语义分割使用深度学习

pretrainedURL =“https://ssd.mathworks.com/万博1manbetxsupportfiles/vision/data/deeplabv3plusResnet18CamVid.zip”;pretrainedFolder = fullfile (tempdir,“pretrainedNetwork”);pretrainedNetworkZip = fullfile (pretrainedFolder,“deeplabv3plusResnet18CamVid.zip”);如果~存在(pretrainedNetworkZip“文件”mkdir (pretrainedFolder);disp (“下载pretrained网络(58 MB)…”);websave (pretrainedNetworkZip pretrainedURL);结束解压缩(pretrainedNetworkZip pretrainedFolder)

pretrainedNetwork = fullfile (pretrainedFolder,“deeplabv3plusResnet18CamVid.mat”);data =负载(pretrainedNetwork);网= data.net;

类=字符串(net.Layers(结束). class)

我= imread (“highway.png”);

inputSize = net.Layers (1) .InputSize;我= imresize(我inputSize (1:2));

C = semanticseg(我,净);

提出= camvidColorMap;我= labeloverlay (C“Colormap”提出,“透明”,0.4);图imshow (B) pixelLabelColorbar(提出、类);

resnet18 ();

imageURL =“http://web4.cs.ucl.ac.uk/staff/g.brostow/MotionSegRecData/files/701_StillsRaw_full.zip”;labelURL =“http://web4.cs.ucl.ac.uk/staff/g.brostow/MotionSegRecData/data/LabeledApproved_full.zip”;outputFolder = fullfile (tempdir,“CamVid”);labelsZip = fullfile (outputFolder,“labels.zip”);imagesZip = fullfile (outputFolder,“images.zip”);如果~存在(labelsZip“文件”)| | ~存在(imagesZip“文件”mkdir (outputFolder) disp (“下载数据集16 MB CamVid标签…”);websave (labelsZip labelURL);解压缩(labelsZip fullfile (outputFolder“标签”));disp (“下载557 MB CamVid数据集图像…”);websave (imagesZip imageURL);解压缩(imagesZip fullfile (outputFolder“图片”));结束

imgDir = fullfile (outputFolder,“图片”,701 _stillsraw_full);imd = imageDatastore (imgDir);

I = readimage (imd, 559);I = histeq(我);imshow(我)

类= [“天空”“建筑”“极”“路”“路面”“树”“SignSymbol”“篱笆”“汽车”“行人”“自行车”];

labelIDs = camvidPixelLabelIDs ();

labelDir = fullfile (outputFolder,“标签”);pxds = pixelLabelDatastore (labelDir、类labelIDs);

C = readimage (pxds, 559);提出= camvidColorMap;我= labeloverlay (C“ColorMap”,提出);imshow (B) pixelLabelColorbar(提出、类);

台= countEachLabel (pxds)

频率= tbl.PixelCount /笔(tbl.PixelCount);栏(1:元素个数(类),频率)xticks(1:元素个数(类))xticklabels (tbl.Name) xtickangle (45) ylabel (“频率”)

[imdsTrain, imdsVal imdsTest、pxdsTrain pxdsVal, pxdsTest] = partitionCamVidData (imd, pxds);

numTrainingImages =元素个数(imdsTrain.Files)

numValImages =元素个数(imdsVal.Files)

numTestingImages =元素个数(imdsTest.Files)

%指定网络图像大小。这通常是一样的训练图像大小。图象尺寸= (720 960 3);%指定类的数量。numClasses =元素个数(类);%创建DeepLab v3 +。numClasses lgraph = deeplabv3plusLayers(图象尺寸,“resnet18”);

imageFreq =(资源。PixelCount。/ tbl.ImagePixelCount;classWeights =值(imageFreq)。/ imageFreq

pxLayer = pixelClassificationLayer (“名字”,“标签”,“类”tbl.Name,“ClassWeights”,classWeights);lgraph = replaceLayer (lgraph,“分类”,pxLayer);

%定义验证数据。dsVal =结合(imdsVal pxdsVal);%定义培训选项。选择= trainingOptions (“个”,…“LearnRateSchedule”,“分段”,…“LearnRateDropPeriod”10…“LearnRateDropFactor”,0.3,…“动量”,0.9,…“InitialLearnRate”1 e - 3,…“L2Regularization”,0.005,…“ValidationData”dsVal,…“MaxEpochs”30岁的…“MiniBatchSize”8…“洗牌”,“every-epoch”,…“CheckpointPath”tempdir,…“VerboseFrequency”2,…“阴谋”,“训练进步”,…“ValidationPatience”4);

dsTrain =结合(imdsTrain pxdsTrain);

xTrans = -10 [10];yTrans = -10 [10];dsTrain =变换(dsTrain @(数据)augmentImageAndLabel(数据、xTrans yTrans));

doTraining = false;如果doTraining[网,信息]= trainNetwork (dsTrain、lgraph选项);结束

I = readimage (imdsTest 35);C = semanticseg(我,净);

我= labeloverlay (C“Colormap”提出,“透明”,0.4);imshow (B) pixelLabelColorbar(提出、类);

expectedResult = readimage (pxdsTest 35);实际= uint8 (C);预期= uint8 (expectedResult);预计imshowpair(实际)

借据= jaccard (C, expectedResult);表(类、借据)

pxdsResults = semanticseg (imdsTest净,…“MiniBatchSize”4…“WriteLocation”tempdir,…“详细”、假);

指标= evaluateSemanticSegmentation (pxdsResults pxdsTest,“详细”、假);

metrics.DataSetMetrics

metrics.ClassMetrics

函数labelIDs = camvidPixelLabelIDs ()%返回标签id对应于每个类。%% CamVid数据集有32类。成11类%原SegNet训练方法[1]。%%的11类:%的“天空”“建筑”,“极”,“路”,“路面”、“树”、“SignSymbol”,%“栅栏”,“车”、“行人”,“骑自行车”。%% CamVid像素标签id作为RGB颜色值提供。集团成% 11类并返回它们的单元阵列M-by-3矩阵。的%原CamVid列出类名与每一个RGB值。请注意%,其他/无效类下面被排除在外。labelIDs = {…%的“天空”(128 128 128;…%的“天空”]%“建设”(000 128 064;…%“桥”128 000 000;…%“建设”064 192 000;…%的“墙”064 000 064;…%的“隧道”192 000 128;…%的“拱门”]%“极”(192 192 128;…%”Column_Pole”000 000 064;…%”TrafficCone”]%的道路(128 064 128;…%的“路”128 000 192;…%”LaneMkgsDriv”192 000 064;…%”LaneMkgsNonDriv”]%“路面”(000 000 192;…%“人行道”064 192 128;…%”ParkingBlock”128 128 192;…%”RoadShoulder”]%的“树”(128 128 000;…%的“树”192 192 000;…%”VegetationMisc”]%”SignSymbol”(192 128 128;…%”SignSymbol”128 128 064;…%”Misc_Text”000 064 064;…%”TrafficLight”]%“栅栏”(064 064 128;…%“栅栏”]%的“汽车”(064 000 128;…%的“汽车”064 128 192;…%”SUVPickupTruck”192 128 192;…%”Truck_Bus”192 064 128;…%“训练”128 064 064;…%”OtherMoving”]%“行人”(064 064 000;…%“行人”192 128 064;…%的“孩子”064 000 192;…%”CartLuggagePram”064 128 064;…%的“动物”]%“自行车”(000 128 192;…%“自行车”192 000 192;…%”MotorcycleScooter”]};结束

函数一会pixelLabelColorbar(提出)% colorbar添加到当前轴。colorbar是格式化%显示类名与颜色。甘氨胆酸colormap(提出)% colorbar添加到当前的图。c = colorbar (“对等”甘氨胆酸,);%为刻度线使用类名。c。TickLabels =一会;numClasses =大小(提出,1);%中心标记标签。c。蜱虫= 1 / (numClasses * 2): 1 / numClasses: 1;%去除刻度线。c。TickLength = 0;结束

函数提出= camvidColorMap ()%定义colormap CamVid所使用的数据集。提出= (128 128 128%的天空128 0 0%的建筑192 192 192%极128 64 128%的道路60 40 222%的人行道上128 128 0%的树192 128 128% SignSymbol64 64 128%的栅栏64 0 128%的车64 64 0%行人0 128 192%骑自行车];%(0 1)之间的正常化。提出255 =提出。/;结束

函数[imdsTrain, imdsVal imdsTest、pxdsTrain pxdsVal, pxdsTest] = partitionCamVidData (imd pxds)%分区CamVid数据通过随机选择60%的数据进行训练。的%休息是用于测试。%设置初始随机状态例如再现性。rng (0);numFiles =元素个数(imds.Files);shuffledIndices = randperm (numFiles);%使用60%的图像进行训练。numTrain =圆(0.60 * numFiles);trainingIdx = shuffledIndices (1: numTrain);%使用20%的图像进行验证numVal =圆(0.20 * numFiles);valIdx = shuffledIndices (numTrain + 1: numTrain + numVal);%使用其余的测试。testIdx = shuffledIndices (numTrain + numVal + 1:结束);%创建图像数据存储进行训练和测试。trainingImages = imds.Files (trainingIdx);valImages = imds.Files (valIdx);testImages = imds.Files (testIdx);imdsTrain = imageDatastore (trainingImages);imdsVal = imageDatastore (valImages);imdsTest = imageDatastore (testImages);%提取类和标签id信息。类= pxds.ClassNames;labelIDs = camvidPixelLabelIDs ();%建立像素标签数据存储进行训练和测试。trainingLabels = pxds.Files (trainingIdx);valLabels = pxds.Files (valIdx);testLabels = pxds.Files (testIdx);pxdsTrain = pixelLabelDatastore (trainingLabels、类labelIDs);pxdsVal = pixelLabelDatastore (valLabels、类labelIDs);pxdsTest = pixelLabelDatastore (testLabels、类labelIDs);结束

函数data = augmentImageAndLabel(数据、xTrans yTrans)%增加图像和像素标签使用随机反射和图像%的翻译。为i = 1:尺寸(数据,1)tform = randomAffine2d (…“XReflection”,真的,…“XTranslation”xTrans,…“YTranslation”,yTrans);%中心视图输出空间中图像的中心%允许翻译将视图的输出图像。溃败= affineOutputView(大小(数据{我1}),tform,“BoundsStyle”,“centerOutput”);%扭曲图像和像素标签使用相同的变换。{我1}= imwarp(数据{1}我,tform,“OutputView”,溃败);{我2}= imwarp(数据{2}我,tform,“OutputView”,溃败);结束结束