• documentation
main content

semantic segmentation of multispectral images using deep learning -凯发k8网页登录

this example uses:

this example shows how to perform semantic segmentation of a multispectral image with seven channels using u-net.

semantic segmentation involves labeling each pixel in an image with a class. one application of semantic segmentation is tracking deforestation, which is the change in forest cover over time. environmental agencies track deforestation to assess and quantify the environmental and ecological health of a region.

deep learning based semantic segmentation can yield a precise measurement of vegetation cover from high-resolution aerial photographs. one challenge is differentiating classes with similar visual characteristics, such as trying to classify a green pixel as grass, shrubbery, or tree. to increase classification accuracy, some data sets contain multispectral images that provide additional information about each pixel. for example, the hamlin beach state park data set supplements the color images with three near-infrared channels that provide a clearer separation of the classes.

this example first shows you how to perform semantic segmentation using a pretrained u-net and then use the segmentation results to calculate the extent of vegetation cover. then, you can optionally train a u-net network on the hamlin beach state parck data set using a patch-based training methodology.

download dataset

this example uses a high-resolution multispectral data set to train the network []. the image set was captured using a drone over the hamlin beach state park, ny. the data contains labeled training, validation, and test sets, with 18 object class labels. the size of the data file is 3.0 gb.

download the mat-file version of the data set using the downloadhamlinbeachmsidata helper function. this function is attached to the example as a supporting file. specify datadir as the desired location of the data.

datadir = fullfile(tempdir,"rit18_data"); 
downloadhamlinbeachmsidata(datadir);

load the dataset.

load(fullfile(datadir,"rit18_data.mat"));
whos train_data val_data test_data
  name            size                         bytes  class     attributes
  test_data       7x12446x7654            1333663576  uint16              
  train_data      7x9393x5642              741934284  uint16              
  val_data        7x8833x6918              855493716  uint16

the multispectral image data is arranged as numchannels-by-width-by-height arrays. however, in matlab®, multichannel images are arranged as width-by-height-by-numchannels arrays. to reshape the data so that the channels are in the third dimension, use the switchchannelstothirdplane helper function. this function is attached to the example as a supporting file.

train_data = switchchannelstothirdplane(train_data);
val_data   = switchchannelstothirdplane(val_data);
test_data  = switchchannelstothirdplane(test_data);

confirm that the data has the correct structure.

whos train_data val_data test_data
  name                size                     bytes  class     attributes
  test_data       12446x7654x7            1333663576  uint16              
  train_data       9393x5642x7             741934284  uint16              
  val_data         8833x6918x7             855493716  uint16

save the training data as a mat file and the training labels as a png file. this facilitates loading the training data using an imagedatastore and a pixellabeldatastore during training.

save("train_data.mat","train_data");
imwrite(train_labels,"train_labels.png");

visualize multispectral data

in this dataset, the rgb color channels are the 3rd, 2nd, and 1st image channels. display the color component of the training, validation, and test images as a montage. to make the images appear brighter on the screen, equalize their histograms by using the function.

figure
montage(...
    {histeq(train_data(:,:,[3 2 1])), ...
    histeq(val_data(:,:,[3 2 1])), ...
    histeq(test_data(:,:,[3 2 1]))}, ...
    bordersize=10,backgroundcolor="white")
title("rgb component of training, validation, and test image (left to right)")

display the last three histogram-equalized channels of the training data as a montage. these channels correspond to the near-infrared bands and highlight different components of the image based on their heat signatures. for example, the trees near the center of the second channel image show more detail than the trees in the other two channels.

figure
montage(...
    {histeq(train_data(:,:,4)),histeq(train_data(:,:,5)),histeq(train_data(:,:,6))}, ...
    bordersize=10,backgroundcolor="white")
title("training image ir channels 1, 2, and 3 (left to right)")

channel 7 is a mask that indicates the valid segmentation region. display the mask for the training, validation, and test images.

figure
montage(...
    {train_data(:,:,7),val_data(:,:,7),test_data(:,:,7)}, ...
    bordersize=10,backgroundcolor="white")
title("mask of training, validation, and test image (left to right)")

visualize ground truth labels

the labeled images contain the ground truth data for the segmentation, with each pixel assigned to one of the 18 classes. get a list of the classes with their corresponding ids.

disp(classes)
0. other class/image border      
1. road markings                 
2. tree                          
3. building                      
4. vehicle (car, truck, or bus)  
5. person                        
6. lifeguard chair               
7. picnic table                  
8. black wood panel              
9. white wood panel              
10. orange landing pad           
11. water buoy                   
12. rocks                        
13. other vegetation             
14. grass                        
15. sand                         
16. water (lake)                 
17. water (pond)                 
18. asphalt (parking lot/walkway)

create a vector of class names.

classnames = [ "roadmarkings","tree","building","vehicle","person", ...
               "lifeguardchair","picnictable","blackwoodpanel",...
               "whitewoodpanel","orangelandingpad","buoy","rocks",...
               "lowlevelvegetation","grass_lawn","sand_beach",...
               "water_lake","water_pond","asphalt"];

overlay the labels on the histogram-equalized rgb training image. add a color bar to the image.

cmap = jet(numel(classnames));
b = labeloverlay(histeq(train_data(:,:,4:6)),train_labels,transparency=0.8,colormap=cmap);
figure
imshow(b)
title("training labels")
n = numel(classnames);
ticks = 1/(n*2):1/n:1;
colorbar(ticklabels=cellstr(classnames),ticks=ticks,ticklength=0,ticklabelinterpreter="none");
colormap(cmap)

perform semantic segmentation

download a pretrained u-net network.

trainedunet_url = "https://www.mathworks.com/supportfiles/vision/data/multispectralunet.mat";
downloadtrainednetwork(trainedunet_url,datadir);
load(fullfile(datadir,"multispectralunet.mat"));

to perform the semantic segmentation on the trained network, use the segmentmultispectralimage helper function with the validation data. this function is attached to the example as a supporting file. the segmentmultispectralimage function performs segmentation on image patches using the function. processing patches is required because the size of the image prevents processing the entire image at once.

predictpatchsize = [1024 1024];
segmentedimage = segmentmultispectralimage(val_data,net,predictpatchsize);

to extract only the valid portion of the segmentation, multiply the segmented image by the mask channel of the validation data.

segmentedimage = uint8(val_data(:,:,7)~=0) .* segmentedimage;
figure
imshow(segmentedimage,[])
title("segmented image")

the output of semantic segmentation is noisy. perform post image processing to remove noise and stray pixels. use the function to remove salt-and-pepper noise from the segmentation. visualize the segmented image with the noise removed.

segmentedimage = medfilt2(segmentedimage,[7,7]);
imshow(segmentedimage,[]);
title("segmented image with noise removed")

overlay the segmented image on the histogram-equalized rgb validation image.

b = labeloverlay(histeq(val_data(:,:,[3 2 1])),segmentedimage,transparency=0.8,colormap=cmap);
figure
imshow(b)
title("labeled segmented image")
colorbar(ticklabels=cellstr(classnames),ticks=ticks,ticklength=0,ticklabelinterpreter="none");
colormap(cmap)

calculate extent of vegetation cover

the semantic segmentation results can be used to answer pertinent ecological questions. for example, what percentage of land area is covered by vegetation? to answer this question, find the number of pixels labeled vegetation. the label ids 2 ("trees"), 13 ("lowlevelvegetation"), and 14 ("grass_lawn") are the vegetation classes. also find the total number of valid pixels by summing the pixels in the roi of the mask image.

vegetationclassids = uint8([2,13,14]);
vegetationpixels = ismember(segmentedimage(:),vegetationclassids);
validpixels = (segmentedimage~=0);
numvegetationpixels = sum(vegetationpixels(:));
numvalidpixels = sum(validpixels(:));

calculate the percentage of vegetation cover by dividing the number of vegetation pixels by the number of valid pixels.

percentvegetationcover = (numvegetationpixels/numvalidpixels)*100;
fprintf("the percentage of vegetation cover is %3.2f%%.",percentvegetationcover);
the percentage of vegetation cover is 51.72%.

the rest of the example shows you how to train u-net on the hamlin beach dataset.

create random patch extraction datastore for training

use a random patch extraction datastore to feed the training data to the network. this datastore extracts multiple corresponding random patches from an image datastore and pixel label datastore that contain ground truth images and pixel label data. patching is a common technique to prevent running out of memory for large images and to effectively increase the amount of available training data.

begin by loading the training images from "train_data.mat" in an . because the mat file format is a nonstandard image format, you must use a mat file reader to enable reading the image data. you can use the helper mat file reader, matread6channels, that extracts the first six channels from the training data and omits the last channel containing the mask. this function is attached to the example as a supporting file.

imds = imagedatastore("train_data.mat",fileextensions=".mat",readfcn=@matread6channels);

create a to store the label patches containing the 18 labeled regions.

pixellabelids = 1:18;
pxds = pixellabeldatastore("train_labels.png",classnames,pixellabelids);

create a from the image datastore and the pixel label datastore. each mini-batch contains 16 patches of size 256-by-256 pixels. one thousand mini-batches are extracted at each iteration of the epoch.

dstrain = randompatchextractiondatastore(imds,pxds,[256,256],patchesperimage=16000);

the random patch extraction datastore dstrain provides mini-batches of data to the network at each iteration of the epoch. preview the datastore to explore the data.

inputbatch = preview(dstrain);
disp(inputbatch)
        inputimage        responsepixellabelimage
    __________________    _______________________
    {256×256×6 uint16}     {256×256 categorical} 
    {256×256×6 uint16}     {256×256 categorical} 
    {256×256×6 uint16}     {256×256 categorical} 
    {256×256×6 uint16}     {256×256 categorical} 
    {256×256×6 uint16}     {256×256 categorical} 
    {256×256×6 uint16}     {256×256 categorical} 
    {256×256×6 uint16}     {256×256 categorical} 
    {256×256×6 uint16}     {256×256 categorical}

create u-net network layers

this example uses a variation of the u-net network. in u-net, the initial series of convolutional layers are interspersed with max pooling layers, successively decreasing the resolution of the input image. these layers are followed by a series of convolutional layers interspersed with upsampling operators, successively increasing the resolution of the input image []. the name u-net comes from the fact that the network can be drawn with a symmetric shape like the letter u.

this example modifies the u-net to use zero-padding in the convolutions, so that the input and the output to the convolutions have the same size. use the helper function, createunet, to create a u-net with a few preselected hyperparameters. this function is attached to the example as a supporting file.

inputtilesize = [256,256,6];
lgraph = createunet(inputtilesize);
disp(lgraph.layers)
  58×1 layer array with layers:
     1   'imageinputlayer'                        image input                  256×256×6 images with 'zerocenter' normalization
     2   'encoder-section-1-conv-1'               2-d convolution              64 3×3×6 convolutions with stride [1  1] and padding [1  1  1  1]
     3   'encoder-section-1-relu-1'               relu                         relu
     4   'encoder-section-1-conv-2'               2-d convolution              64 3×3×64 convolutions with stride [1  1] and padding [1  1  1  1]
     5   'encoder-section-1-relu-2'               relu                         relu
     6   'encoder-section-1-maxpool'              2-d max pooling              2×2 max pooling with stride [2  2] and padding [0  0  0  0]
     7   'encoder-section-2-conv-1'               2-d convolution              128 3×3×64 convolutions with stride [1  1] and padding [1  1  1  1]
     8   'encoder-section-2-relu-1'               relu                         relu
     9   'encoder-section-2-conv-2'               2-d convolution              128 3×3×128 convolutions with stride [1  1] and padding [1  1  1  1]
    10   'encoder-section-2-relu-2'               relu                         relu
    11   'encoder-section-2-maxpool'              2-d max pooling              2×2 max pooling with stride [2  2] and padding [0  0  0  0]
    12   'encoder-section-3-conv-1'               2-d convolution              256 3×3×128 convolutions with stride [1  1] and padding [1  1  1  1]
    13   'encoder-section-3-relu-1'               relu                         relu
    14   'encoder-section-3-conv-2'               2-d convolution              256 3×3×256 convolutions with stride [1  1] and padding [1  1  1  1]
    15   'encoder-section-3-relu-2'               relu                         relu
    16   'encoder-section-3-maxpool'              2-d max pooling              2×2 max pooling with stride [2  2] and padding [0  0  0  0]
    17   'encoder-section-4-conv-1'               2-d convolution              512 3×3×256 convolutions with stride [1  1] and padding [1  1  1  1]
    18   'encoder-section-4-relu-1'               relu                         relu
    19   'encoder-section-4-conv-2'               2-d convolution              512 3×3×512 convolutions with stride [1  1] and padding [1  1  1  1]
    20   'encoder-section-4-relu-2'               relu                         relu
    21   'encoder-section-4-dropout'              dropout                      50% dropout
    22   'encoder-section-4-maxpool'              2-d max pooling              2×2 max pooling with stride [2  2] and padding [0  0  0  0]
    23   'mid-conv-1'                             2-d convolution              1024 3×3×512 convolutions with stride [1  1] and padding [1  1  1  1]
    24   'mid-relu-1'                             relu                         relu
    25   'mid-conv-2'                             2-d convolution              1024 3×3×1024 convolutions with stride [1  1] and padding [1  1  1  1]
    26   'mid-relu-2'                             relu                         relu
    27   'mid-dropout'                            dropout                      50% dropout
    28   'decoder-section-1-upconv'               2-d transposed convolution   512 2×2×1024 transposed convolutions with stride [2  2] and cropping [0  0  0  0]
    29   'decoder-section-1-uprelu'               relu                         relu
    30   'decoder-section-1-depthconcatenation'   depth concatenation          depth concatenation of 2 inputs
    31   'decoder-section-1-conv-1'               2-d convolution              512 3×3×1024 convolutions with stride [1  1] and padding [1  1  1  1]
    32   'decoder-section-1-relu-1'               relu                         relu
    33   'decoder-section-1-conv-2'               2-d convolution              512 3×3×512 convolutions with stride [1  1] and padding [1  1  1  1]
    34   'decoder-section-1-relu-2'               relu                         relu
    35   'decoder-section-2-upconv'               2-d transposed convolution   256 2×2×512 transposed convolutions with stride [2  2] and cropping [0  0  0  0]
    36   'decoder-section-2-uprelu'               relu                         relu
    37   'decoder-section-2-depthconcatenation'   depth concatenation          depth concatenation of 2 inputs
    38   'decoder-section-2-conv-1'               2-d convolution              256 3×3×512 convolutions with stride [1  1] and padding [1  1  1  1]
    39   'decoder-section-2-relu-1'               relu                         relu
    40   'decoder-section-2-conv-2'               2-d convolution              256 3×3×256 convolutions with stride [1  1] and padding [1  1  1  1]
    41   'decoder-section-2-relu-2'               relu                         relu
    42   'decoder-section-3-upconv'               2-d transposed convolution   128 2×2×256 transposed convolutions with stride [2  2] and cropping [0  0  0  0]
    43   'decoder-section-3-uprelu'               relu                         relu
    44   'decoder-section-3-depthconcatenation'   depth concatenation          depth concatenation of 2 inputs
    45   'decoder-section-3-conv-1'               2-d convolution              128 3×3×256 convolutions with stride [1  1] and padding [1  1  1  1]
    46   'decoder-section-3-relu-1'               relu                         relu
    47   'decoder-section-3-conv-2'               2-d convolution              128 3×3×128 convolutions with stride [1  1] and padding [1  1  1  1]
    48   'decoder-section-3-relu-2'               relu                         relu
    49   'decoder-section-4-upconv'               2-d transposed convolution   64 2×2×128 transposed convolutions with stride [2  2] and cropping [0  0  0  0]
    50   'decoder-section-4-uprelu'               relu                         relu
    51   'decoder-section-4-depthconcatenation'   depth concatenation          depth concatenation of 2 inputs
    52   'decoder-section-4-conv-1'               2-d convolution              64 3×3×128 convolutions with stride [1  1] and padding [1  1  1  1]
    53   'decoder-section-4-relu-1'               relu                         relu
    54   'decoder-section-4-conv-2'               2-d convolution              64 3×3×64 convolutions with stride [1  1] and padding [1  1  1  1]
    55   'decoder-section-4-relu-2'               relu                         relu
    56   'final-convolutionlayer'                 2-d convolution              18 1×1×64 convolutions with stride [1  1] and padding [0  0  0  0]
    57   'softmax-layer'                          softmax                      softmax
    58   'segmentation-layer'                     pixel classification layer   cross-entropy loss

select training options

train the network using stochastic gradient descent with momentum (sgdm) optimization. specify the hyperparameter settings for sgdm by using the (deep learning toolbox) function.

training a deep network is time-consuming. accelerate the training by specifying a high learning rate. however, this can cause the gradients of the network to explode or grow uncontrollably, preventing the network from training successfully. to keep the gradients in a meaningful range, enable gradient clipping by specifying "gradientthreshold" as 0.05, and specify "gradientthresholdmethod" to use the l2-norm of the gradients.

initiallearningrate = 0.05;
maxepochs = 150;
minibatchsize = 16;
l2reg = 0.0001;
options = trainingoptions("sgdm",...
    initiallearnrate=initiallearningrate, ...
    momentum=0.9,...
    l2regularization=l2reg,...
    maxepochs=maxepochs,...
    minibatchsize=minibatchsize,...
    learnrateschedule="piecewise",...
    shuffle="every-epoch",...
    gradientthresholdmethod="l2norm",...
    gradientthreshold=0.05, ...
    plots="training-progress", ...
    verbosefrequency=20);

train the network or download pretrained network

to train the network, set the dotraining variable in the following code to true. train the model by using the (deep learning toolbox) function.

train on a gpu if one is available. using a gpu requires parallel computing toolbox™ and a cuda® enabled nvidia® gpu. for more information, see gpu computing requirements (parallel computing toolbox). training takes about 20 hours on an nvidia titan x.

dotraining = false; 
if dotraining
    net = trainnetwork(dstrain,lgraph,options);
    modeldatetime = string(datetime("now",format="yyyy-mm-dd-hh-mm-ss"));
    save(fullfile(datadir,"multispectralunet-" modeldatetime ".mat"),"net");
end

evaluate segmentation accuracy

segment the validation data. 

segmentedimage = segmentmultispectralimage(val_data,net,predictpatchsize);

save the segmented image and ground truth labels as png files. the example uses these files to calculate accuracy metrics.

imwrite(segmentedimage,"results.png");
imwrite(val_labels,"gtruth.png");

load the segmentation results and ground truth using .

pxdsresults = pixellabeldatastore("results.png",classnames,pixellabelids);
pxdstruth = pixellabeldatastore("gtruth.png",classnames,pixellabelids);

measure the global accuracy of the semantic segmentation by using the function.

ssm = evaluatesemanticsegmentation(pxdsresults,pxdstruth,metrics="global-accuracy");
evaluating semantic segmentation results
----------------------------------------
* selected metrics: global accuracy.
* processed 1 images.
* finalizing... done.
* data set metrics:
    globalaccuracy
    ______________
       0.90411

the global accuracy score indicates that just over 90% of the pixels are classified correctly.

references

[1] kemker, r., c. salvaggio, and c. kanan. "high-resolution multispectral dataset for semantic segmentation." corr, abs/1703.01918. 2017.

[2] ronneberger, o., p. fischer, and t. brox. "u-net: convolutional networks for biomedical image segmentation." corr, abs/1505.04597. 2015.

[3] kemker, ronald, carl salvaggio, and christopher kanan. "algorithms for semantic segmentation of multispectral remote sensing imagery using deep learning." isprs journal of photogrammetry and remote sensing, deep learning rs data, 145 (november 1, 2018): 60-77. https://doi.org/10.1016/j.isprsjprs.2018.04.014.

see also

(deep learning toolbox) | (deep learning toolbox) | | | | | | |

related topics

external websites

    网站地图