pixelLabelImageDatastore
(删除)语义分割数据存储网络
描述
pixelLabelImageDatastore将在将来的版本中被删除。使用imageDatastore和pixelLabelDatastore对象和结合函数来代替。
创建
语法
描述
pximds= pixelLabelImageDatastore (gTruth)groundTruth对象或数组的groundTruth对象。使用输出pixelLabelImageDatastore对象的深度学习工具箱™函数trainNetwork(深度学习工具箱)训练卷积神经网络的语义分割。
pximds= pixelLabelImageDatastore (洛桑国际管理发展学院,pxds)洛桑国际管理发展学院是一个ImageDatastore对象代表了训练网络的输入。pxds是一个PixelLabelDatastore对象,该对象代表所需的网络输出。
pximds= pixelLabelImageDatastore (___,名称,值)DispatchInBackground和OutputSizeMode属性。对于二维数据,还可以使用指定名称-值对ColorPreprocessing,DataAugmentation,OutputSize增加属性。您可以指定多个名称-值对。在报价附上每个属性的名字。
例如,pixelLabelImageDatastore (gTruth PatchesPerImage, 40)创建一个随机生成的像素标签图像数据存储40补丁从每个地面实况对象gTruth。
输入参数
属性
对象的功能
结合 |
合并来自多个数据存储的数据 |
countEachLabel |
数出现的像素或盒标签 |
hasdata |
确定数据可用来读 |
partitionByIndex |
分区pixelLabelImageDatastore根据指数 |
预览 |
预览数据存储中数据的子集 |
读 |
从数据存储中读取数据 |
readall |
读取所有数据存储中的数据 |
readByIndex |
指定的索引读取数据pixelLabelImageDatastore |
重置 |
数据存储重置为初始状态 |
洗牌 |
返回版本的数据存储 |
变换 |
变换数据存储 |
例子
火车一个语义分割网络
加载训练数据。
dataSetDir = fullfile (toolboxdir (“愿景”),“visiondata”,“triangleImages”);imageDir = fullfile (dataSetDir,“trainingImages”);labelDir = fullfile (dataSetDir,“trainingLabels”);
创建一个图像的图像数据存储。
imd = imageDatastore (imageDir);
创建一个pixelLabelDatastore地面真理像素标签。
一会= [“三角形”,“背景”];labelIDs = (255 0);一会,pxds = pixelLabelDatastore (labelDir labelIDs);
想象训练图像和地面实况像素标签。
我=读(imd);C =阅读(pxds);I = imresize (5);L = imresize (uint8 (C {1}), 5);imshowpair (L,我“蒙太奇”)
创建一个语义分割网络。这个网络使用一个简单的语义分割网络基于一个将采样和upsampling设计。
numFilters = 64;filterSize = 3;numClasses = 2;32 32层= [imageInputLayer ([1]) convolution2dLayer (filterSize numFilters,“填充”(1)reluLayer) maxPooling2dLayer (2,“步”(2)convolution2dLayer filterSize numFilters,“填充”(1)reluLayer) transposedConv2dLayer (4 numFilters“步”2,“种植”1);numClasses convolution2dLayer(1日);pixelClassificationLayer softmaxLayer () ()];
设置培训选项。
选择= trainingOptions (“个”,…“InitialLearnRate”1 e - 3,…“MaxEpochs”,100,…“MiniBatchSize”,64);
结合图像和像素标签数据存储进行训练。
pxds trainingData =结合(imd);
培训网络。
网= trainNetwork (trainingData层,选择);
培训在单CPU。初始化输入数据规范化。| = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = | | | |时代迭代时间| Mini-batch | Mini-batch |基地学习| | | | (hh: mm: ss) | | |损失精度率| | = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = | | 1 | 1 |就是58.11% | | 1.3458 | 0.0010 | | 17 | 50 | 00:00:12 | 97.30% | 0.0924 | 0.0010 | | 100 | | 00:00:24 | 98.09% | 0.0575 | 0.0010 | | 150 | | 00:00:37 | 98.56% | 0.0424 | 0.0010 | 67 | 200 | | 00:00:49 | 98.48% | 0.0435 | 0.0010 | 84 | 250 | | 00:01:02 | 98.66% | 0.0363 | 0.0010 | 100 | 300 | | 00:01:14 | 98.90% | 0.0310 | 0.0010 | | = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = |培训完成:达到最终的迭代。
读取和显示测试图像。
testImage = imread (“triangleTest.jpg”);imshow (testImage)
段测试图像和显示结果。
C = semanticseg (testImage,净);B = labeloverlay (testImage C);imshow (B)
提示
的
pixelLabelDatastorepxds和imageDatastore洛桑国际管理发展学院位于一个文件夹的文件存储在辞典编纂的秩序。例如,如果您有十二个文件命名“file1.jpg”,“file2.jpg”、…“file11.jpg”,“file12.jpg”在这个订单,那么文件存储:文件存储在一个单元阵列读取存储在相同的顺序。“file1.jpg”“file10.jpg”“file11.jpg”“file12.jpg”“file2.jpg”“file3.jpg”…“file9.jpg”如果订单的文件
pxds和洛桑国际管理发展学院是不一样的,那么你可能会遇到当你读不匹配和相应的地面实况图像标签数据使用pixelLabelImageDatastore。如果发生这种情况,然后重命名像素标签文件,以便他们有正确的顺序。例如,重命名“file1.jpg”、…“file9.jpg”来“file01.jpg”、…“file09.jpg”。提取语义分割的数据
groundTruth生成的对象贴标签机视频,可以使用pixelLabelTrainingData函数。
