收稿日期:2021-05-18修回日期:2021-06-05第40卷第4期计算机仿真2023年4月文章编号:1006-9348(2023)04-0005-06改进卷积神经网络的SAR图像噪声抑制算法冯博迪,杨海涛∗,张长弓,高宇歌(航天工程大学航天信息学院,北京101416)摘要:SAR图像由于其主动成像的特点,不可避免的存在相干斑噪声,噪声的存在让SAR图像的后续解译处理变得较为复杂,为了便于SAR图像的解译处理和广泛应用,改进了一种基于CNN的斑点噪声抑制算法,使用Inception结构和扩张卷积来增大网络的感受野,使用非对称卷积组取代传统对称卷积来增强网络的特征提取能力,同时引入跳跃结构进行残差学习,将仿真数据输入网络,学习干净图像和噪声图像之间的映射关系,使用常用的评价指标对网络进行评估并与其它的噪声抑制算法进行对比,实验结果表明,改进的算法具有较好的去噪效果,对比其它去噪算法,上述方法不仅可以有效去除斑点噪声,并且能够较好的保留纹理信息。关键词:图像去噪;卷积神经网络;残差学习中图分类号:TP183文献标识码:BImprovedConvolutionalNeuralNetwork-BasedSARImageCoherentSpeckleNoiseSuppressionAlgorithmFENGBo-di,YANGHai-tao∗,ZHANGChang-gong,GAOYu-ge(SchoolofSpaceInformation,SpaceEngineeringUniversity,Beijing101416,China)ABSTRACT:TheexistenceofnoisemakesthesubsequentdecompressionprocessingofSARimagesmorecomplicat-ed.InordertofacilitatethedecompressionprocessingandwideapplicationofSARimages,thispaperimprovesaCNN-basedspecklenoisesuppressionalgorithm.TheInceptionstructureanddilationconvolutionwereusedtoin-creasetheperceptualfieldofthenetwork.Theasymmetricconvolutiongroupwasusedtoreplacethetraditionalsym-metricconvolutiontoenhancethefeatureextractionabilityofthenetwork,andatthesametime,thejumpstructurewasintroducedforresiduallearning,thesimulationdatawasinputintothenetwork,themappingrelationshipbetweencleanandnoisyimageswaslearned,andthenetworkwasevaluatedusingcommonevaluationmetricsandcomparedwithothernoisesuppressionalgorithms.Theexperimentalresultsshowthattheimprovedalgorithminthispaperhasabetterdenoisingeffect,andcomparedwithotherdenoisingalgorithms,theabovemethodcannotonlyef-fectivelyremovespecklenoise,butalsocanretainbettertextureinformation.KEYWORDS:Imagedenoising;Convo...
