W2D: 2 dimensional warping for image recognition

Description

W2D is a software package for appearance based image recognition supporting different pixel to pixel deformation models:

no deformation
image distortion model (IDM)
warped wake two-dimensional hidden Markov model (WW)
pseudo two-dimensional hidden Markov model (P2DHMM)
pseudo two-dimensional hidden Markov distortion model (P2DHMDM)
simulated annealing two-dimensional hidden Markov distortion model (SA)
NEW: Hungarian distortion model (HDM)

Authors

W2D was developed in the course of the diploma thesis of Christian Gollan in 2003 (tutor: D. Keysers, supervisor: H. Ney). It has been (and continues to be) extended by Daniel Keysers and Thomas Deselaers with some new deformation models.

License

W2D is released unter the GNU General Public License Version 2.

Download

w2dsrc.tar.bz2

You are welcome to use the code under the terms of the licence for research or commercial purposes, however please acknowledge its use with a citation to one of the publications below, e.g. [1] for handwritten characters or general images and [2] for medical images.

The code is in constant development and more features will be added in the future. If you find bugs or have questions please contact one of the authors.

Publications

D. Keysers, C. Gollan, and H. Ney. Local Context in Non-linear Deformation Models for Handwritten Character Recognition. In ICPR 2004, International Conference on Pattern Recognition, Cambridge, UK, August 2004.
D. Keysers, C. Gollan, and H. Ney. Classification of Medical Images using Non-linear Distortion Models. In BVM 2004, Bildverarbeitung für die Medizin 2004, Berlin, Germany, pages 366-370, March 2004. (.pdf)
C. Gollan. Non-Linear Deformation Models for Image Object Recognition (in German). Diploma Thesis, Lehrstuhl für Informatik VI, RWTH Aachen University, Aachen, Germany, September 2003. .pdf, .ps, .ps.gz)

Results

USPS: 2.4-1.9%
MNIST: 0.54% using -MID -W2 -Dsobelv:euc2rec3x3:1,sobelh:euc2rec3x3:1 and the 28x28 sized images
IRMA (not publicly available) 6.6-5.3%

For more results see one of the publications above.

Usage

Example usage to obtain a result of 2.4% error rate on the USPS corpus with the image distortion model:
```
W2D usps_train.jf usps_test.jf -Dsobelv:euc2rec3x3:0.5,sobelh:euc2rec3x3:0.5 -MID -W2 | classifier -
```
meaning:
- use usps_train.jf as train data, usps_test.jf as test data
- use the following features: vertical sobel with 3x3 sub images and squared Euclidean distance and weight 0.5, same for horizontal gradient
- use the distortion method ID, i.e. image distortion model method
- use warp range 2

Example usage to obtain a result of 2.2% error rate on the USPS corpus using the Hungarian distortion model:
```
W2D usps_train.jf usps_test.jf -Dsobelv:euc2rec3x3:0.5,sobelh:euc2rec3x3:0.5,posx:euc2:0.005,posy:euc2:0.005 -MHUNG | classifier -
```
meaning:
- use usps_train.jf as train data, usps_test.jf as test data
- use the following features: vertical sobel with 3x3 sub images and squared Euclidean distance and weight 0.5, same for horizontal gradient, x-position with squared Euclidean distance and weigth 0.005, same for y-position
- use the distortion method HUNG, i.e. Hungarian distortion method

Example usage to obtain a result of 1.9% error rate on the USPS corpus using the pseudo-2D hidden Markov distortion model:
```
W2D usps_train.jf usps_test.jf -Dsobelv:euc2rec3x3:0.5,sobelh:euc2rec3x3:0.5 -MP2DHMD -O3 -LNO-N1000-Dimage:euc2:1.disdat.gz | classifier -k3  -
```
meaning:
- use usps_train.jf as train data, usps_test.jf as test data
- use the following features: vertical sobel with 3x3 sub images and squared Euclidean distance and weight 0.5, same for horizontal gradient
- use the distortion method P2DHM, i.e. pseudo-2D hidden Markov distortion model
- add a border of 3 pixels around the images
- use the first 1000 nearest neighbors according to the file NO-N1000-Dimage:euc2:1.disdat.gz (which was produced using NO distortion and "pure" Euclidean distance)
- use the 3-nearest neighbor classifier