基于RIME-VDSR神经网络的声场超分辨率重建

打开文本图片集

中图分类号：TP183；0426.4 文献标识码：A 文章编号：2096-4706（2026）03-0045-07

Super-resolution Reconstruction of Sound Field Based on RIME-VDSR Neural Network

JIA Hui¹， WANG Xun234， LIANG Shengde³， GAO Liru²， LOU Fengfei² （1.NalangNine-Year School in Gansu Province，Zhuoni 747602，China;

2.SchoolofAeronautics，ShanghaiDianjiUniversityShanghai 201306，China;3.SchoolofEnergy andPowerEnginering，

GansuMinzuNormalUiversityHezuo747ooo，China;4.StateKeyLaboratoryfcousticsandarineInformationtiute of Acoustics， Chinese Academy of Sciences， Beijing 100190， China）

Abstract： This paper investigates the reconstruction problem of ultrasonic fields in liquids based on a Very Deep SuperResolution （VDSR） Deep Neural Network.The COMSOL-MATLAB co-simulation method is adopted to simulate the sound fields intheliquidundertheradiationof transducers withdiferent positionsand diferentoperating frequencies.The simulation dataare saved toconstructadataset.Itconstructs a VDSRDepNeuralNetwork，integratestheRIMEotimizationalgorithm， andutilizes thedatasettocompletetheneuralnetworktrainingandtesting.TheresearchfindsthatusingtheRIMEoptimization algorithmcan improve thereconstruction precision.Furthermore，the paperanalyzes thereconstructionoflow-resolutionsound fields obtained through down-sampling with varying scaling factors.Itreveals that reconstruction accuracy gradually decreases asthe scalingfactorseduceandthereconstructionacuracyofhigh-frequencysoundfieldismoresensitive tothescalingfactor thanthatoflow-frequency sound field.Finaly，the methodiscompared with the existing sound feldreconstructionmethod. The results showthatthereconstruction precisionof the proposed method is slightlybeterthantheexisting methods forlowfrequency sound field，andtheadvantages of the proposed method are more significant for high-frequencysound field.

Keywords： neural network; super-resolution; finite element simulation; sound field reconstruction

0 引言

超声波指的是频率在 20kHz 以上的声波[]，它具有穿透能力强、指向性好、分辨率高和无辐射等优点。（剩余13385字）