基于VAE-RefineNet算法流程的GPR杂波抑制和目标成像

Clutter suppression of GPR B-scan and target imaging based on VAE-RefineNet algorithm process

在探地雷达勘探中,地下层位和杂波会对管线异常双曲线波形产生的干扰,为了提高目标成像质量,本文提出了一种基于混合损失函数的VAE和RefineNet网络算法流程用于优化GPR剖面.首先,使用VAE网络消除剖面中的地层信息,将输入数据通过编码映射到潜在空间,再从潜在空间中随机抽取样本并解码,以生成不含地层信息的剖面;然后,将该剖面输入RefineNet网络中,对剖面中的杂波干扰进行抑制,RefineNet网络通过增加混合注意力模块、残差卷积单元、像素混洗、链式残差池化和多尺度金字塔模块,提高了网络对于细节特征的捕获能力,在有效抑制杂波干扰的同时,还能关注到目标信号,提高目标信号连续性并且增强目标信号.通过数值模拟中不同损失函数的处理效果对比,验证了本文所提出算法流程对于处理GPR剖面的有效性和适应性.并且成功应用于实测资料中,提高了目标成像质量,使得目标异常体的定位信息更加准确,对实际工程应用具有指导意义.

In Ground Penetrating Radar(GPR)exploration,subsurface layers and clutter can interfere with the anomalous hyperbolic waveforms of pipelines.In order to improve target imaging quality,this paper proposes a hybrid loss function-based VAE and RefineNet network algorithm process for optimizing GPR profiles.First,the stratigraphic information in the profile is eliminated using the VAE network,and the input data is mapped to the potential space by coding,then random samples are taken from the potential space and decoded to generate a profile without stratigraphic information;then,the profile is input into the RefineNet to suppress the clutter interference in the profile,RefineNet improves the network's ability to capture detailed features,focus on the target signal,improve the target signal continuity and enhance the target signal while effectively suppressing clutter interference by adding hybrid attention module,residual convolution unit,pixel shuffler,chained residual pooling and multiscale pyramid module.The effectiveness and adaptability of the algorithm process proposed in this paper for processing GPR profiles are verified by comparing the processing effects of different loss functions in numerical simulations.And it has been successfully applied to the measured data,improving the imaging quality of the target and making the positioning information of the target anomaly more accurate,which has guiding significance for practical engineering applications.