基于深度学习的图像超分辨率重建研究

毕业设计(论文)基于深度学习的图像超分辨率重建研究院别数学与统计学院专业名称信息与计算科学班级学号5133117学生姓名楚文玉指导教师张琨2017年06月10日东北大学秦皇岛分校毕业设计(论文)第I页基于深度学习的图像超分辨率重建研究摘要人工神经网络凭借其超强的学习能力，使得人工智能得到迅猛的发展，让人工神经网络再次成为研究热点。目前深度学习已经广泛应用于计算机视觉，语音处理，自然语言处理等各个领域，甚至在某些领域已经起到了主导作用。单一图像超分辨率重建技术旨在将一个低分辨率图像经过一系列算法重构出对应的高分辨率图像。目前比较成熟的方法有基于频域法，非均匀图像插值法，凸集投影法，最大后验概率法以及稀疏表示法。本文主要研究利用深度学习实现单一图像超分辨率重建。本文首先简要介绍人工神经网络的发展历程，然后介绍深度学习在计算机视觉方面的应用。然后介绍神经网络的一些理论知识，最后介绍深度学习中的卷积神经网络（CNN,ConvolutionalNeuralNetwork）。本文研究如何利用卷积神经网络实现超分辨率重建。卷积神经网络分为三层结构，第一层的作用是特征块的提取和表示，第二层的作用是非线性映射，第三层的作用是重建出高分辨率图像。本文首先将一个图像降采样再双三次插值作为低分辨率图像，作为卷积神经网络的输入，而高分辨率图像作为卷积神经网络的输出，利用卷积神经网络建立低分辨率，高分辨率之间的映射。最后针对该模型进行改进，再加入一层作为特征提取。最后利用深度学习框架TensorFlow实现上述模型。最后研究快速超分辨率重建模型，并针对模型层数和过滤器大小进行改进，与先前实验做比对。关键字：超分辨率重建，卷积神经网络，深度学习，TensorFlow东北大学秦皇岛分校毕业设计(论文)第II页ImageSuper-ResolutionUsingDeeplearningAuthor:ChuWen-yuTutor:ZhangKunAbstractArtificialNeuralNetworkbecauseofitsstrongabilitytolearn,getrapiddevelopmentofartificialintelligence,lettheArtificialNeuralNetworkbecometheresearchupsurgeagain.Deeplearninghasbeenwidelyusedincomputervision,speechprocessing,naturallanguageprocessingandsoon.Thesuper-resolution(SR)techniqueisdesignedtorefactoralow-resolutionimagethroughaseriesofalgorithmstoreconstructthecorrespondinghigh-resolutionimage.Currently,themethodoffrequencydomain,Non-uniformimageinterpolation,Projectionontoconvexset(POCS),Maximumaposterior(MPA)andsparsematrixmethodarethemorematuremethods.Thispapermainlyresearchestherealizationofsuper-resolution(SR)reconstructionusingdeeplearning.Inthisthesis,firstisabriefintroductionofthedevelopmentofartificialneuralnetwork,thenintroducestheapplicationofdeeplearningincomputervision.Withthatintroducessometheoreticalknowledgeofneuralnetwork,andfinallyintroducestheconvolutionneuralnetwork(CNN)indeeplearning.Thisarticlemainlyresearcheshowtousetheconvolutionneuralnetwork(CNN)togetthesuper-resolutionreconstruction.Theconvolutionneuralnetworkcontainsthreestructures,theeffectofthefirstlayerisPatchextractionandrepresentation,thesecondisthefunctionofNon-linearmapping,theroleofthethirdlayeristhehigh-resolutionimagereconstruction.Firsttodownscaleandbicubicinterpolationanimageasthelow-resolutionimagesastheinputoftheconvolutionneuralnetwork,andthehigh-resolutionimageastheoutputoftheconvolutionneuralnetwork,usingconvolutionneuralnetworkestablishedend-to-endmappingbetweenthelow-resolutionandhigh-resolution.Finally,themodelisimproved,andthenalayeris东北大学秦皇岛分校毕业设计(论文)第III页addedasfeatureextraction.ThemodelimplementsusingdeeplearningframeTensorFlow.Finally,learnmoreabouttheacceleratesuper-resolutionreconstructionmodelandimprovethemodellayerandfiltersize,andcomparewiththepreviousexperiment.KeyWords:Super-Resolution,Convolutionneuralnetwork,Deeplearning，TensorFlow东北大学秦皇岛分校毕业设计（论文）第IV页目录1绪论......................................................................................................................................11.1课题背景及意义...........................................................................................................................................11.2国内外研究现状...........................................................................................................................................21.3论文的内容结构...........................................................................................................................................32深度学习理论......................................................................................................................52.1人工神经网络理论......................................................................................................................................52.1.1神经网络基础理论......................................................................................................................62.1.2BP反向传播算法..........................................................................................................................92.1.3随机梯度下降法.........................................................................................................................122.2深度神经网络理论...................................................................................................................................132.2.1深度学习的核心思想...............................................................................................................132.2.2卷积神经网络..............................................................................................................................142.3TensorFlow简介.......................................................................................................................................193基于SRCNN的超分辨率重建算法研究.........................................................................213.1SRCNN模型简介.....................................................................................................................................213.2SRCNN模型的改进................................................................................................................................223.3改进的模型的实现...................................................