data,respectively,to simulate complex image degradation and lighting changes in the real world,enhancing the model’s adaptability to different scenes and making its performance more robust.Experimental results demonstrate that the proposed method outperforms existing methods on five widely used benchmark datasets,with average PSNR,SSIM,and LMSE values of 25.929 dB,0.917,and 0.0036,respectively,surpassing the best comparison methods.It performs well in structural preservation while maintaining a relatively low number of network parameters.(2)A sharpness-aware and gradient-matching-based generalization enhancement method,along with a registration-based data augmentation strategy,is proposed to further improve the model’s robustness and generalization ability in the single-image reflection removal task.The former addresses the issue of parameter update inconsistency caused by the high diversity and large distribution differences of reflection data,reducing the model’s sensitivity to disturbances and enhancing its generalization capability.The latter solves the challenge of limited real training set size and the difficulties associated with precise alignment through physical means,simplifying the data collection process for reflection images.By combining sharpness-awareness and gradient matching,this method constructs a hybrid loss function that guides the model parameters toward more stable and consistent directions.For training data collection,based on the unique characteristics of reflection images,the SIFT-Flow method is introduced to align non-registered datasets,offering an alternative to the traditionally used physical alignment methods for reflection image pair collection,reducing both the difficulty and cost of data acquisition.Experimental results on two benchmark models show that the proposed hybrid loss significantly enhances the network’s generalization ability.Compared to training methods using only alignment-invariant losses,models trained with registered images show better performance in reflection removal across multiple test sets.This demonstrates the broad potential of the registration strategy in real-world applications,providing a new technical approach for reflection removal tasks.更多还原