国外电子测量技术北大中文核心期刊DOI:10.19652/j.cnki.femt.2204312融合图像局部和退化表征信息的盲超分辨重建*刘建军郝敏钗李建朝胡雪花(河北工业职业技术大学智能制造学院石家庄050091)摘要:针对假设的退化模型与实际模型不一致时图像超分辨性能显著降低的问题,提出了一种融合图像空间局部和退化表征信息的深度卷积神经网络模型。首先对低分辨图像提取初始特征和退化表达量;然后构建级联的空间局部信息和退化信息模块以及特征融合块,这些模块进一步级联组成特征变换子网络;最后,利用反卷积层得到高分辨率图像。在基准测试数据集上的实验表明,当高斯核宽度不为0时,算法在采样因子为×2和×4的盲超分辨重建中均取得了较当前主流算法更高的峰值信噪比值(PSNR),其中×2盲超分辨时最高的PSNR值为37.56,×4盲超分辨时最高的PSNR值为31.87,并且与主流算法相比也有较高的重建效率和较好重建视觉效果。关键词:盲超分辨;级联结构;卷积神经网络;深度学习中图分类号:TN391文献标识码:A国家标准学科分类代码:520.604Blindsuper-resolutionreconstructionbasedonfusionoflocalinformationanddegradationrepresentationofimageLiuJianjunHaoMinchaiLiJianchaoHuXuehua(CollegeofIntelligentManufacturing,HebeiVocationalUniversityofIndustryandTechnology,Shijiazhuang050091,China)Abstract:Aimingattheproblemthatthesuper-resolutionperformanceofimageissignificantlyreducedwhentheassumeddegradationmodelisinconsistentwiththeactualmodel,adeepconvolutionneuralnetworkmodelintegratingthelocalanddegradationrepresentationinformationofimagespaceisproposed.First,theinitialfeaturesanddegradedexpressionsareextractedfromthelow-resolutionimage,andthenthecascadedspatiallocalinformationanddegradedinformationmodulesandfeaturefusionblocksareconstructed.Thesemodulesarefurthercascadedtoformthefeaturetransformationsubnetwork.Finally,thehigh-resolutionimageisobtainedbyusingthedeconvolutionlayer.Theexperimentsonthebenchmarktestdatasetshowthatthealgorithmachieveshigherpeaksignal-to-noiseratio(PSNR)thanthecurrentmainstreamblindsuper-resolutionalgorithmsforboth×2and×4samplingfactorswhentheGaussiankernelwidthisnot0,withthehighestPSNRvalueis37.56for×2blindsuper-resolutionand31.87for×4blindsuper-res...
