Pr

1
A Generic Approach for Image Classification Based
on Decision Tree Ensembles and Local Sub-windows
Raphaël Marée, Pierre Geurts, Justus Piater,
Louis Wehenkel University of Liège,
Belgium maree_at_montefiore.ulg.ac.be

• Problem
• Many application domains require classification
  of characters, symbols, faces, 3D objects,
  textures,
• Specific feature extraction methods must be
  manually adapted when considering a new
  application
• Approach
• Recent and generic ML algorithm based on
  decision tree ensembles and working directly on
  pixel values
• Extension with local sub-window extraction
• Results
• Competitive with the state of the art on four
  well known datasets MNIST, ORL, COIL-100, OUTEX
• Encouraging results for robustness
  (generalisation, rotation, scaling, occlusion)

Abstract
2
Image classification

• Many different kind of problems

• Usually tackled using
• Problem-specific feature extraction
• ie. extracting a reduced set of  interesting 
  features from the initially huge number of
  pixels
• Learning or matching algorithm

• Our generic approach
• Working directly on pixel values ie. without
  any feature extraction ie. images are described
  by integer values (grey or RGB intensities)
  of all pixels
• Ensemble of decision trees

3
Global generic approach

• Ensemble of extremely randomized trees
  (extra-trees)
• Learning
• Top-down induction algorithm like classical
  decision tree (with tests at the internal nodes
  of the form ak,l lt ath that compare the value
  of the pixel at position (k,l) to a threshold
  ath) but
• Test attributes and thresholds in internal nodes
  are chosen randomly,
• Each tree is fully developed until it perfectly
  classifies images in the learning sample,
• Several extra-trees are built from the same
  learning sample.
• Testing
• Propagate the entire test image successively into
  all the trees (involves comparing pixel values to
  thresholds in test nodes) and assign to the image
  the majority class among the classes given by the
  trees.

4
Local generic approach

• Extra-trees and Sub-windows
• Learning
• Given a window size w1 x w2 and a large number
  Nw
• Extract Nw sub-windows at random from learning
  set images and assign to each sub-window the
  classification of its parent image
• Build a model to classify these Nw sub-windows
  by using the w1 x w2 pixel values that
  characterize them
• Testing
• Given the window size w1 x w2
• Extract all possible sub-windows of size w1 x w2
  from test image
• Apply the model on each sub-window
• Assign to the image the majority class among the
  classes assigned to the sub-windows by the model

5
Experiments description

• Database specificationEvery image in each
  database is described by all its pixel values and
  belong to one class.

DBs images features classes
MNIST 70000 784 (28x28x1) 10
ORL 400 10304 (92x112x1) 40
COIL-100 7200 3072 (32x32x3) 100
OUTEX 864 49152 (128x128x3) 54

• Database protocolsSeparation of each database
  in two independent sets the learning set (LS) of
  pre-classified images used to build a model and
  the test set (TS) used to evaluate the model.

• MNIST
• LS first 60000 images
• TS last 10000 remaining images
• ORL
• 100 random runs
• LS 200 images
• TS 200 remaining images
• COIL-100
• LS 1800 images (k20, k0..17)
• TS 5400 remaining images
• OUTEX
• LS 432 images
• TS 432 remaining images

6
Experiments results

• Error rates on test sets

DBs Extra-trees Extra-trees Sub-windows State-of-the-art
MNIST 3.26 2.63(w1w224) 12 0.7 1
ORL 4.56 1.43 2.13 1.18(w1w232) 7.5 0 2
COIL-100 2.04 0.39(w1w216) 12.5 0.1 3
OUTEX 64.35 2.78(w1w24) 9.5 0.2 4
1 Y. LeCun and L. Bottou and Y. Bengio and P.
Haffner, Gradient-based learning applied to
document recognition, 1998 2 R. Paredes and A.
Perez-Cortes, Local representations and a direct
voting scheme for face recognition, 2001 3 S.
Obrzalek and J. Matas, Object Recognition using
Local Affine Frames on Distinguished Regions,
2002 4 T. Mäenpää, M. Pietikäinen, and J.
Viertola, Separating color and pattern
information for color texture discrimination, 2002

• Computing times
• Learning on OUTEX
• Extra-trees 5 sec
• Extra-trees Sub-Windows 8min
• Testing on OUTEX (one image)
• Extra-trees lt 1 msec
• Extra-trees Sub-Windows 0,6 sec

7
Evaluation of Robustness

• Generalisation
• Rotation
• Scaling
• Occlusion

Considering different learning sample sizes
(COIL-100)
Image-plane rotation of the test images (COIL-100)
Scaled version of the test images, with model
built from 32x32 images (COIL-100)
Erasing right parts of the test images (COIL-100)
8
Conclusion

• Novel, generic, and simple method
• Competitive accuracy
• Our local generic method (Extra-trees
  Sub-windows) is close to state-of-the-art methods
  without any problem-specific feature extraction
  but still slightly inferior to best results
• In practice, is it necessary to develop specific
  methods to have a slightly better accuracy ?
• Invariance
• Robustness to small transformations in test
  images
• Local approach more robust than global approach
  (many local feature vectors are left more or less
  intact by a given image transformation)

Future work directions

• Improving robustness
• Augmenting the learning sample with transformed
  versions of the original images
• Normalization of sub-window sizes and
  orientations
• Speed/accuracy trade-off for prediction
• Combining Sub-windows with other Machine
  Learning algorithms