perClass Documentation
version 5.1 (31-May-2017)

Feature extraction, table of contents

Chapter 8.2: Feature extraction in local image regions

8.2.1. Introduction ↩

Region feature extractors process square image neighborhoods and represent its central pixel by the resulting feature vector. This is useful to account for local spatial information and structure in images. Typical applications are texture classification and image segmentation (distinguishing edges).

local image feature extraction in Matlab

Region features help you to characterize:

Region feature extraction framework allows you to:

8.2.2. Quick example ↩

For this example we load a dice image sddata object I:

>> I=sdimage('dice01.jpg','sddata')
101376 by 3 sddata, class: 'unknown'

It becomes a data set with three features corresponding to R,G,B bands respectively.

>> sdimage(I)

ans =

 1

Dice image (showing the first of the three bands i.e. the 'R' channel)

The blue color is the label layer that is uniform, because we have only one "unknown" class by default.

We want to characterize local image structure. The simplest approach is to compute mean and standard deviation in small image neighborhoods with the 'moments' extractor:

 >> A=sdextract( I(:,1), 'region', 'moments')
 96945 by 2 sddata, class: 'unknown'

Note that 'moments' extractor operates on a single image band and we, therefore, provide it only with the first feature ('R' channel). The extraction domain is 'region' and the extractor name is 'moments'. By default, 'moments' features are extracted in 8x8 sliding windows. The resulting data set A is still an image (see this chapter)

We can visualize it with sdimage. We open two separate figures, one showing the first feature (local mean), the other showing the second feature (local standard deviation):

>> sdimage(A)

ans =

 2

>> sdimage(A(:,2))

ans =

 3

Local image features - mean and standard deviation in image neighborhoods

Note, that the extracted image contains a thin border without data. This is because perClass computes region features only if the neighborhood fully covers image data. For example, the left-upper corner pixel cannot be represented because the neighborhood "sticks out" from available data. Therefore, the number of samples in A is slightly smaller that in the original image I.

Why is local feature extraction useful? Let us try to separate objects and background by clustering. Move to the Figure 2 that contains data set A with both bands and select 'Cluster with k-means' from 'Image' menu. Enter the desired number of clusters, for us 2 and click OK.

Clustering local image features is more robust than direct RGB pixel clustering

We can observe is that clustering on mean/std features robustly identifies the dice objects including the dots due to larger spatial context. For comparison, perform clustering on the original RGB image I or its first band I(:,1).

8.2.3. Region feature types ↩

8.2.3.1. Raw neighborhood values ↩

Example:

>> im1
262144 by 1 sddata, class: 'unknown'
>> out=sdextract(im1,'region','raw','block',3)
260100 by 9 sddata, class: 'unknown'

8.2.3.2. Local mean and standard deviation ↩

Example: See above.

8.2.3.3. Local histograms ↩

The histogram feature extractor requires explicitly specified data range. It is important that the 'range' is identical in the entire classification problem i.e. on the training set and in the production or test set. Therefore, you should not use data-driven definitions such as 'range',[+min(im) +max(im)]! This holds also for other histogram-like features ('histfeat' and 'cm').

Examples:

>> data=sdextract(I(:,1),'region','hist','range',[0 255])
96945 by 8 sddata, class: 'unknown'

>> data=sdextract(I(:,1),'region','hist','range',[0 255],'bins',16)
96945 by 16 sddata, class: 'unknown'

8.2.3.4. Features extracted from local histograms ↩

The 'histfeat' extractor computes five statistical features characterizing shape of local histograms. These are: 'mean','2nd moment','skewness','kurtosis', and 'entropy'

>> data=sdextract(I(:,1),'region','histfeat','range',[0 255])
96945 by 5 sddata, class: 'unknown'
>> data.featlab
sdlab with 5 entries, 5 groups: 
'Feature 1,mean'(1) 'Feature 1,2nd moment'(1) 'Feature 1,skewness'(1) 'Feature 1,kurtosis'(1) 'Feature 1,entropy'(1) 

8.2.3.5. Co-occurrence matrices ↩

Co-occurrence matrix is a two-dimensional histogram estimating probability that a pixel has a specific gray-level while a displaced pixel exhibits another gray-level. Co-occurrence matrix encodes structural information which is useful for derivation of informative data representation in texture classification problems.

In this example, we compute co-occurrences on R channel, subsampling intensities to 4 bins:

>> data=sdextract(I(:,1),'region','cm','range',[0 255],'bins',4)
96945 by 16 sddata, class: 'unknown'

Let is identify pixel row 101, column 162 in our data set. We can use standard Matlab sub2ind function and imsize of our image:

>> ind=sub2ind(getiminfo(data,'imsize'),101,162)

ind =

   46469

Now we get the sample with pixel index ind:

>> data( data.pixel==ind )
1 by 16 sddata, class: 'unknown'

Co-occurrence as feature vector:

>> +data( data.pixel==ind )

ans =

  Columns 1 through 7

     0    0.0089    0.0089    0.0089    0.0089         0         0

  Columns 8 through 14

0.0179    0.0089         0    0.0179    0.0268    0.0089    0.0179

  Columns 15 through 16

0.0268    0.8393

and reshaped into 4x4 matrix:

>> reshape(ans,[4 4])

ans =

     0    0.0089    0.0089    0.0089
0.0089         0         0    0.0179
0.0089         0    0.0179    0.0268
0.0089    0.0179    0.0268    0.8393

8.2.3.6. Gaussian filter and its derivatives ↩

The 'gauss' feature extractor convolves input image with Gaussian filter or its derivative and returns its response in the central pixel of each image region.

By default, sigma of 2 and block of 8 pixels are used:

% using the dice01.jpg image loaded above
>> data=sdextract(I(:,1),'region','gauss')
block: 8, sigma: 2.0 yields kernel coverage 91.5%
96945 by 1 sddata, class: 'unknown'

Note the message displaying the actual coverage of the convolution kernel on the image region. Gaussian kernel has in general an infinite support. In practice, we only use the convolution values in the neighborhood defined by the 'block' option. Therefore our filter only covers the sliding window partially.

If we use larger sigma, the coverage gets lower which means that our actual filtering result is getting further away from ideal Gaussian shape:

>> data=sdextract(I(:,1),'region','gauss','sigma',3)
block: 8, sigma: 3.0 yields kernel coverage 67.2%
96945 by 1 sddata, class: 'unknown'

In order to increase the coverage, we may increase the 'block' size:

>> data=sdextract(I(:,1),'region','gauss','sigma',3,'block',12)
block: 12, sigma: 3.0 yields kernel coverage 91.3%
94457 by 1 sddata, class: 'unknown'

Example on 'detergent1' data set illustrating the noise suppression:

>> load detergent1
>> imdet=sdimage(im1m,'sddata')
16384 by 7 sddata, class: 'unknown'

>> data=sdextract(imdet(:,1),'region','gauss','sigma',3,'block',12)
block: 12, sigma: 3.0 yields kernel coverage 91.3%
13689 by 1 sddata, class: 'unknown'
>> sdimage(imdet)

ans =

 5

>> sdimage(data)

ans =

 6

Gaussian derivatives are also supported with 'der' option:

>> data2=sdextract(imdet(:,1),'region','gauss','sigma',3,'block',12,'der',[1 0])
block: 12, sigma: 3.0 yields kernel coverage 91.3%
13689 by 1 sddata, class: 'unknown'
>> sdimage(data2)

ans =

 7

8.2.3.7. Sobel filter ↩

Sobel filter computes local image gradient in 3x3 image neighborhoods.

>> data=sdextract(imdet,'region','sobel')
15876 by 2 sddata, class: 'unknown'

>> sdimage(imdet)
>> sdimage(im)

The data set contains two features. The first describes gradient magnitude and the second its orientation:

>> data.featlab
sdlab with 2 entries, 2 groups: 'Feature 1,gradient magnitude'(1) 'Feature 1,gradient orientation'(1) 

In order to highlight larger-scale structures, we may want to filter the image first with a Gaussian:

>> data1=sdextract(imdet(:,1),'region','gauss','sigma',3,'block',16)
block: 16, sigma: 3.0 yields kernel coverage 98.5%
12769 by 1 sddata, class: 'unknown'

>> data2=sdextract(data1,'region','sobel')
12321 by 2 sddata, class: 'unknown'

>> sdimage(data2(:,1))
>> sdimage(data2(:,2))

The result of directional Sobel filters may be extracted using the 'sobel.x' and 'sobel.y' options, respectively:

>> x=sdextract(imdet(:,1),'region','sobel.x')
15876 by 1 sddata, class: 'unknown'

8.2.3.8. User-defined filter bank ↩

Filter bank is specified with 3D matrix K (block x block x filter_count). The 'fbank' extractor supports maximum response filtering with 'join' option. To define it, a vector F should be supplied with one entry per filter. Each entry in F specifies to which final output (joined filter) is this filter merged.

8.2.3.9. Leung-Malik multi-orientation/multi-scale filter bank ↩

Example:

>> im1
986049 by 1 sddata, class: 'unknown'
>> out=sdextract(im1,'region','fbank:LM')
893025 by 48 sddata, class: 'unknown'

LM filters:

multi-scale multi-orientation filter bank for feature extraction and texture classification

Reference: http://www.robots.ox.ac.uk/~vgg/research/texclass/filters.html

8.2.3.10. Schmid rotationally-invariant filter bank ↩

Schmid filters:

multi-scale rotationally-invariant filter bank for feature extraction and texture classification

8.2.3.11. Maximum Response filter bank combining multiple orientations ↩

In this example, we extract MR8 filter bank on an image combining multiple Brodatz textures:

 >> im1=sdimage('texmos1.p512.tiff','sddata')
 262144 by 1 sddata, class: 'unknown'

 >> out=sdextract(im1,'region','fbank:MR8')
 215296 by 8 sddata, class: 'unknown'

 >> sdimage(out); sdimage(im1);

The figure shows original image and one output feature i.e. combined output of multiple orientation filters.

maximum response filter bank feature extractor combining multiple orientations

8.2.3.12. Maximum Response filter bank combining multiple orientations and scales ↩

8.2.3.13. Gray-level morphology features ↩

Gray-level morpology enables straightforward edge detection in single-band images. It uses a square structure element with a size defined by the 'block' option. At the low level, erosion and dilation operations are defined similar to the classical black and white morphology. On top of these, opening and closing operations are specified. The high-level 'edge' extractor uses opening and closing to enhance edge regions.

8.2.3.14. User-defined feature extractors ↩

Apart from built-in feature extractors, sdextract can execute a custom Matlab function on each image neighborhood. This allows us to quickly experiment with arbitrary external code for feature computation while benefiting from the flexible image grid definition and masking.

To use custom feature extractor, we need to create a Matlab function following a specific syntax. As an example, lets define a function custom_extract_hist.m that will compute a local image histogram in each region.

The custom feature extractor is invoked by providing function handle instead of extractor name:

>> data=sdextract(im,'region',@custom_extract_hist)
..........
14641 by 8 sddata, class: 'unknown'

In our example, we want to be able to set the number of histogram bins, and data range for the histograms estimated. These are the extractor parameters.

To provide parameters, we simply pass a cell array with name, value pairs after the function handle:

>> data=sdextract(im,'region',@custom_extract_hist,{'bins',16,'min',0,'max',255})
..........
14641 by 16 sddata, class: 'unknown'

How to implement the custom feature extractor function?

Any feature extraction function needs to do two things:

  1. Set extraction parameter and return feature names (called once)
  2. Extract feature vector from a given neighborhood (called many times)

Our function will have, as any custom callback function, two inputs and two outputs:

function [data,par]=custom_extract_hist(data,par)

if isempty(data)

    % 1. parameter setting
    %     - (optional) set any default params or precomputed data used
    %       during block processing
    %     - return feature labels

    % set default number of bins
    if ~isfield(par,'bins')
    par.bins=8;
    end
    % set default minimum and maximum value
    if ~isfield(par,'min')
    par.min=0;
    end
    if ~isfield(par,'max')
    par.max=255;
    end

    % store the histogram bins definition
    par.histbins=par.min:floor((par.max-par.min+1)/par.bins):par.max;

    % return feature labels
    data=sdlab('Hist ',1:par.bins);

else

    % 2. block processing
    %     - process input block matrix data and return feature vector
    % data is a image block (matrix)
    % par is a struct with parameters

    A=data(:);

    data=hist(A,par.edge)./sum(A);

end 

The first part is called by sdextract once, before the block-by-block feature computation, to initialize any parameters our extractor need and return the feature labels. The data input argument is empty, the par argument is a structure with some basic fields describing the extraction process, such as block size and step.

We may add any further fields to par. In our example, we set the bins, min, and max fields to default values, in not already present.

We also create anything we need for feature extraction process. In our example, it is useful to define the histogram binning vector histbins once.

Finally, we return the feature labels.

The second part of the extractor function is called for each image neighborhood. The data argument is now a neighborhood matrix and par our parameter structure with all fields we set. Now we perform the computation and return the feature vector.

Note, that feature extraction process is not limited to single-band images. If the input image data set, passed to sdextract has multiple bands, the custom function will receive a 3D matrix for each neighborhood. In this way, you may implement e.g. color texture features or process hyperspectral images in one pass.

8.2.4. Grid definition ↩

Region features are computed on a regular grid laid over an image. The grid is defined by two parameters:

Block size is an scalar defining the square sliding window. Default value is 8 (8x8 pixel regions). It may be changed using the 'block' option.

The default step is 1. Large step is useful for representing large images as it yields sparse smaller representation.

>> B=sdextract(I(:,1),'region','moments','block',16)
92001 by 2 sddata, class: 'unknown'

>> C=sdextract(I(:,1),'region','moments','block',16,'step',2)
23153 by 2 sddata, class: 'unknown'

The getiminfo method can be used to access detailed information on grid definition for any image data set.

>> getiminfo(I)

ans = 

imsize: [288 352]

>> getiminfo(C)

ans = 

  imsize: [288 352]
    grid: 1
   block: 16
    step: 2
gridsize: [137 169]

Typical use-case for getiminfo is to return directly image size:

>> s=getiminfo(C,'imsize')

s =

   288   352

8.2.5. Visualization of feature images computed on a grid ↩

The sdimage command visualizes any image data set in the context of the ''original image''. Therefore, data extracted on a grid with step higher than one will be sparse:

>> sdimage(C)

ans =

 1

Note the gaps between data points. We may wish to remove the grid pattern for the sake of visualization purposes. This may be done using 'Shrink to grid' command in Image menu:

A new figure window opens with out image constrained to grid.

We can achieve the same from the command prompt with the sdimage 'grid' option:

>> D=sdimage(C,'grid')
23153 by 2 sddata, class: 'unknown'
>> sdimage(D)

ans =

 2

>> getiminfo(D)

ans = 

imsize: [137 169]

8.2.6. Computing features based on foreground mask ↩

Often, we only want to process certain part of an image, e.g. object that was already segmented out from the conveyor belt by low-level image processing.

sdextract accepts the 'mask' option that specifies where to compute the features.

Lets assume we have a object mask as a set of labels in out data set:

Our mask label M:

>> M
sdlab with 16384 entries, 2 groups: 'object'(8207) 'background'(8177) 

We may extract features only from the region defined by this mask. The 'mask' option takes, as argument a logical vector or image matrix.

If our mask is represented by label object, we may just compare it to the desired class resulting in a logical vector:

>> out=sdextract(imdet(:,1),'region','moments','mask',M=='object')
6667 by 2 sddata, class: 'unknown'
>> sdimage(out)

Features are computed only for neighborhoods that fully cover available image data. Therefore, the resulting data set out contains less samples than the 'object' class in M.

If our mask is a matrix such as objmask:

>> objmask=reshape(-M,getiminfo(imdet,'imsize'));
>> size(objmask)

ans =

   128   128

Numerical representation of our labels contain 1-based indices:

>> unique(objmask(:))

ans =

 1
 2

Because 'object' was first in the M.lab.list (see above), we can define our mask with:

>> out=sdextract(imdet(:,1),'region','moments','mask',objmask==1)
6667 by 2 sddata, class: 'unknown'

8.2.7. Propagating image labels ↩

When applied to sddata data set with image data, sdextract propagates image labels and all sample properties to the output data.

We may, for example, define labels my hand, painting directly in the sdimage figure. In this example, we painted background, particle and "interesting texture" regions. We save the data set in Matlab workspace using Image menu:

 >> sdimage(imdet)

>> Creating data set data2 in the workspace.
16384 by 1 sddata, 3 classes: 'background'(7629) 'particle'(7760) 'interesting texture'(995) 

We may now compute the features on data set data2. The class labels get propagated to the pixels that serve as block centers in feature computation:

>> a=sdextract(data2(:,1),'region','histfeat','block',4)
15625 by 5 sddata, 3 classes: 'background'(7517) 'particle'(7113) 'interesting texture'(995) 

>> sdimage(a)

8.2.8. Visualizing regions in original image ↩

When extracting features from local image neighborhoods, we often want to visualize the exact block position in the original image. For example, this is very useful in order to understand, what parts of the object are misclassified.

The label image shows the image neighborhoods in the context of the original image, used for feature extraction.

Say, we start from the particle image with the segmented background:

We save the image data set into Matlab workspace as data2:

>> Creating data set data2 in the workspace.
16384 by 7 sddata, 2 'cluster' groups: 'C1'(8263) 'C2'(8121) 

We only use the C1 cluster (the particle foreground):

>> sub=data2(:,:,'C1')
8263 by 7 sddata, 'cluster' lab: 'C1'

Now, we compute features on the first band of the particle:

>> b=sdextract(sub(:,1),'region',@custom_extract_hist,'block',8,'step',10)
..........
60 by 8 sddata, 'cluster' lab: 'C1'

To visualize block positions, use sdimage with 'labim' option. It will extract RGB image of the same size at the original image and impaint blocks with colors based on the labels b.lab:

>> LI=sdimage(b,'labim');
>> figure; imagesc(LI)

Note the gap inside the particle. As mentioned above, sdextract computes features only in regions without holes.

We may wish to visualize the label image blended with the original image with 'blend' option. We provide image matrix with <0,1> values with the same dimensions as the original image (imsize in getiminfo(b)).

>> LI=sdimage(b,'labim','blend',I./256);
>> figure; imagesc(LI)

To adjust image blending, use 'alpha' parameter:

>> LI=sdimage(b,'labim','blend',I./256,'alpha',0.8);
>> figure; imagesc(LI)

Finally, to use different set of labels than the b.lab, you may provide the any sdlab object with the same length directly after the 'labim' option. This is useful when visualizing decisions (we assume a classifier p trained on extracted features):

>> dec=b*p
>> LI=sdimage(b,'labim',dec,'blend',I./256,'alpha',0.8);
>> figure; imagesc(LI)