Interferometric phase denoising combining global context and fused attention

目的 干涉相位去噪是合成孔径雷达干涉测量(interferometric synthetic aperture radar,InSAR)技术中的关键环节,其效果对测量精度具有重要影响。针对现有的干涉相位去噪方法大多关注局部特征以及在特征提取方面的局限性,同时为了平衡去噪和结构保持两者之间的关系,提出了一种结合全局上下文与融合注意力的相位去噪网络 GCFA-PDNet (global context and fused attention phase denoising network)。方法 将干涉相位分离为实部和虚部依次输入到网络,先从噪声相位中提取浅层特征,再将其映射到由全局上下文提取模块和融合注意力模块组成的特征增强模块,最后通过全局残差学习生成去噪图像。全局上下文提取模块能提取全局上下文信息,具有非局部方法的优势;融合注意力模块既强调关键特征,又能高效提取隐藏在复杂背景中的噪声信息。结果 所提出的方法与对比方法中性能最优者相比,在模拟数据结果的平均峰值信噪比(peak signal to noise ratio,PSNR)和结构相似性(struc-tural similarity,SSIM)指标分别提高了 5.72% 和 2.94%,在真实数据结果的平均残差点减少百分比(percentage ofresidual point reduction,PRR)和相位标准偏差(phase standard deviation,PSD)指标分别提高了 2.01% 3.57%。结合定性与定量分析,所提出的方法优于其他 5 种不同类型的相位去噪方法。结论 提出的去噪网络较其他方法具有更强大的特征提取能力,此外由于关注全局上下文信息和强调关键特征,网络能够在增强去噪能力的同时保持原始相位细节。;Objective Interferometric is introduced by three types of inherent factors:1)system noise, such as thermal radar(SAR)speckle noise;2)decoherence problems, including baseline, temporal, spatial decoherence;3)signal processing errors, misregistration.The existence increases the difficulty unwrapping even causes process fail, thereby seriously interfering with final interferometric result.Therefore, a key link in SAR(InSAR)technology.Its effect has an important influence on accuracy measurement results.The existing algorithms still have many defects.First insufficient ability capture global contextual information.Some ignore information or only focus local derived from few pixels.They also lack information.This feature manifested unstable detail preservation results.Second, researchers pay dimension channel image result improve performance networks.However, they do not use dimensions combination.Third, high-level features extracted deep layers convolutional neural network rich semantic ambiguous details.In comparison, low-level shallow contain considerable pixel-level information.However, these are isolated one another;thus, cannot be fully used.Method Most methods features, limitations extraction.A called proposed solve problems while balancing relationship between structure preservation.This combines attention.The method separates interference into real imaginary parts inputs them network.First, phase.Then, mapped enhancement module composed extraction module.The concurrently.Finally, denoised generated through residual learning.Four modules four used whole network.The core block, which can extract advantages nonlocal methods.The fuses its two submodules:the block block.It emphasizes efficiently extracts hidden complex backgrounds.Result We present six experimental results:Goldstein, interferogram estimator(NL-InSAR), block-matching 3D(InSAR-BM3D), learning framework for SAR restoration coherence estimation (DeepInSAR), filtering network, GCFAPDNet.Different evaluation indicators selected different datasets evaluate disadvantages various objectively.For experiments simulated images, peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)are indicators.A large PSNR indicates little difference filtered clean phase.However, it does consider correlation pixels image.Therefore, SSIM employed assess images overall quality images.Compared comparative methods, improves average data results 2.94%, respectively.For phase, above evaluating because no noise-free reference.The number residues (NOR)and deviation(PSD)can objective judge subjectively images.NOR reflect suppress noise.The smaller NOR is, stronger is.PSD measure discrete degree distribution.The PSD value more concentrated distribution better is.In results, compared percentage reduction 3.57%, respectively.The visual observation shows that achieves best qualitative quantitative analyses indicate outperforms five other methods.Conclusion The designed study attention.Thus, certain related algorithms.It powerful than focuses features.Thus, maintain original details enhancing ability, achieving results.