Docstoc

Hou

Document Sample
Hou Powered By Docstoc
					Chessmen Position Recognition Using
Artificial Neural Networks


            Jun Hou
           Dec. 8, 2003.
Scenario
  Augmented Reality chess                      Hidden pieces
game – detect the position
of all black chess pieces
 Problem – moving camera,
hidden pieces
    Constraints: 1-2 piece
    moves each time;
    Initial position known

    Problem reduced to:
    whether there is a piece on
    a given square
                                  Fig 1. Illustration of hidden pieces
 Task – Generate
synthesized images, train
ANN, test recognition rate
Feature Selection
   Normalized chessboard squares
     In 2D: find region of interest – the chessboard,
     calculate 3D positions.
     In 3D: divide each chessboard square into m*m
     smaller squares (64*m*m)
     Map 3D square positions onto 2D

     Use average of each square as input


   Camera angles and chessboard square positions
        To compensate for the black ratio difference

Prior probability of each square occupied
by a chess piece
    Neural Network Design

ANN applied:
(1) Tradition GD;
                       m*m                Square       Camera       Prior
(2) Gradient Descent   Image data         position     parameters   probability

    with variable
                                                                    of     the
                       m*m input          2 input      6 input      square
    learning rate;                                                  1 input

(3) Gradient Descent
    with momentum                   Fig 2. Neural Network Design
Experimental Results – Data
generation and preprocessing
Use Blender Version 2.30 to model the pieces and
use Python scripts to generate the synthesized
images
    Change the chessboard state – play one random move
    each time, and observe the chessboard from different
    viewpoints
    Image properties – clean, high contrast, little noise, no
    distortion, no lighting variation.
    Generate 1500 images with 640*480 resolution



Data filtering

    Use threshold to select the squares that have at least a
    portion of black, exclude completely blank squares for
    training
      Experimental Results – Data
      generation and preprocessing (cont.)
(a)                                             (b)




  Fig.3 (a) Original synthesized 640 * 480 3D image and (b) converted 64*64 2D
  image. Only none white squares are used for training. Trying to restore the 3D
  image to 2D image. The 2D chessboard looks as if rotated 45.
      Experimental Results – Train ANN




     Fig.4 Training using GD with         Fig.5 Training using GD
     variable learning rate               with momentum

Recognition rate: 72%, GD with variable learning rate 1.05
No significant recognition rate difference between the GD, GD-VLR, GD-M
GD-VLR and GD-M are faster than GD
     Experimental Results
The more training samples,
the more similar are the
training and test sets
As the number of training
samples increases, the
training set accuracy and
test set accuracy merge
                             Fig.6 Training set accuracy and test
                             set accuracy. GD with variable
                             learning rate
Conclusion and Future Work
   Conclusion
     Variations of Feed Forward ANNs are used to
     recognize the positions of chess pieces.

   Future work
     Adjusting ANN parameters
         Size of square divisions, effect of randomizing the training
         data, learning rate, momentum, etc.
     Use Confidence measure
     Use constraints to help recognition
         Look at successive frames to find out the piece moved;
         use frames of different viewpoints for one chessboard state,
         etc.
Thank you! *^_^*

				
DOCUMENT INFO