perClass Documentation
version 5.1 (31-May-2017)

Feature extraction, table of contents

Chapter 8.3: Representing objects identified in an image

8.3.1. Introduction ↩

Object feature extraction derives one feature vector for each object identified in the original data set.

Typical use-case is image-based object recognition task such as defect detection. Original image is first processed by local (region) feature extraction. A local classifier is then trained to distinguish e.g. object/background or detecting a defect. When applied to a new image, each pixel is assigned to one of classes of interest (e.g. object/background).

In the second step, the decisions are spatially segmented using connected component analysis. Each object (connected spatial area) is then represented by a set of features and an object-level classifier is build. This helps us to distinguish objects based on content (original local features) or shape.

This chapter concerns derivation of object-level representation using sdextract command.

object feature extraction in Matlab

8.3.2. Quick example of the entire process ↩

We will use dice data set.

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

We will: Preparing pixel classifier ↩

First, we will train dice/background local classifier. We will paint dice and background labels:

Painting image labels to train pixel classifier.

We save the labeled image back to Matlab workspace into variable I by pressing 's' key or via Image menu / Create data set in workspace.

>> Creating data set I in the workspace.
101376 by 3 sddata, 2 classes: 'dice'(4020) 'background'(97356) 

We will train a Gaussian detector on the 'dice' class, and make sure the non-targets will be still called background (without 'non-target' option, the default would be 'non-dice'):

>> pd=sddetect(I,'dice',sdgauss,'non-target','background') 
  1: back   -> background
  2: dice   -> dice
sequential pipeline       3x1 'Gaussian model+Decision'
 1 Gaussian model          3x1  full cov.mat.
 2 Decision                1x1  ROC thresholding on dice (2000 points, current 1)

It is useful to visualize detector behaviour on an image:

>> sdimage(I,pd,'roc')

We will set the detector operating point to allow some errors in the background, rather that having holes in the dice. By pressing 's' we can save the setting back into the pd detector pipeline.

Tuning detector trade-off not to loose objects in an image. Applying pixel classifier ↩

Our dice detector may be now applied to a new image (open sdimage figure use "Apply classifier" form "Image" menu or press 'd' and enter pd as the name of a pipeline to apply):

>> I2=sdimage('dice02.jpg','sddata')
101376 by 3 sddata, class: 'unknown'
>> sdimage(I2)

Tuning detector trade-off not to loose objects in an image. Segmenting objects ↩

We will segment the objects based on decisions with "Image" / "Connected components" / "Find connected component" command.

Segmenting connected components based on classifier decisions.

Segmented objects are defined by a new set of 'object' labels. We can save the data set in Matlab workspace as A:

Connected regions identified from classifier decisions

>> Creating data set A in the workspace.
101376 by 3 sddata, 17 'object' groups: [135    996    932    895    126    939    947    126    377    915    156    212    158    290   1512  92583     77]
>> A.lab'
 ind name                        size percentage
   1 dice-object 30               135 ( 0.1%)
   2 dice-object 43               996 ( 1.0%)
   3 dice-object 57               932 ( 0.9%)
   4 dice-object 75               895 ( 0.9%)
   5 dice-object 82               126 ( 0.1%)
   6 dice-object 87               939 ( 0.9%)
   7 dice-object 98               947 ( 0.9%)
   8 dice-object102               126 ( 0.1%)
   9 dice-object103               377 ( 0.4%)
  10 dice-object105               915 ( 0.9%)
  11 dice-object121               156 ( 0.2%)
  12 dice-object122               212 ( 0.2%)
  13 dice-object138               158 ( 0.2%)
  14 dice-object145               290 ( 0.3%)
  15 dice-small objects          1512 ( 1.5%)
  16 background-object 1        92583 (91.3%)
  17 background-small objects      77 ( 0.1%)

The object labels group pixels corresponding to each spatially connected object.

A command-line alternative is:

 >> A=sdsegment( I2.*pd ,'minsize',100)
 101376 by 3 sddata, 17 'object' groups: [135    996    932    895    126    939    947    126    377    915    156    212    158    290   1512  92583     77] Extracting object features ↩

Object features may be extracted with sdextract using:

 >> obj=sdextract(A,'object','mean')
 Extracting a feature vector from each of 17 objects ('object' labels):
 17 by 3 sddata, class: 'unknown'

For general syntax of sdextract, look here.

The mean object extractor computes one mean vector for each object, defined by 'object' labels. In our RGB image example, this corresponds to mean color information.

We open the extracted data set obj in scatter plot. We can observe that a set of dice objects is separated from green background (change to 'object' labels in 'Scatter' menu).

 >> sdscatter(obj)

 ans =


Feature space with data extracted from individual objects.

You can visualize a specific object in sdimage:

 >> sub=A(:,:,'/57')
 932 by 3 sddata, 'object' lab: 'dice-object 57'
 >> sdimage(sub)

The object-level classifier would be then trained on properly labeled data set obj.

8.3.3. Object features ↩ Object size ↩ Mean of object pixels ↩ Sum of object pixels ↩ Histogram of a specific input feature per object ↩ Shape features on object mask ↩ Shape features on object content ↩ Example of computing per-object histogram of local gradient ↩

In this example, we separate dice objects from background regions based on their strong gradient structure. We characterize it by a per-object histogram of gradient.

We apply the dice detector trained above and segment objects preserving all even single-pixel islands:

>> B=sdsegment(I2.*pd,'minsize',1)
101376 by 3 sddata, 182 'object' groups

We smooth the image with Gaussian filter and then extract gradient with sobel operator:

>> G=sdextract(B(:,1),'region','gauss','sigma',2)
block: 8, sigma: 2.0 yields kernel coverage 91.5%
96945 by 1 sddata, 167 'object' groups
>> S=sdextract(G,'region','sobel')
95697 by 2 sddata, 167 'object' groups

For more details on local filter features, see this chapter.

When we visualize local gradient (first feature in S), we can observe high variability within dice objects and low in the background:

>> sdimage(S)

Visualizing gradient magnitude.

Therefore, we extract per-object histograms, bining the values into three output features (very low, middle and high gradient).

>> obj=sdextract(S(:,1),'object','hist','range',[0 255],'bins',3)
Extracting a feature vector from each of 167 objects ('object' labels):
167 by 3 sddata, class: 'unknown'

We obtain one output feature vector for each of the 167 connected components in the data set S.

>> sdscatter(obj)

Extracting gradient per object to classify objects with strong internal edge structure.

The dice objects are having small probability of low gradients (1st feature) and high probability of middle gradients (2nd feature). The segmentation islands in the background fall into high probable area on the first feature.

8.3.4. Copying labels into object data set ↩

When extracting features per object with sdextract, it is often useful to copy object-related meta-data into the output set.

In our dice example, we may wish to preserve the information from which image is a specific object from. We already may have this information available in the original image set, e.g. as an image label:

>> I=sdimage('dice01.jpg','sddata')
101376 by 3 sddata, class: 'unknown'
>> I.image
sdlab with 101376 entries from 'dice01.jpg'

Also any data set created by region feature extraction will contain this information:

>> B=sdsegment(I.*pd,'minsize',100)
101376 by 3 sddata, 18 'object' groups: [115    134    102    948    998    114    926    158    943    934    246    997    358    126    109   1532  92579     57]

>> sub=B(:,:,'/74')
926 by 3 sddata, 'object' lab: 'dice-object 74'
>> sub.image
sdlab with 926 entries from 'dice01.jpg'

Once we extract object features, a new per-object data set is created and only 'object' labels are copied by default:

>> obj=sdextract(B,'object','size')
Extracting a feature vector from each of 18 objects ('object' labels):
18 by 1 sddata, class: 'unknown'

>> obj'
18 by 1 sddata, class: 'unknown'
sample props: 'lab'->'class' 'class'(L) 'object'(L) 'bbox'(N)
feature props: 'featlab'->'featname' 'featname'(L)
data props:  'data'(N)

We may copy other labels sets using the 'copy' option:

>> obj2=sdextract(B,'object','size','copy','image')
Extracting a feature vector from each of 18 objects ('object' labels):
18 by 1 sddata, class: 'unknown'
>> obj2'
18 by 1 sddata, class: 'unknown'
sample props: 'lab'->'class' 'class'(L) 'object'(L) 'image'(L) 'bbox'(N)
feature props: 'featlab'->'featname' 'featname'(L)
data props:  'data'(N)

>> obj2(1).image
sdlab with one entry: 'dice01.jpg'

Only sdlab label sets may be copied, not general meta-data.

Multiple sets of labels may be specified in a cell array:

>> obj3=sdextract(B,'object','size','copy',{'image','class'})
Extracting a feature vector from each of 18 objects ('object' labels):
18 by 1 sddata, 2 classes: 'background'(2) 'dice'(16) 

Note, that label set may be copied to object data set only if a value is always unique in each object (otherwise, the single per-object value would be undefined). In our example, we're lucky as the dice/background 'class' labels are always unique in any connected object found.

8.3.5. Bounding box of objects ↩

sdextract adds to each extracted per-object feature vector a new 'bbox' meta-data containing information on a object bounding box in the original image.

>> obj=sdextract(B,'object','mean','copy','image')
Extracting a feature vector from each of 18 objects ('object' labels):
18 by 3 sddata, class: 'unknown'

>> obj'
18 by 3 sddata, class: 'unknown'
sample props: 'lab'->'class' 'class'(L) 'object'(L) 'image'(L) 'bbox'(N)
feature props: 'featlab'->'featname' 'featname'(L)
data props:  'data'(N)

>> obj(9).bbox

ans =

     144         220          37          36

The information is stored in the [row column height width] format where row and column correspond to the upper left corner, respectively.

We may use this information to highlight a specific object in the original image, based on extracted data obj only:

The image name is available as we copied image labels with 'copy' option above:

>> obj(9).image
sdlab with one entry: 'dice01.jpg'
>> im=imread(+obj(9).image);

>> figure; imagesc(im)

>> h=rectangle('position',[obj(9).bbox(2) obj(9).bbox(1) obj(9).bbox(4) obj(9).bbox(3)])
>> set(h,'edgecolor',[1 1 1])

Visualizing bounding box of an object in an image